strands-labs/robots roadmap | v0.3.9 → v0.4.0 → v0.5.0

[cagataycali]

Pinned roadmap -- releases, merged PRs, active work, what's next.

Current tag: v0.5.1 (shipped 2026-08-06) -- the latest published tag. Previous: v0.5.0 (2026-08-04), v0.4.1 (2026-07-01).
State: v0.5.0 (2026-08-04) and v0.5.1 (2026-08-06) are both out; the planned v0.4.2 patch was never cut, so the post-v0.4.1 wave shipped as a minor tag. v0.5.0 published the Isaac Sim backend absorption (#1156, closing epic #1144), the remote policy inference client/server split (#1142), the floating-base state lane, the locomotion reward-term DSL composed into runnable go2/g1/t1 benchmarks, the terrain curriculum, the describe() discovery surface and the lerobot>=0.6.0 floor (#1114). v0.5.1 is a correctness-only patch over three days, headlined by the lerobot>=0.6.1 floor that makes stream_dataset(repo_type="bucket") resolver-guaranteed rather than lucky (#1930). Since the v0.5.1 tag a further 1,451 PRs (#1987 -> #3750) have accrued on a fresh [Unreleased] wave -- 981 fixes + 158 test pins against 151 features, 78.5% correctness. Its capability surfaces are the native hardware-driver programme (twelve buses across 25 registry robots, #2734 seam onward), the agent-facing Unitree G1 verb surface ported from neon-the-g1 and then consolidated 79 modules -> 22 (#3037/#3070), two humanoid policy providers (KimodoPolicy #2270, ProtoMotionsPolicy #2287), Isaac backend parity, photoreal 3DGS + GS-derived lighting, the VLM-judge batch, one drive contract for every mobile base (#2689/#2696), the dashboard -- which in the newest band finally got a server (#3695), taking that package to fifteen modules plus a served static/ bundle -- FastTD3 (#3206), the rollout observer lane (#3351) and a deployable RL checkpoint through the policy factory (#3414). Roughly 116,000 lines have come out across thirteen net-negative refactors, and the newest band ends the featureless run at nine bands: 73 PRs, 277 files, +18,945 / -682 -- net +18,263 lines, the largest net-positive band this page has recorded, carrying seven feat commits after nine bands with none. The removals ledger is nonetheless unchanged for a fifth band (vera, lerobot_calibrate, motionbricks, strands_robots.transforms, the MolmoAct2 normalization port), the net-negative refactor table holds at thirteen, and for a fifth band the absence of a ! or a BREAKING CHANGE trailer is correct rather than a gap -- nothing public left main. It also closed the oldest small item on this page: the 0000- fragment placeholder defect is fixed (#3715/#3721/#3722), so changelog.d/ holds none for the first time since they appeared and sequencing item 6 is retired. Its recurring fix idioms each have a section below: a stop reports its outcome rather than asserting one; a published budget is measured on a clock that cannot step; a shared value domain replaces a hand-rolled bound; a posture flag is read as a value, not by truthiness; a refusal names a remedy that works; the whole-tree graders grade the whole tree; a span that resolves to no work is refused; a cache is served only while its key still names what the entry is about; a run that stopped early states how much never ran; one owner answers one question; a verdict names the thing it was written for; a store is written whole, or the previous one is still there; a telemetry field the frame does not carry is not a reading; an operator is asked before a command reaches a real actuator (#3418/#3420/#3430/#3431, widened to every unknown verb by #3508); a docs page states each contract once; a value that can never be honored is refused at the entry point (thirteen fixes, #3502 the sharpest); a resource a failed bring-up adopted is released (#3519/#3522/#3526/#3530/#3532); an action naming a key the surface cannot carry is refused (#3534/#3537/#3539); a name is held to the alphabet its consumers can carry (#3557/#3558/#3559/#3565/#3570); a run that commanded nothing is refused rather than scored (#3584/#3608/#3612/#3613/#3620, and now at the rollout entry point itself by #3680); a shutdown that cannot complete does not keep the resource (#3590/#3596/#3602/#3603); an offline snapshot names what the live registry registers (#3614/#3616/#3618/#3623); a criterion already satisfied at reset is reported, not read as a result (#3629/#3633/#3637/#3639/#3642); a plan that commands nothing is a failure, not a plan (#3643/#3652/#3655/#3657); a ROS 2 name must be able to name a topic (#3628/#3635/#3636/#3638/#3648/#3651); a reader grades the size its own payload declares (#3627/#3650/#3658); a refusal applies nothing from its own call (#3622/#3625/#3645); and now the band's own four: an unreadable reply names the peer, not the codec (#3670/#3672/#3676/#3677), a configuration entry that is not a string is refused (#3666/#3667/#3674/#3679), a motion that did not reach names what stopped it (#3703/#3705/#3727/#3733) and a camera name answers for the camera it names (#3683/#3691/#3693/#3730). The tracker moved again after a band in which nothing moved at all: two issues were filed and closed inside the band against merged code (#3664 via #3666, #3684 via #3686) and two more were filed and are still open (#3734, #3744), so open issues went 3 -> 5. The review queue went 11 -> 21, the largest this page has recorded, and the extra-approval remedy is now the board's normal path rather than a novelty: seven heads read APPROVED (#3729, #3743, #3746, #3749, #3751, #3752, #3753). What still waits on a second account is #3205, #3343, #3427, #3496, #3515, #3577, #3615, #3714, #3726 and #3737 -- each green, armed, every thread resolved, and therefore still REVIEW_REQUIRED under require_extra_approval_for_unattributed_changes.
Status: CLOSED as completed 2026-08-14 -- this roadmap is retired. It stays pinned as the historical record of the v0.3.9 -> v0.4.0 -> v0.4.1 -> v0.5.0 -> v0.5.1 arc; live tracking is the 0.5.0 milestone and the open issues, not this document. Reconciled once more below so a pinned page does not read as current plan.
Last synced: 2026-09-16 against main @ #3750 -- 73 PRs have merged since the previous marker, taking the wave to 1,451 PRs (#1987 -> #3750) since the v0.5.1 tag: 1,378 + 73 = 1,451. The band spans #3661 -> #3750 with two out-of-order merges (#3661 and #3662, both rows on the previous Open-PRs table). Caveats for anyone recounting. The counting streak broke again, and in the direction that under-reports: the tip commit is #3740 while #3750 is the highest merged number, so a marker read off git log's newest subject is ten short. Last sync's marker was also its tip, so a range taken from the #3660 commit yields exactly 73 commits for 73 new pull requests -- no trap this time. The clean sweep broke after two bands, and it broke on the shape #3169 measured: seventeen numbers inside #3661 -> #3750 are absent from main -- nine open pull requests (#3673, #3714, #3726, #3729, #3737, #3739, #3743, #3746, #3749), two open issues (#3734, #3744), two issues closed in band against merged code (#3664, #3684), and four pull requests closed unmerged, every one a second implementation (#3665 against merged #3666, #3671 against merged #3672, #3736 against merged #3731, and #3732 descoped into the fix #3733 plus the tracker item #3734) -- instances sixteen to nineteen of that shape, after two bands at zero. The 0000- placeholder defect is closed: changelog.d/ holds none, where it held four for three syncs, and the guard now refuses the number at every surface (#3715/#3721). One bookkeeping stray remains: #3702's body carries Closes #767, which names a pull request merged in the v0.4.1 wave rather than an issue, so it closed nothing. The two older caveats still stand: one wave commit carries no PR reference at all (cff130dc, a direct test(isaac) push), so git log shows 1,452 commits for 1,451 PRs -- measured exactly at this marker -- and #2801's squash title carries a shell-quoting artifact ((#'2801')), so a naive (#NNNN) grep returns 1,450.

Where we are (2026-09-16 reconciliation)

No new tag since v0.5.1 (2026-08-06) -- it is 41 days old and still the latest published release, with 1,451 PRs (#1987 -> #3750) accrued behind it. The day since the last sync added 73 PRs, the largest band since 2026-09-11. Milestone 0.5.0 held at 11 closed / 0 open (100%) for a nineteenth sync. Open issues 3 -> 5. Open PRs 11 -> 21, the largest review queue this page has recorded.

The featureless run is over at nine bands, and it ended on volume rather than on one feature. 73 PRs: 58 fix + 7 feat + 4 test + 3 docs + 1 refactor, 277 files, +18,945 / -682 -- net +18,263 lines, the largest net-positive band here by a factor of nearly two. feat moves 144 -> 151, fix 923 -> 981, test 154 -> 158, docs 109 -> 112, refactor 24 -> 25, and correctness still climbs, 78.2% -> 78.5%, for a fourteenth consecutive sync -- because the fix count grew faster than the feature count even in the band that broke the streak. The one refactor is examples-scoped (#3709), so the net-negative refactor table holds at thirteen and the removals ledger is untouched for a fifth band: vera, lerobot_calibrate, motionbricks, strands_robots.transforms and the MolmoAct2 normalization port are still gone from main and still present in the last published tag. No commit carries a ! or a BREAKING CHANGE trailer, which is again correct rather than a gap. _SHIPPED_DRIVERS holds at twelve buses across 25 registry robot names and strands_robots/tools/g1 at 22 modules; strands_robots/dashboard/ moves eleven -> fifteen and gains a served static/ bundle.

The seven features are the news, and six of the seven are one theme: the library says more about the thing it just handed you. #3695 is the largest -- the dashboard finally has a server (19 files, +1,609 / -74), the core wired to the rails the previous nine bands landed (auth, access, consent, safety_state, settings, log_redaction, validate_scope, agent_hitl) and failing closed, which is what the #2848 decomposition argument was for. Beside it #3728 (tool_frame declares the tool point a site-less model lacks), #3727 (get_robot_state says which way the arm extends from its base), #3694 (SO-arm joints can be addressed by label rather than by index), #3724 (step feeds an open dataset recording, so a hand-stepped scene records like a rollout), #3731 (Robot(tool_name=) names the tool, so two robots of one type can share one Agent) and #3697 (render_dir on the constructor moves the render sandbox per Simulation).

Its dominant fix shape is a refusal telling the truth about which side failed: an unreadable reply names the peer, not the codec. Four clients decoded a peer's response inside a try that reported a codec error, so an inference server, a Cosmos3 world model, a MoveIt2 planner and a GR00T service each answered a malformed or truncated reply by blaming the local decoder -- which sends the caller to the wrong side of the wire. #3670, #3672, #3676 and #3677 each name the peer and what it sent.

Second, a configuration entry that is not a string is refused. The mirror of the posture-flag lane, on the value side rather than the boolean side: a credential, a name-list entry or a body arrives as None, a number or a nested object, is stored, and is then dropped by whatever consumes it. #3666 is the one with a tracker item behind it -- a DDS Security credential that is not a string was accepted and the participant QoS then dropped it, so a run believed it was authenticated (closes #3664) -- with #3667 (a dashboard settings list entry), #3674 (a mesh ACL cert_common_names entry) and #3679 (a Reachy daemon body that is not an object).

Third, a motion that did not reach names what stopped it. The verified-halt lane asked whether a stop had stopped; this asks the same question of a move. #3733 makes a move_to that fell short name the contact or the joint limit that stopped the servo, rather than reporting the pose it was asked for; #3705 makes set_gripper say what the fingers closed on; #3703 makes attach_bodies report whether the bodies actually touch; and the feature #3727 answers the question underneath all three, which is which way the arm is pointing. #3732 is the instructive part of this lane: it was the feat version of the same idea, and it was closed unmerged in favour of the narrower fix #3733 plus a tracker item, #3734, for the rotate_wrist half -- a descope recorded rather than dropped.

Fourth, a camera name answers for the camera it names. Four fixes on the surface that has produced more naming defects than any other: #3683 (every camera name the sim lists can be rendered through it), #3693 (a camera enumeration offers each camera once, not once per name), #3691 (add_camera/remove_camera accept camera_name where every sibling does) and #3730 (start_recording names the cameras by a name that answers for them).

• The featureless run ends at nine bands: 73 PRs, 58 fix + 7 feat + 4 test + 3 docs + 1 refactor, +18,945 / -682, net +18,263; feat 144 -> 151, correctness 78.2% -> 78.5%.
• An unreadable reply names the peer, not the codec (4 fixes, new lane): fix(inference): a frame the client cannot read names the peer, not the codec #3670, fix(policies/cosmos3): an unreadable frame names the peer, not the codec #3672, fix(policies/moveit2): a reply the client cannot read names the peer, not the codec #3676, fix(policies/groot): a reply the client cannot read names the peer, not the codec #3677.
• A configuration entry that is not a string is refused (4 fixes, new lane): fix(ros_telemetry): a DDS Security credential that is not a string is refused #3666 (closes ros_telemetry: _validate_dds_security_config accepts a None credential the participant QoS then drops #3664), fix(dashboard/settings): a list entry is a string or the list is refused #3667, fix(mesh/acl): a cert_common_names entry is a non-empty string or the ACL is refused #3674, fix(drivers/reachy): a daemon body that is not an object is refused, not read as one #3679.
• A motion that did not reach names what stopped it (4, new lane): fix(simulation): attach_bodies reports whether the bodies touch #3703, fix(simulation): set_gripper says what the fingers closed on #3705, feat(mujoco): get_robot_state says which way the arm extends from its base #3727, fix(simulation): a move_to that did not reach names the contact or joint limit that stopped the servo #3733.
• A camera name answers for the camera it names (4): fix(simulation/mujoco): every camera name the sim lists can be rendered through it #3683, fix(sim): add_camera/remove_camera accept camera_name for name #3691, fix(simulation/mujoco): a camera enumeration offers each camera once, not once per name #3693, fix(recording): start_recording names the cameras by a name that answers for them #3730.
• Seven features, the dashboard server among them: feat(sim): SO-arm joints can be addressed by label #3694, feat(dashboard): the server core, wired to the landed rails, fails closed #3695, feat(sim): render_dir on the constructor moves the render sandbox per Simulation #3697, feat(simulation/mujoco): step feeds an open dataset recording #3724, feat(mujoco): get_robot_state says which way the arm extends from its base #3727, feat(registry): tool_frame declares the tool point a site-less model lacks #3728, feat(robot): Robot(tool_name=) names the tool, so two robots of one type share an Agent #3731.
• The 0000- fragment defect is closed, and sequencing item 6 with it: fix(changelog): a 0000 fragment is a placeholder and is refused #3715 refuses the placeholder number, fix(changelog): the placeholder-number rule has one owner, so every surface refuses it #3721 gives that rule one owner so every surface refuses it, fix(changelog): the mesh discovery fragment names the PR that carries it #3722 renames the one fragment that named no PR. changelog.d/ holds zero placeholders for the first time since they appeared.
• A rollout is exclusive, and one that commanded nothing is an error: fix(simulation): run_policy and eval_policy are refused while another rollout drives the robot #3741 refuses run_policy/eval_policy while another rollout drives the robot; fix(simulation): a rollout that never commanded the robot is an error, not a completed one #3680 carries the fix(dataset_recorder): an action column present as None is refused, not recorded as 0.0 #3584 -> fix(simulation): an evaluation that never commanded the robot is refused #3620 lane to the rollout entry point itself.
• A surface offers a next step the thing it just made can take: fix(simulation/mujoco): add_robot offers a next step the model it added can take #3686 (closes simulation: add_robot advertises run_policy for a model with zero actuators #3684 -- add_robot had advertised run_policy for a model with zero actuators), fix(simulation): the sim tool description describes the world the agent is joining #3687 (the sim tool description describes the world the agent is joining), fix(command_gate): a paused real-robot call tells the script how to resume it #3719 (a paused real-robot call tells the script how to resume it), fix(policies/groot): an unreachable server fails fast and names the remedy #3742 (an unreachable GR00T server fails fast and names the remedy).
• The dashboard's own correctness pass, same band as its server: fix(dashboard): a caller-supplied value cannot forge a log entry #3710 (a caller-supplied value cannot forge a log entry), fix(dashboard): a credential store that parses but cannot serve as one is recovered, not answered with a 500 #3738 (a credential store that parses but cannot serve as one is recovered, not answered with a 500).
• Four test pins: test(drivers/earthrover): every SDK-facing verb answers at the same doors #3662, test(docs): an env var read through a family prefix is graded, and the dashboard auth knobs are documented #3669, test(simulation/mujoco): an actuator target id outside its own space names no body #3675, test(dashboard): the sealed-store cell enrolls passkeys it can then withhold #3745.
• The rest: fix(recorder): a frame missing every declared column is refused like one missing one #3663, fix(registry): talos names the model that carries its 32 actuators #3681, fix(tools): use_rosbridge refuses a missing name before the dial #3682, fix(simulation/mujoco): a robot's action keys are its owned actuators, not the namespace-matching ones #3685, fix(policies): MockPolicy commands inside each actuator's ctrlrange #3688, fix(examples): mesh discovery reads the peer row's type and waits a heartbeat #3689, fix(mesh): startup posture warnings are said once per process #3690, fix(dyld): a quiet import; the ffmpeg remedy is said when video is decoded #3692, docs(install): create and activate the venv before uv pip install #3696, fix(training): the Training overview names the relative-action roster the gate accepts #3698, fix(sim): step under an active recording says it captures nothing #3699, fix(simulation): eval_policy without a success criterion prints no success fraction #3700, fix(simulation): the bundled benchmark roster is derived, not hand-written #3701, fix(tools): train_policy refuses an unknown action first #3702, fix(policies/microduck): a bare weight name is fetched from Pollen's Hub repository #3704, fix(simulation/mujoco): renderers a script leaves alive are closed before the EGL display terminates #3706, fix(policies/lerobot_local): the state-key remedy stops naming an embodiment this policy rejected #3707, fix(examples/locomotion): scripted_g1 delivers every segment it records #3708, fix(simulation/mujoco): a relative scene name means the scenes directory, on write and on read #3711, fix(examples/microduck): render_video places the ball where the kick weights were trained to find it #3712, fix(examples/microduck): render_video says when a ride runs off the rendered floor #3713, docs(policies/wbc): say how to record a walking G1 - add a camera that follows before the rollout #3716, docs(getting-started): the Jetson install path drops the numpy<2 pre-pin the next line undid #3717, fix(simulation/mujoco): get_features counts the list it is introducing #3718, fix(policies/lerobot_local): the inert-normalization warning names the stats the checkpoint ships #3720, fix(simulation/mujoco): a joint write says what unit its numbers are in #3723, fix(mujoco): add_robot records the registry entry its model came from #3725, fix(tools): a joint's configured travel is the scale pose_tool drives it on #3735, fix(robot): a hardware-only keyword on a simulated Robot() is refused, naming the mode #3740, fix(policies): create_policy refuses a misspelled provider kwarg, naming the parameter meant #3747, fix(hardware_robot): real-mode status names the device, and an idle arm is not an error #3748, fix(docs,mesh): the quickstart tells the peer which policy to run - the wire refuses an execute without one #3750.
• The earlier bands are unchanged and still the bulk of the wave: the 81-PR g1 verb port (consolidated to 22 modules), the twelve-bus native-driver programme, the Isaac asset-reader run, the shared-value-domain sweep, Microduck hardening, the doctor lane, the whole-tree grader audit, the cache-validity lane, the mobile-base drive contract, FastTD3, the observer lane, the operator-consent lane, the docs diet, and the refuse-what-cannot-be-honored, release-what-a-failed-bring-up-adopted, checked-posture-flag, refused-payload-key, name-alphabet, commanded-nothing, failed-shutdown, offline-snapshot, criterion-already-satisfied, plan-that-commands-nothing, ROS-2-name, declared-size and refusal-applies-nothing lanes. Their sections below are unedited.

Release hygiene is still the only standing gap, and it is now 1,451 pull requests wide. There is no ## [0.5.0] or ## [0.5.1] heading -- CHANGELOG.md has not been edited since #1690 (2026-07-28) -- [Unreleased] is still frozen at 168 entries, and the real accrual is still not in it: changelog.d/ now holds 1,601 .md files, of which 1,600 are fragments (README.md is not one). The band added 70 and deleted none: 1,530 + 70 = 1,600 reconciles exactly, and scripts/assemble_changelog.py --check reports OK (1600 pending) -- collapsing both shipped tags remains a release action rather than a repair. 73 PRs produced 70 fragments, the four test commits owing none and one PR carrying two, which is the convention working as intended. What did change is the naming: the four 0000- placeholders are gone and the guard now rejects the number at every surface, so sequencing item 6 -- unactioned for three syncs -- is closed. The gap between the fragments and the tag remains the page's oldest unaddressed item -- 50 days since the last CHANGELOG edit, and every one of those 1,600 fragments is a release note nobody has read.

flowchart LR
    A["v0.3.8<br>Feb 20"] --> B["v0.3.9 mesh security<br>folded into v0.4.0"]
    B --> C["v0.4.0 SHIPPED<br>Jun 17"]
    C --> D["v0.4.1 SHIPPED<br>Jul 1<br>ROS2 / training+RL / VERA / RTC / Newton / verify-dataset / security / EGL-GPU"]
    D --> E["v0.5.0 SHIPPED<br>Aug 4 -- 805 PRs<br>Isaac backend absorbed + remote inference + floating-base state + loco reward DSL + go2 walk/strafe/turn + g1/t1 benchmarks + terrain curriculum + describe() surface + lerobot 0.6"]
    E --> F["v0.5.1 SHIPPED<br>Aug 6 -- 51 PRs<br>latest tag; correctness-only patch (lerobot >=0.6.1 floor, hf_hub >=1.5, Isaac reset, LIBERO deadlock)"]
    F --> H["post-v0.5.1 wave (Unreleased)<br>1,451 PRs #1987-#3750<br>981 fixes + 158 test pins vs 151 features<br>+ twelve native drivers across 25 robots (Dynamixel, Feetech, Franka, G1, Go2, Reachy Mini, Microduck, Robotiq, Booster, UR, Crazyflie, EarthRover)<br>+ agent-facing G1 verb surface ported from neon-the-g1, then consolidated 79 modules -> 22<br>+ net-negative refactors: -34,302 (#3037) -33,213 (#3056) -15,204 (#3389 vera provider) -8,571 (#3070) -6,064 (#3426 augmentation transforms) -4,583 (#3419 motionbricks provider) -4,425 (#3388) -2,798 (#3391) -2,313 (#3386 lerobot_calibrate) -2,179 (#3411); ~116,000 lines out<br>+ verified-halt lane reaches drivers, sessions, the inference server and every backend's rollout release<br>+ whole-tree graders audited: roster derived from the tree, walk roots resolved; then one linter rule replaces fourteen copies (#3417)<br>+ refuse a span that resolves to no work + dashboard slices land, the 481-file draft #2848 closed<br>+ a cache is served only while its key still names what the entry is about<br>+ a run that stopped early states how much never ran; coverage on the PEP 669 core (+8.2% vs +76.6%)<br>+ the dashboard becomes the feature lane: HITL gate over hardware-moving tool calls, e-stop Lockout type, preflight scope, consent built from a refusal code (6 modules -> 11)<br>+ one owner answers one question; a key names the entry it is about; a verdict names the thing it was written for<br>+ FastTD3 lands as a native off-policy RL provider (#3206); an RL checkpoint becomes deployable through the factory (#3414)<br>+ a shared value domain reaches both owners of one value, and both halves of a dialled address<br>+ a value is graded before it is converted; a packaging declaration reaches the install that needs it<br>+ a store is written whole, or the previous one is still there; a dataset root resolves to the directory its repo_id recorded to<br>+ the rollout observer lane lands (#3351) and #3212's peer retention lands (#3352), both attributed<br>+ a posture flag is checked, so a word for no cannot select the other branch (closes #3356)<br>+ a telemetry field the frame does not carry is not a reading (Unitree + Booster decoders)<br>+ an operator is asked before a command reaches a real actuator (#3418 #3420 #3430 #3431)<br>+ the docs state each contract once: README to 120 lines, seven pages, -3,575<br>+ a value that can never be honored is refused at the entry point, not scored 0% (#3502 and twelve siblings)<br>+ a resource a failed bring-up adopted is released (#3519 #3522 #3526 #3530 #3532)<br>+ an evaluation reports how close it came and how often it holds (#3521, both routes by #3528)<br>+ first band that removed nothing: no refactor, +10,234 / -483<br>+ a posture flag is checked, not read by truthiness, on nine more surfaces including every policy provider<br>+ an action naming a key the surface cannot carry is refused (#3534 #3537 #3539)<br>+ a name is held to the alphabet its consumers can carry (#3557 #3558 #3559 #3565 #3570)<br>+ a stop grades the verdict it was given, on every robot_mesh route (#3560 #3566 #3571)<br>+ one camera-routing precedence honoured on every router (#3545 #3550 #3555)<br>+ a private name must have a reader (#3580 #3583)<br>+ a run that commanded nothing is refused, not scored (#3584 #3608 #3612 #3613 #3620)<br>+ a shutdown that cannot complete does not keep the resource (#3590 #3596 #3602 #3603)<br>+ an offline snapshot names what the live registry registers (#3614 #3616 #3618)<br>+ a door refuses a collaborator it cannot use (#3617 #3619)<br>+ a criterion already satisfied at reset is reported, not read as a result (#3629 #3633 #3637 #3639 #3642)<br>+ a plan that commands nothing is a failure, not a plan (#3643 #3652 #3655 #3657)<br>+ a ROS 2 name must name a topic; a refused call leaves no node behind (#3628 #3635 #3636 #3638 #3648 #3651)<br>+ a reader grades the size its own payload declares (#3627 #3650 #3658)<br>+ a refusal applies nothing from its own call (#3622 #3625 #3645)<br>+ eleven test pins in one band (#3621 #3626 #3632 #3634 #3641 #3644 #3646 #3647 #3656 #3659 #3660)<br>+ an unreadable reply names the peer, not the codec (#3670 #3672 #3676 #3677)<br>+ a configuration entry that is not a string is refused (#3666 #3667 #3674 #3679)<br>+ a motion that did not reach names what stopped it (#3703 #3705 #3727 #3733)<br>+ a camera name answers for the camera it names (#3683 #3691 #3693 #3730)<br>+ the dashboard gets a server, wired to the landed rails and failing closed (#3695)<br>+ the featureless run ends at nine bands: 7 feat (#3694 #3695 #3697 #3724 #3727 #3728 #3731), newest band +18,945 / -682"]
    H --> I["Next tag<br>uncut; no [0.5.0] or [0.5.1] heading, 1,600 changelog.d fragments pending, CHANGELOG untouched 50 days<br>the only standing gap -- open PRs 11 -> 21, the largest queue recorded; seven heads read APPROVED (#3729 #3743 #3746 #3749 #3751 #3752 #3753)<br>#3205 #3343 #3427 #3496 #3515 #3577 #3615 #3714 #3726 #3737 are green + armed + threads resolved and still REVIEW_REQUIRED on require_extra_approval_for_unattributed_changes<br>provider direction being executed (#3075): GR00T + MolmoAct stay; vera and motionbricks removed, lerobot_async / curobo / moveit2 still to go<br>removals section must still be assembled by hand -- five bands running have removed nothing<br>the 0000- fragment placeholder defect is FIXED (#3715 #3721 #3722): zero placeholders, guard refuses the number at every surface<br>the clean sweep broke: 4 PRs closed unmerged in #3661-#3750, each a second implementation (#3665 #3671 #3732 #3736)<br>plus a 108-file Isaac contribution ready for review (#3343), unchanged for seven syncs"]

    C --> G["0.5.0 milestone DRAINED<br>11 closed / 0 open (100%) -- held 15 syncs<br>#469 #472 #473 #474 closed 2026-08-24<br>substantiation issue #2720 and epic #858 both closed NOT_PLANNED 2026-09-10"]
    style A fill:#2ea44f,color:#fff
    style B fill:#2ea44f,color:#fff
    style C fill:#2ea44f,color:#fff
    style D fill:#2ea44f,color:#fff
    style E fill:#2ea44f,color:#fff
    style F fill:#2ea44f,color:#fff
    style H fill:#d29922,color:#fff
    style I fill:#0969da,color:#fff
    style G fill:#2ea44f,color:#fff

Loading

Published releases

Tag	Date	Notes
v0.3.8	2026-02-20	prior baseline
v0.4.0	2026-06-17	mesh security (v0.3.9 scope) + docs/policies/motion-planners/HW-compat, collapsed into one tag. MoveIt2 #305 + hardware_robot coverage #453 landed before the cut.
v0.4.1	2026-07-01	the large post-v0.4.0 wave: ROS 2/RTPS interop, end-to-end training + from-scratch RL (PPO/FastSAC, VecSimEnv, BaseRLAlgo.evaluate()), VERA, RTC/chunking correctness, recording + verify-dataset integrity, sim discovery/validation, EGL-GPU render reliability, filesystem/dependency security hardening, Newton GPU-backend parity. Tag at #947; release-enabled by #950 (PyPI vera-extra fix). CHANGELOG [Unreleased] collapsed into the dated 0.4.1 heading in #977.
v0.5.0	2026-08-04	the post-v0.4.1 wave as a minor tag (the planned v0.4.2 patch was never cut): Isaac Sim backend absorbed (#1156, closing epic #1144), remote policy inference client/server split (#1142), floating-base state lane, locomotion reward-term DSL composed into runnable go2/g1/t1 benchmarks, terrain curriculum, describe() discovery surface, lerobot>=0.6.0 floor (#1114). Published 2026-08-04T01:32:07Z.
v0.5.1	2026-08-06	latest published tag (published 2026-08-06T20:23:39Z, tag commit 00cccb06). A correctness-only patch, 51 commits over three days, no API changes: raises the lerobot floor to >=0.6.1,<0.7.0 across [lerobot]/[lerobot-async]/[molmoact2] so stream_dataset(repo_type="bucket") is resolver-guaranteed rather than lucky (#1930), raises the [wbc] huggingface_hub floor to >=1.5.0 for the hf buckets/hf sync CLI (#1957), fixes Isaac 6.0.x reset() articulation-handle revival (#1901, closes #1895) and the LIBERO eval Kit/agent thread deadlock (#1899, closes #1896), and continues the numeric-input hardening pass across training, tools, inference, hardware and the sim backends.

Post-v0.5.1 -- the current [Unreleased] wave (1,451 PRs, #1987 -> #3750)

Merged to main after the v0.5.1 tag (2026-08-06). 981 fix + 151 feat + 158 test + 112 docs + 25 refactor + 17 ci + 4 security + 1 perf + 1 examples + 1 chore -- 78.5% correctness and contract-pinning, climbing for a fourteenth sync. The newest band is 58 fix + 7 feat + 4 test + 3 docs + 1 refactor, so fix moved 923 -> 981, feat 144 -> 151, test 154 -> 158, docs 109 -> 112 and refactor 24 -> 25 while ci held at 17. feat moving at all is the band's defining fact: it ends a nine-band featureless run, and at 277 files, +18,945 / -682 -- net +18,263 lines it is the largest net-positive band on this page, with none of its 73 commits carrying a ! or a BREAKING CHANGE trailer -- correct here, because nothing public was removed. Its one refactor is examples-scoped (#3709), so the net-negative refactor table holds at thirteen. No new vendor bus, so _SHIPPED_DRIVERS holds at twelve across 25 registry robot names and strands_robots/tools/g1 at 22 modules; strands_robots/dashboard/ moves eleven -> fifteen on #3695's server.py, access.py, routes_auth.py and cli.py, plus a served static/ bundle. strands_robots/transforms/ no longer exists, nor does strands_robots/policies/motionbricks/. perf remains a scope with exactly one member (#3149). Heaviest scopes across the wave: simulation (111), tools/g1 (86), mesh (86), simulation/mujoco (55), tools (48), training (37), drivers (33) plus 32 more under a drivers/* sub-scope (65 in the family), policies (30), isaac (29), simulation/isaac (24), ci (21), sim (20), rendering (19), policies/microduck (17), hardware (17), drivers/g1 (17), policies/vera (17, a removed surface), device_connect (15), registry (13), tests (12), scripts (12), policies/wbc (12), policies/cosmos3 (12), examples/fleet (12), dashboard (12), transforms (11, a removed surface), g1 (11), agents (11), training/rl (10), policies/protomotions (10), dataset (9), assets (8).

Isaac backend parity (the largest feature lane)

PR	What
#2164	promote run_multi_policy to the Simulation base with shared validation (the parity groundwork)
#2175	extract a backend-agnostic MotionPrimitivesCore from the MuJoCo mixin
#2178	synchronized multi-robot run_multi_policy control loop on Isaac (no recording)
#2200	Isaac joint-space motion primitives -- set_gripper / rotate_wrist
#2212	Isaac merged multi-robot recording for run_multi_policy
#2218	Isaac IK-backed move_to via the shared MinkIKBridge

The stated scope of both parity issues has landed -- #2122 (synchronized multi-robot rollout) via #2178/#2212 and #2123 (move_to/set_gripper/rotate_wrist) via #2200/#2218 -- but neither issue auto-closed, so both still read as not-started. Both parity issues have since been closed (#2156 with the parity suite in #2233; #2221, the body-mounted add_camera(parent_body=...), on 2026-08-24), so nothing on this lane is open. The backend's own work continued in a different direction -- a 22-PR asset-reader run, tabled below.

Off-policy RL -- FastTD3 lands (epic #858, #2720)

PR	What
#3206	FastTD3 -- twin delayed DDPG as a native off-policy RL provider, with vectorized collection and the trainer env-leak fix (+2,199 / -95 across 29 files)

The first of epic #858's off-policy deliverables to reach main, and the first code answer to #2720's claim that three of them are absent from the tree. It joins the from-scratch PPO (#877) and FastSAC (#886) backends shipped in v0.4.1 and the VecSimEnv / BaseRLAlgo.evaluate() groundwork (#918, #920). It closed neither issue -- the pull request declares no closing reference -- so #858 is still the last unscheduled epic and #2720 is still open against the milestone; the bookkeeping is owed. Its two changelog fragments are named for the issues they answer (862-fast-td3-trainer.md, 2720-rl-trainers-close-their-env.md) rather than for the pull request, the first in the wave to do that.

Humanoid policy providers

PR	What
#2270	KimodoPolicy -- text-to-motion diffusion for the Unitree G1
#2279	Kimodo action-key bridge for lerobot's Unitree G1 driver
#2287	ProtoMotionsPolicy -- native ONNX GTP whole-body tracker for the Unitree G1
#2271	Policy.required_bodies -- a policy declares the body poses the runtime must supply (the seam both trackers need)

Milestone issue #469 was closed COMPLETED on 2026-08-24. That close is one of the 57 #2720 flags as carrying no linked PR, and the reservation this document recorded for four syncs still stands on the code: the provider merged, but the issue's namesake kinematic-constraints API did not -- there is no constraint module under strands_robots/policies/kimodo/, no multi-prompt period splitting, and the registry config_keys enumerate sampling knobs only.

Fleet-orchestration examples (epic #2179)

PR	What
#2197	01_skill_dispatch_multi_vendor -- capability-based dispatch across heterogeneous robots
#2191	03_failover_and_degraded_ops -- peer-loss reassignment + dispatcher-down safety
#2207	04_emergency_evacuation -- three-phase evacuate protocol, benchmark-scored
#2188	05_work_order_dispatch -- work-order ingress onto capability manifests

All five landed (02_cross_zone_transport via #2198), and the epic #2179 closed 2026-08-24. #2506 then added a --serve-web live Rerun viewer for headless hosts, and #2643 corrected the docs to describe a dashboard that is named rather than promised.

Merge gates / CI / supply chain

PR	What
#2039	refuse a lockfile that does not resolve the manifest
#2043	compare the whole recorded requirement set, not two properties of it
#2019	report whether an open pull request already claims an issue
#2005	last-push-approval reports a finding without failing the job
#2241	bound every workflow job so a hung step cannot hold a runner silently
#2244	sweep the open pull-request set for untested compositions
#2044	declare the advisory floors that keep four transitive packages patched
#2769 #2797 #2830	collide duplicate work on the identity a created path declares; then on a module one of the branches' tests names; then on the change a branch describes rather than the file it creates

Dependency floors (new lane)

Every one of these was measured against the capability the code actually imports, rather than bumped speculatively.

PR	What
#2335	raise the robot_descriptions floor to a release that ships every registry robot
#2342	raise the websockets floor to the release that ships the name the server imports
#2345	raise the eclipse-zenoh floor to the release that ships wire source attribution
#2349	bound the vera-sim imageio floor at the release that reads the clip duration in milliseconds
#2340	the same lane inverted -- rclpy is not on PyPI, so the refusal now names a remedy that can supply it
#3236	an abbreviated written version floor is not below the floor pyproject declares (extends #3188)
#3234	the lane read from the other end -- a test pinned an np.cross behaviour removed in numpy 2.5, so the upper half of the declared numpy range was red; the premise now holds across it (closes #3229)
#3601	the lane read one layer further: a pin whose whole admitted range has no wheel for the requires-python the project declares. [ros2] pinned cyclonedds<1.0.0, whose wheels stop at cp310, so the one transport documented as a self-contained wheel was a source build on every supported Python (closes #3599)

Agent-facing discovery + actionable refusals

The dominant theme of the #2308 -> #2387 window. Two shapes: a capability the schema documents but never publishes, and a refusal that dead-ends.

PR	What
#2311	publish set_geom_properties' payload contract in the tool schema
#2320	publish the ellipsoid shape the agent schema already documents
#2308	advertise set_timestep under the parameter name it accepts (Newton)
#2329	advertise every base-contract method the Newton backend delivers
#2372	report the provider aliases create_policy accepts
#2317	diagnose a missing asset instead of suggesting the name it refused
#2364	route the create_world refusal by the request instead of offering reset
#2369	name the hardware route when a registered robot has no sim model
#2379	make a relative output_path refusal name a remedy that works
#2382	offer a joint-discovery action an agent can call
#2377	name the action a misspelled action probably meant
#2327	point the WBC missing-session refusal at the seam that installs one

Security + gating

PR	What
#2313	use_ros gates every verb that can command a robot, not just publish -- supersedes the long-open #1087; #2315 tracks the halt-vs-drive contract it exposed
#2358	mesh security admits every policy provider the registry can build
#2381	gate every spelling of a remote-code provider, not just its canonical name
#2365	let the caller decide whether a Kimodo checkpoint runs its own code
#2855 #2859	escape a line break where an observation-derived value enters a log record, and read the escape's shape off the value the caller receives

Scene-rebuild state + lifecycle (MuJoCo)

A coherent run: a rebuild or a teardown must not silently drop state the caller established.

PR	What
#2360	carry ctrl, act, an unnamed freejoint and the clock across a robot eject
#2370	keep a latched apply_force wrench latched across a scene rebuild
#2380	carry a latched joint force across a scene grow
#2384	apply the actuator command a keyframe pairs with its pose
#2386	refuse to double-actuate a joint a tendon already drives
#2371	report the placement add_robot compiled, not the one it was asked for
#2387	export a scene whose asset references still resolve
#2355	refuse remove_robot when the cooperative stop lapses instead of un-tracking a live worker
#2333	a PolicyServer stop must not kill the thread serving it
#2362	release the robot on Ctrl+C instead of exiting straight out of run()

Monotonic clocks (new lane)

One root cause, six sites: a published time budget was measured by subtracting two time.time() readings, so an NTP correction or a manual clock change mid-budget silently truncated or extended the budget the caller asked for.

PR	What
#2404	measure a published wait budget on a clock that cannot step (agent-callable tools)
#2406	measure a hardware control-loop time budget on a clock that cannot step
#2408	measure mesh peer liveness + publish intervals on a clock that cannot step
#2412	measure a teleop session's duration budget from the end of setup, not including it
#2413	measure a sim rollout's frame pacing + RTC latency on a clock that cannot step
#2425	pace the MuJoCo rendering loops on a clock that cannot step

The lane filed its own remainder rather than leaving it implicit, and has since closed it: #2414 named the three pacers still deciding on time.time() -- the MJPEG stream, the multi-camera recorder and Isaac's idle-render gate -- and was closed on 2026-08-18 (no closing PR recorded; #2432 moved Isaac's preview cadence and its reported recording duration onto a clock that cannot step).

ProtoMotionsPolicy hardening

The ONNX GTP whole-body tracker merged in #2287 one wave earlier; this is its post-merge correctness pass. No public-surface change.

PR	What
#2416	read a quaternion as the rotation it encodes
#2417	fill the reference-motion cache's body rows in the order the tracker reads them
#2418	use the ground an MJCF already declares instead of adding a second one
#2411	name the frame the reference-motion cache's velocity channels are in (docs)

Photoreal 3DGS + GS-derived lighting (new lane)

The five-PR batch tracked as #2321/#2322/#2323/#2325 landed in one window, taking the hybrid Gaussian-splat render path from silently-optional to a declared contract. Four of the five tracking issues closed with it.

PR	What
#2336	fail loud when the photoreal 3DGS background cannot initialize, instead of demoting to the procedural gradient (closes #2321)
#2332	composite premultiplied splat RGB once, alpha-normalize gsplat depth, honor the SPZ antialiased flag (closes #2322)
#2331	light the robot with the scene it stands in -- GS-derived IBL, shadow catcher, colour-pipeline contract (closes #2323)
#2326	evaluate SH view-dependence and swap the GS presets to fully-trained captures (closes #2325)
#2328	forward render_mode to SimulationApp; delete the abandoned renderer helpers (closes #2324)

Its correction pass followed inside the same wave -- the shape every new surface here takes:

PR	What
#2441	state the scale of an sRGB decode input instead of guessing it
#2443	check derive_key_light's search-region flag instead of reading it by truthiness
#2446	check the environment-map resolution knobs before baking with them
#2452	check the panorama bake's resolutions before it names a file or loads splats

Both remaining items -- #2337 (regenerate the shipped gallery) and the parent #1537 (lift the GS layer out of examples) -- were closed 2026-08-24, so the lane is fully closed out. rendering carries 12 PRs across the wave.

Posture flags read as values (new lane)

A flag is a posture the caller chose, so reading it by truthiness silently inverts the choice: "false", "no" and 0 are all truthy strings or a falsy int, and the surface acts on the opposite of what was asked. Checked rather than coerced across four surfaces in one window, continuing the sub-shape #2436/#2443/#2448 opened.

PR	What
#2473	check the ROS 2 command-surface posture flags instead of reading them by truthiness
#2487	check randomize()'s per-axis flags instead of reading them by truthiness
#2475	check the dataset-recorder creation API's posture flags instead of reading them by truthiness
#2485	check GsplatBackground's alignment flags instead of reading them by truthiness
#2489	its resolution-knob sibling: check the value that sizes a heightfield before generating it

Native hardware drivers (a programme, 18 feature PRs) -- twelve buses shipped

Every real robot used to be built by lerobot. The seam that made that optional landed in the #2734/#2739 pair; the #2750 -> #2836 band built three vendor buses and a complete G1 write path behind it, the #2837 -> #2879 band added a second robot, the #2880 -> #3057 band added a sixth driver plus a coverage report that derives the remaining gap -- and the #3037 -> #3091 band doubled the set in a day, landing every driver that had been queued for review. _SHIPPED_DRIVERS now registers twelve buses across 25 registry robot names: Dynamixel, Feetech, Franka, G1, Go2, Reachy Mini, Microduck, Robotiq, Booster, UR, Crazyflie and EarthRover. This is the wave's one genuinely new capability surface and it is now its largest feature lane.

PR	What
#2734	choose which driver builds a real robot -- the seam itself
#2739	native CycloneDDS driver for the Unitree G1 -- first driver that does not route through lerobot
#2750	native Dynamixel Protocol 2.0 codec + driver skeleton
#2786	native Feetech STS/SCS Protocol 1 codec
#2762	native daemon driver for the Pollen Reachy Mini
#2757	DDSPublisher -- the G1's write path
#2767	send_action publishes LowCmd_ on rt/lowcmd
#2779	a 500 Hz control loop behind run_policy / stop_task / get_task_status
#2856	register Pollen Microduck, the 14-DOF open biped
#2862	Microduck end to end -- policy provider, native robotd driver, sim + real
#3039	native CycloneDDS driver for the Unitree Go2 quadruped -- the sixth shipped bus
#3042	list_driver_coverage joins the robot registry and the driver registry, so a driver gap is derived rather than asserted
#3040	native Franka FCI driver for panda, fr3 and fr3_v2
#3053	native Modbus TCP driver for the Robotiq 2F-85 gripper (robotiq_2f85, robotiq_2f85_v4)
#3045	native Booster Robotics T1 driver (booster_t1)
#3058	native RTDE driver for the UR5e and UR10e
#3046	native Crazyflie 2.x CRTP driver
#3081	native EarthRover Mini+ driver over the earth-rovers-sdk HTTP API
#3043	wire the Feetech SCS bus so an SO-arm moves and reports joints
#2947 #2949	cache rt/mainboardstate and rt/pressuresensorstate on the G1, each with its own agent verb
#2745	survive a teardown after the device bring-up failed
#2719	hold the ROS 2 bridge QoS depth to a domain the transport can carry

Its own hardening arrived in the same band: #2802 stops the Dynamixel codec citing a vendor unstuffer that loses a byte; #2780 applies the vendor's per-joint gains on rt/lowcmd rather than one scalar; #2804 holds the battery floor to a numeric domain before it is stored; #2773 records the DDS interface ensure_dds actually bound rather than the one it was asked for; #2755 verifies a Feetech reply instead of reading it at fixed offsets and #2820 refuses the Feetech broadcast where serial_tool is about to read one reply; #2752 and #2756 stop reporting a LiDAR point cap the summary never applies and read the fields LidarState_ actually declares; #2817 names a robot's native driver instead of lerobot's vocabulary. The newest band is almost entirely this lane's own accounting: #2865 makes the write refusal cite a resolvable issue and documents the soft-stop helper's real call site, #2863 and #2866 make the module docstring and the FSM-set docstrings name the verbs and the topic send_action actually writes, and #2872/#2877 stop two deferrals citing changes this repository has already landed. Four test pins hold the un-wired half honest -- #2847 and #2850 grade the rt/lowcmd wire-frame contract and the per-joint gain cells on CI rather than only where the vendor SDK is installed, #2841 pins the battery floor as unreachable behind the un-wired FSM, and #2861/#2874 pin the acceptance criterion as an XPASS trigger whose marker cites the refusal the driver actually returns. A five-PR DDS endpoint-lifecycle sub-run closes what it opens: #2768 closes the endpoints instead of only forgetting them, #2783 releases one whose construction failed part way, #2811 one a concurrent teardown left unreachable, #2816 reads whether the control loop actually halted before releasing the wire, and #2828 makes the stop verb report the halt outcome instead of asserting it.

Three of the four items this lane filed against itself have now closed against merged code: #2749 via #2763 (which reads joints from a driver that owns its bus, pinned by #2754), #2765 via #2888/#2916 (the motion-switcher decoder plus the FSM-producer wiring answered its five rt/lowcmd wire-format decisions), and #3062 via #3064 -- the sport-mode gate now opens its RPC client under the shared _DDS_INIT_LOCK, with #3067/#3071/#3072/#3073 closing the same class across the T1 and the Go2. What the programme is still carrying is #2753 (Dynamixel model-name resolution, against a live servo) and #2812 (the Feetech position field bounded and reported at STS full scale, 4x the SCS-series servo the registry declares) -- both need hardware, and #3074 has since recorded the word order as an STS/SMS property so #2812 is a scale question alone -- plus two decisions the g1 port filed fresh: #2891 (the FSM read cadence, still the thing blocking the send_action wire step) and #2928 (the tools-bundle port has no seam decided, now partly answered in practice by #3037's use_unitree dispatcher).

A stop reports its outcome instead of asserting one (new lane -- the band's dominant fix shape)

One root cause at eight sites: a teardown or halt verb returned success because it had issued the stop, not because the thing had stopped. A loop could still be writing, or a device could be closed underneath one that was.

PR	What
#2809 #2815	report whether stop_teleoperate / InputPublisher.stop actually stopped the loop
#2816 #2828	read whether the G1 control loop halted before releasing the wire; the stop verb reports the outcome
#2821 #2826	do not close the devices under a teleop loop that is still writing; join the loop before detach_teleop disconnects what it reads
#2831	report an emergency stop the robot refused instead of discarding it
#2825 #2832	report the joints an interpolated move actually commanded; answer for every joint a whole-arm reading was asked for
#2834 #2835	abandon a rollout whose bring-up a shutdown already cancelled; halt teleoperation as well as policies on a simulation emergency stop
#2839	report a timed drive whose trailing stop was refused

The lane filed its own remainder as #2833 (a cooperative stop issued before a rollout's first frame is discarded, and every stop path still reports the halt), which is still open -- its first two answers, #2834 and #2835, have since merged.

One drive contract for every mobile base (new lane)

Both of the long-stalled external PRs this document tracked for months landed here -- re-landed as fresh PRs rather than rebased, after which the originals were closed unmerged.

PR	What
#2689	AckermannRosRobot -- Ackermann ROS 2 cars (AWS DeepRacer) as strands robots (the #1035 scope)
#2696	MobileBaseRobot + Transport -- one drive contract for every transport (the #1722 scope)
#2700	name every class that carries a scoped drive limit
#2705	let a RosbridgeRobot command carry an operator decision
#2693	gate every ROS 2 transport's commands, not just use_ros (security; broadens #2313)

Isaac asset readers read what the compiler reads (the largest single theme, 22 PRs)

One shape throughout: an MJCF or URDF attribute was read literally at the site that declares it, where the compiler resolves it through defaults, inheritance and alternative spellings. The backend loaded a model that was subtly not the model on disk.

PR	What
#2669 #2672 #2674 #2675 #2691	read a geom's shape, a joint's every attribute and a mesh asset's scale through the <default> class they inherit; read the root class under both spellings; merge several top-level <default> elements
#2680 #2686 #2692	read every MJCF orientation spelling rather than quat alone; read a link's rotation from it; report a quat normalized, as the compiler does
#2660 #2665	a fromto collision geom carries its length and its offset; read a fromto link's extent from its endpoints
#2656 #2667	read an MJCF's mesh registry and its bodies from the whole model, not just the top file
#2712	read both MJCF spellings of a free joint
#2701 #2694	read a URDF joint's axis default from URDF rather than MJCF; read a link's pose from the joint that reaches it
#2671 #2662 #2659	refuse a body reached by more than one joint; refuse an ASCII-STL facet with no endfacet; read an OBJ keyword off the split
#2498 #2575	LIBERO objects render as real meshes instead of box proxies; refuse a render whose camera_name no camera carries
#2800 #2813 #2827	keep only the joint types URDF declares in the movable set; refuse a non-finite geom, and a non-finite body placement, instead of measuring a body around them
#2838 #2845	refuse a non-finite mesh scale instead of measuring through it -- then refuse it wherever it is carried, not only where it is measured
#3264 #3267	a declaration the format leaves implicit: a URDF joint that states no type is refused rather than welded, and a typeless MJCF geom takes the format's documented default

Shared value domains (the wave's dominant fix idiom)

A parameter validated ad hoc, beside siblings that already share a domain, gets held to that domain -- so one spelling of "what is a legal value here" covers the whole surface.

PR	What
#2547	hold a patch op's claimed name to the shared entity-name domain
#2555 #2559	hold gr00t_inference's selectors and lerobot_camera's documented vocabularies to a domain, like the numerics beside them
#2560 #2563	hold both trainers' device to torch's own device-string domain
#2690	hold every orientation parameter to one quaternion domain
#2706 #2708	hold lerobot_train's save_freq to an int lerobot can decode, and TrainSpec.save_freq to one shared step-cadence domain
#2715 #2731	hold the terrain seed to the shared stream-name domain, and the RL run-size factors to the shared count domain
#2722	grade every field-scoped domain guard, not the five named by one helper spelling
#2751 #2778 #2804 #2810 #2824	hold FastSAC's replay counts, a holdout failure-mode tag, the G1 battery floor, history_length and a background's alignment numbers to the domain each already had beside it
#2837	hold a stored pose to the travel its joints are driven within
#2849 #2860	hold num_envs to the shared count domain on both RL backends, and learning_starts to the domain its relation's other operand already uses
#2858	hold every WBC observation dimension to the shared count domain
#3242 #3244	an infinite max_tokens escapes the dashboard-settings domain that refuses an infinite temperature; a per-ip challenge cap at or above the global cap loses the fairness it exists for
#3238 #3239 #3241	a hidden layer of width zero is not an architecture; a served policy's advertised required_bodies are held to a domain; a skeleton the caller named is not one the builder derived
#3247 #3249 #3251	a dataset identity the provenance record could not hold; a VERA readiness budget with no domain (and the variable its own timeout tells you to raise); an empty cmd_scale resolves from its owner table, not a bare 1.0
#3243	a non-finite wire timestamp is neither stale nor future, so four mesh freshness gates accepted it
#3266 #3275 #3277	two owners of one value: a configured gravity, a configured camera resolution and the parallel environment count each take the domain their call-time argument already had
#3284	the same shape on the RL side -- the counts a vectorized env is built from take the shared count domain
#3256 #3258 #3260 #3270	an RL run's checkpoint interval, FastSAC's target temperature, the rate a target network tracks its online one, and the device every backend places its tensors on
#3263 #3253	the period a reference motion is resampled onto; a Kimodo seed the sampler and the buffered-motion key disagreed about
#3269 #3259	an Isaac stage path is a USD prim path, not free text; an unknown VERA embodiment is refused rather than resolved as mimicgen
#3261 #3262	three VERA sampler knobs that reached the server as unchecked text -- one of them in only one of the two launch modes
#3257	a narrowed dashboard session window is honoured or refused, never widened
#3303 #3304 #3319	graded before it is converted: only the integer 0 means a port was not supplied; a session record's pid is not converted to a nearby process id; its start time is graded rather than subtracted from the clock
#3301 #3294 #3290	a dataset's declared task count is not converted into the absent case; a per-call Kimodo sampler knob; the legacy Isaac timestep shortcut, graded before the conversion
#3302 #3295 #3292	a de-normalization quantile and a facing goal are graded in whatever spelling they arrive; the speed a serial bus is opened at takes the shared count domain
#3347 #3350	stream_dataset refuses a flag that is not a boolean rather than inverting it; a turn goal outside the measurable heading range is refused rather than compiled

A payload that cannot be encoded is reported, not dropped

PR	What
#2638	report a payload that can never reach the wire, at a level an operator sees
#2730 #2741	coerce a sensor topic's and a safety event's readings so they reach the wire at all
#2744	a sensor record's identity is the publisher's, not the provider's
#2621 #2733	a degraded state probe names itself in the published snapshot; survey each sensor provider under its own guard
#2729 #2735	report which robot is running a policy on the state topic; report a refused rollout when a fleet stop aggregates
#2842	count a teleop frame the follower refused instead of reporting it applied

Dataset transforms, labeling + continuable refusals (new lane)

The forward-looking batch filed on 2026-08-19 shipped in full, and took its own correction pass inside the same wave.

PR	What
#2480	episode augmentation as a dataset-transform surface -- LeRobotDataset in/out, provenance-marked, verdict-gated
#2486	VLM judge agent for recorded-episode labeling on top of deterministic verdicts
#2508	07_data_augmentation.ipynb -- three shipped mechanisms, CPU-only
#2610 #2612 #2616 #2617 #2618	the correction pass: resolve a runtime-registered transform through the public resolver, record every read a verdict can make, refuse a provenance record that cannot answer, hold every determinism-key input to its domain, report a failed probe instead of raising it
#2623	a continuable refusal carries a stable code and its subject (graded by #2634/#2636/#2641)

Refuse a value that cannot be honored (new lane -- the band's dominant shape)

The #2837 -> #2879 band's own idiom, at seven sites: a non-finite or out-of-domain number was accepted, carried through arithmetic that cannot express it, and reported as a result. An all-NaN configuration returned as a successful solve is the clearest case -- nothing raised, and the caller could not tell.

PR	What
#2869	refuse an IK knob that cannot be honored instead of solving with it
#2875 #2879	refuse a non-finite pose or seed instead of solving with it; refuse a non-finite configuration instead of reporting a pose for it (closes #2873)
#2876	read the solve premise off the returned configuration rather than the fake's own storage
#2846	refuse a non-finite direction goal and name the key it came from
#2838 #2845	refuse a non-finite mesh scale where it is measured, then wherever it is carried (also in the asset-reader table above)

Its own remainder is open as PRs #2881 (a wrong-shaped pose or seed, which is a different refusal from a non-finite one) and #2882/#2883 on the Microduck provider.

Docs name what the module actually reads (new lane)

Five PRs with one shape: a page documented the extra that installs a feature, or the credential a link carries, without naming the variable that selects it -- so an operator could set everything the docs mention and still not turn the thing on.

PR	What
#2864 #2867	name the transport selector STRANDS_MESH_BACKEND, not only the extra that installs it
#2870 #2871	name the IoT credentials the transport reads, and the audit-log environment variables the module reads
#2857	the policy catalogue named an install extra that cannot exist

An asset path resolves without fetching the corpus (new lane)

PR	What
#2851	let a caller resolve an asset path without fetching the corpus
#2868	let the directory resolver decline the fetch it cannot use

A remote-inference connection is discarded when its exchange did not complete (new lane)

PR	What
#2843	hold the wire lock across the connect replay
#2854	discard a connection whose exchange did not complete, rather than reusing it

Reference SO-101 pick

PR	What
#2685	a reference SO-101 pick that lifts the cube -- closes the long-open #2145, which held that grasp quality has three conditions and the library exposed none of them

An agent-facing Unitree G1 verb surface (81 PRs -- the largest single theme in the wave, since consolidated)

The neon-the-g1 deployment's operator verbs were ported into strands_robots/tools/g1 across 81 PRs, peaking at 79 modules. Three shapes, one contract: an agent gets the vocabulary, the envelope that bounds it, and a verb that refuses a value outside it. The port has since been consolidated to 22 modules -- #3037 folded the 92 lookup verbs into one use_unitree dispatcher and #3070 folded 13 per-verb execution modules into a table-driven g1_actions -- so the PR table below records how each verb arrived, not the file that carries it today.

PR	What
#2888 #2916	the motion-switcher decoder for the FSM id producer, then wiring that producer to the arm-SDK gate -- together these answered #2765's five rt/lowcmd wire-format decisions, and it closed against them
#2932 #2933	surface the driver's joint map, and its motion-gate admission sets, to agents
#2934 #2938 #2939 #2941 #2943	live-handle read verbs -- g1_get_state, g1_battery, g1_imu, g1_lidar_state, g1_lidar_summary
#2947 #2949	cache rt/mainboardstate + g1_mainboard; cache rt/pressuresensorstate + g1_pressure
#2955 #2956 #3016 #3004 #3005	the task and command verbs -- g1_get_task_status, g1_stop_task, g1_start_task, g1_send_action, g1_run_policy
#3025 #3031 #3032 #3033 #3034 #3035 #3041 #3055	posture and locomotion verbs -- g1_set_fsm, g1_set_stand_height, g1_set_swing_height, g1_balance_stand, g1_release_arm, g1_stop_move, g1_wave_hand_loco, g1_move_velocity
#3044 #3048 #3052	the safe posture transitions -- squat-to-stand, lie-to-stand, stand-to-squat
#2959 #2960 #2961 #2962	lookup pairs for arm actions, FSM targets, loco tasks and error codes
#2965 #2966 #2970 #2972 #2973 #2974	envelope pairs -- velocity, stand height, swing height, locomotion duration, walk-forward, turn
#2967 #2971 #2975 #2976 #2978 #2980 #2987	balance modes, arm-ready mode machines, damp transitions, shake-hand stages, wave-hand turn flags, joint-variant notes, safe-posture FSM gates
#2969 #2982 #2984 #2986 #3002 #3020 #3021	the DDS families -- topics and roles, dangerous publish topics, categories, IDL types, max-buffer envelope, topic descriptions, BMS state topics
#2968 #2983 #2988 #2989 #2991 #2995 #2996 #2998 #2999 #3015 #3019 #3023 #3036	the audio families -- volume envelope, voice providers, bidi-audio dependencies / mic keywords / sample rates / stream delay / frame size, capture-rate candidates, speak VAD and silence-duration envelopes, ASR payload keys, audio call API ids, speak actions
#2992 #2994 #3001	loco call API ids, LiDAR switch states, LiDAR max-points envelope
#3000 #3006 #3008 #3010 #3011 #3013 #3014 #3026 #3027 #3029	the SLAM family -- map-name rules, relocalize, ICP fitness, cloud range, map liveness, save, voxel dedup, pose history, frame queue, map-chunk compaction
#2946 #2948 #2950 #2953	the refusal run -- a short-form role names a member rather than a call expression; a live-handle verb, g1_get_state, and g1_battery/g1_mainboard each refuse a wrong driver handle
#2997 #3028 #2951	fail the live-handle sweep when its own discovery narrows; drop a stale deferral once its blocker landed; release the subscribers a failed bring-up already built

The port filed its own remainder rather than leaving it implicit: #2891 (the FSM read cadence is undecided, and that is what now blocks the send_action wire step -- not the write-frame decisions #2765 carried) and #2928 (the tools-bundle port has no seam decided). Both have since been narrowed in practice by the consolidation: #3037 merged, folding 92 of the lookup verbs into one use_unitree dispatcher (-34,302 LOC), and #3070 folded 13 per-verb execution modules into a table-driven g1_actions (-8,571).

Net-negative refactors (13 merged, 1 refused) -- and now they reach public surface

Every band before #2880 -> #3057 added surface. Seven reductions have now merged -- roughly 101,000 lines removed -- and the one that was stopped was stopped by an argument about public surface rather than by a conflict, which is why the argument mattered: it produced a ruling, and the ruling is what licensed the two removals in the #3352 -> #3406 band.

PR	What
#3056	remove the vendored LIBERO benchmark adapter, suite and drivers -- 96 files, +336 / -33,213; strands_robots/benchmarks/ no longer exists
#3037	consolidate 92 g1 lookup verbs into the use_unitree dispatcher -- 108 files, +1,303 / -34,302
#3070	consolidate 13 per-verb g1 execution modules into one table-driven g1_actions -- 30 files, +761 / -8,571
#3389	refactor(policies)! -- drop the vera provider, git-only and the last reason for the gymnasium resolution fork -- 132 files, +424 / -15,204; uv.lock 258 -> 237 packages, the five [tool.uv] conflicts rows gone, extras 34 -> 33
#3388	move the pull-request triage tools to .github/scripts/ and their replay suites out of tests/ -- 16 files, +110 / -4,425
#3391	the training core states the domain it checks, not the history of the check -- 5 files, +483 / -2,798
#3386	refactor(tools)! -- drop lerobot_calibrate; calibrating is LeRobot's own CLI -- 22 files, +227 / -2,313
#3426	remove the augmentation-transforms package -- strands_robots.transforms (1,943 lines) had no importer anywhere in the package; LeRobotDataset(image_transforms=...) is the supported path -- 32 files, +29 / -6,064
#3419	drop the motionbricks provider -- the NVlabs model is not on PyPI, so the extra installed only support libraries while the model had to be cloned and pip install -e'd, and nothing imported it -- 58 files, +163 / -4,583
#3411	stop porting MolmoAct2's normalization -- lerobot owns it -- 12 files, +232 / -2,179
#3443	the Training page reads in one sitting; the reward wrappers go -- 10 files, +243 / -960
#3446	the episode reader joins its verifier; streaming trusts its floor -- 27 files, +537 / -848
#3459	the two mesh safety subscribers share one envelope path and one audit wrapper -- 7 files, +615 / -844
#3065 #3089 #3384	drop the docstring roles citing the removed lookup modules; a failed use_unitree call still reports what it was; resolve the body name a spec wrote rather than a synthesized <name>_main
#3047	remove the LIBERO adapter and the GR00T provider (276 files, -50,835) -- closed unmerged 2026-08-31T19:10Z

Why #3047 was refused, and how that was settled. #3075 was filed one minute before it closed: the "lerobot ships it upstream" rationale is true of GR00T's local mode (lerobot ships policies/groot, so lerobot_local covers model_path=), but service mode -- the default, which dials Isaac-GR00T's own ZMQ server -- has no upstream replacement, lerobot_async speaks a different transport, and neither the Docker-lifecycle tool nor the 27 data_configs.json embodiments have one either. Removing a documented provider whose modes have unequal upstream coverage is a breaking public-surface decision, not a cleanup. #3075 closed NOT_PLANNED on 2026-09-10T04:21:21Z with the decision recorded in words: "GR00T stays." The same comment states the direction the removals belong to -- keep GR00T and MolmoAct, drop lerobot_async, curobo, moveit2 and motionbricks -- so a provider's fate is now a written ruling rather than an open question, and #3389 is the first of the four to land.

A removed surface takes its own history with it (new lane)

Two removals in one band, both !, and both answer the objection #3047 could not: the surface being dropped has an upstream owner that ships the capability, and the removal states which one.

PR	What
#3389	the vera provider goes -- distributed only as a git repository, so it could never declare a dependency on the package it launched as a subprocess (PyPI rejects VCS metadata); its install column read _(none - git-only)_, and it was the sole reason for gymnasium==0.29.1, gym-pusht, robosuite and a five-row resolution fork. discover_ee_frame stays in simulation.ik with its three-rung ladder pin moved to the shared home; three shared guards were re-anchored on the surviving providers rather than deleted; create_policy("vera") now raises the registry's ordinary unknown-provider error with no migration shim
#3386	lerobot_calibrate goes -- 859 LOC that listed, viewed, backed up and "analyzed" the JSON LeRobot writes, while recording a calibration means moving a physical arm by hand, which LeRobot ships as lerobot-find-port / lerobot-setup-motors / lerobot-calibrate / lerobot-find-joint-limits, all four present in the version [lerobot] pins. It also removed logging.basicConfig() from all four lerobot_* tool modules -- reaching a tool is an import, so that call took over the root logger of whatever process imported it -- and pinned the two remaining import-time mkdir() calls by name

| #3419 | the motionbricks provider goes -- NVlabs' GR00T-WholeBodyControl motionbricks/ is distributed only as a source checkout, so the [motionbricks] extra installed the PyPI support libraries its model code imports while the model itself had to be cloned, pip install -e'd and have its checkpoints pulled through git-LFS. Nothing in the package imported it. create_policy("motionbricks") raises the registry's ordinary unknown-provider error, with no shim -- and it is the first of the four providers #3075's ruling named as going |
| #3426 | the augmentation-transforms package goes -- 1,943 lines with no importer anywhere in strands_robots; only its own tests, one docs page and one notebook reached it, and LeRobotDataset(image_transforms=...) is LeRobot's supported augmentation path. Deleted with tests/transforms, tests_integ/transforms, docs/data/transforms.md, its nav entry, the 07_data_augmentation notebook and four cross-references in other tests |

What the second pair did not do is mark itself. #3389 and #3386 both carried !; #3419 and #3426 carry neither ! nor a BREAKING CHANGE trailer, and no commit in #3407 -> #3488 does. Two documented public surfaces therefore left main with nothing in the commit metadata to say so, which pushes the work onto whoever assembles the tag's removals section by hand.

Both of the first pair's removals deleted their own changelog.d fragments, 23 of them for vera, matching how 3056-remove-vendored-libero-benchmark.md handled the LIBERO ones -- which is why the fragment count moved by less than the PR count for the first time in the wave.

A posture flag is checked, so a word for "no" cannot select the other branch

The lane #2473/#2487/#2475/#2485 opened, finished across the surfaces an agent actually calls. "false", "no" and "off" are all truthy strings, so a flag read by truthiness silently selects the opposite of the posture the caller asked for -- and on one of these the opposite posture destroys data.

PR	What
#3355	lerobot_train refuses a posture that is not a boolean instead of reading it as its opposite
#3360	a calibration restore refuses an overwrite posture it can only misread
#3364	lerobot_camera refuses a camera posture it cannot read
#3365	pose_tool's smooth is checked, so a word for "no" cannot change the trajectory
#3367	resume and streaming are checked, so a word for "no" cannot select the other posture
#3368	the rollout facades check their posture flags (closes #3356)
#3369	the field-scoped guard sweep discovers a guard that calls the shared read rule directly

#3356 enumerated seven such tools and named lerobot_calibrate's overwrite as the destructive one. That row resolved by deletion rather than by a guard: #3386 removed the tool, so #3361 -- the pull request adding its check -- closed unmerged with its subject gone.

An operator is asked before a command reaches a real actuator (new lane)

The dashboard's HITL gate (#3184) established that an agent tool call which would move real hardware pauses for a human. Four agent-callable surfaces were still dispatching straight through, and each of them writes to a physical device.

PR	What
#3418	use_unitree asks an operator before a mutative RPC reaches the robot
#3420	serial_tool asks an operator before a write reaches the servo bus
#3430	pose_tool asks an operator before a goal position reaches the arm
#3431	the hardware Robot tool asks an operator before execute / start reach real actuators
#3403	the other end of the same contract: a consent request is built from a refusal's code, never its prose, so what the operator is asked cannot drift from what the surface refused
#3508	the lane widened from a named list to a default: use_unitree gates every verb it does not know to be a read, so a verb added later is gated until it is declared safe

A value that can never be honored is refused at the entry point (new lane -- the band's dominant fix shape)

Thirteen fixes, one shape, and it is the next question after the shared-value-domain sweep. Those values were out of domain; these are in domain -- finite, in range, correctly typed, past every existing guard -- and still cannot produce what the caller asked for. So the surface accepted them, did nothing, and reported a result.

PR	What
#3502	the sharpest of them: a benchmark success, failure or dense_reward clause naming an entity that cannot exist compiles, never fires, and evaluate_benchmark publishes success_rate: 0.0 beside success_measured: true -- a number certifying a success it could never observe, and on a failure clause the same defect reads as 100% success
#3497	a stop_when base predicate on a fixed-base robot can never fire -- 29 of the 61 constructible sim robots report no floating base, so base_tipped on an SO-100 is permanently False
#3514	a predicate-DSL entity name that cannot name anything is refused rather than resolving to a constant
#3501	an on_frame hook that is not callable is refused before the rollout starts, not at the first frame
#3511	a policy_object the rollout cannot drive is refused at the entry point
#3494	start_recording refuses a camera it cannot render, rather than opening a dataset it will not fill
#3510	a terrain resolution the requested kind cannot draw is refused
#3516	a pose too large to be a joint coordinate is refused instead of being written
#3492	a policy provider start_task cannot build is refused in an envelope, not raised past the tool boundary
#3493	an unknown Reachy move library is refused instead of collapsing to dances
#3504 #3506	a tuning strategy the backend cannot forward is refused -- Cosmos3, then GR00T, where expert_only was accepted and reached no flag, silently full-finetuning the backbone while reporting success; pinned by #3507
#3517	a posture flag is checked, not coerced with bool(), so "false" cannot select the permissive branch and move a joint

A resource a failed bring-up adopted is released (new lane)

The counterpart of the verified-halt lane, and the question it did not ask. That lane asked whether a stop had stopped; this asks what frees a thing whose start never finished -- because the success path is usually the only place the resource is adopted, so a failed bring-up leaves it running with nothing holding a reference to it. Every one of the five is a leak that no later verb can reach.

PR	What
#3530 #3532	both routes out of init_device_connect_sync: a bring-up that raises, and one that returns no runtime, each left an event loop parked in run_forever on an unreachable thread -- and the second route reported None against a non-optional return, so the caller's own status line sent an operator to a warning that was never written
#3526	Reachy teardown joins the loop thread and closes the loop it previously only stopped -- loop.stop() is not the counterpart of new_event_loop(), and close() on a running loop raises, so the join is what makes the close possible
#3522	a link handshake the driver abandons is released, rather than left subscribed for the life of the process
#3519	leaving the mesh waits for the loops that publish on it -- the roster of loop threads had four writers and no reader, so stop() returned while nine loops were still live and one published after the session was released

A posture flag is checked, not read by truthiness -- the lane reaches every policy provider

The sweep #2473/#2487/#2475/#2485 opened and #3355 -> #3369 carried through the agent-callable tools, finished across the providers, the trainers, the recorder and the sim facade. "false", "no" and "off" are all truthy strings and 0 is a falsy int, so a flag read by truthiness acts on the opposite of the posture the caller chose -- and on two of these the truthy branch is the permissive one.

PR	What
#3538	lerobot_local's strict_keys and cache_model are checked
#3543	GR00T's strict and strict_keys are checked (closes #3540)
#3546	Kimodo's trust_remote_code is checked -- the truthy branch is the one that executes checkpoint code
#3548	the three RL posture flags are checked
#3549	dataset_recorder's strict is checked
#3553	create_world's ground_plane is checked
#3554	add_object's is_static is checked
#3574	WBC's allow_missing_models is checked -- the truthy branch proceeds with models it could not find
#3575	gr00t_inference's posture flags are checked
#3582	the read-side twin: a rover's sensor summary read a headlamp by truthiness, so a field the frame did not carry and a value the write door had already refused both rendered as off

An action naming a key the surface cannot carry is refused (new lane)

One level below #3213. That fix refused an action verb a driver does not declare; these refuse a key inside the payload of a declared verb. Each key had been silently dropped, so a command naming a joint, axis or intent the robot does not have was answered as though it had been applied.

PR	What
#3534	a Crazyflie action naming a key no setpoint carries is refused
#3537	a Reachy action naming a key no axis answers is refused
#3539	a Microduck action naming a key no intent carries is refused

A name is held to the alphabet its consumers can carry (new lane)

The containment question #3296/#3298 asked of a filesystem join, asked of a name. A camera name is interpolated into an MJCF attribute, a dataset column and an MQTT topic, each with its own grammar, and none of the doors that accept one graded it.

PR	What
#3559	a camera's name is a bare token at every door on the hardware Robot that accepts one
#3565	add_camera holds a camera's name to the alphabet its consumers can carry (closes #3562)
#3570	the recording docs name that alphabet, and every documented name is graded against it
#3557	a shadow-topic segment must be a single MQTT topic level
#3558	only a pointer topic gets the camera-frame drop exemption, so a real frame topic cannot inherit it

A stop grades the verdict it was given (the verified-halt lane reaches robot_mesh)

The lane has asked since the #2750 -> #2836 band whether a stop had actually stopped. These three are the case where the caller held the answer and did not read it: every route reported success and audited ok=True over a delivered refusal, a timeout and a lockout alike.

PR	What
#3560	Device Connect emergency_stop grades each target's stop verdict
#3566	the single-target Device Connect stop grades the target's verdict (closes #3561)
#3571	the mesh-path stop grades what the envelope send returned (closes #3568)

One camera-routing precedence, honoured on every router that implements it

The lerobot-local page states a three-rung ladder -- camera_key_map when given, then the embodiment's obs_rename, then a name heuristic. Four functions implement rung one, and three of them did not.

PR	What
#3545	an explicit camera_key_map outranks another camera's exact-name match
#3550	the sim's implicit free view does not take a declared image slot -- it had been seated first and a real camera dropped to make room
#3555	camera_key_map routes cameras on the declarative embodiment path, where an embodiment= caller's map was discarded outright and the model received no image input at all

A private name must have a reader (new lane)

Two cross-cutting invariant fixes, each deleting state that is written and never read, and adding the grader that keeps the next one out. Both are net-positive, so the net-negative refactor table above is unchanged.

PR	What
#3580	private instance state must have a reader -- six write-only attributes across seven files, with a ledger naming the two legitimate exceptions
#3583	the same rule for a private module-level name

A run that commanded nothing is refused, not scored (new lane -- the band's dominant shape)

The outcome form of the #3492 -> #3517 lane. That one refused a value that could never be honored at the entry point; these five refuse the result. Every argument passed every door, the run then applied nothing or the dataset recorded nothing, and a number was published over it -- so a caller could not tell a working rollout from an empty one.

PR	What
#3620	the sharpest: evaluate and _evaluate_with_spec tolerated an empty action chunk on every step and still published success_rate, avg_reward and pass_hat_k under status="success" for a rollout in which send_action was never called. steps_advanced and actions_applied are now separate numbers, and a run that commanded nothing is an error naming its own cause -- the rule two sibling consumers of the same condition already applied
#3612	replay refuses an episode whose frames carry no recorded action -- one frame with no action is tolerated so it keeps its time slice, but an episode where no frame carried one returned text byte-identical to a real replay; the payload now reports frames_with_action beside frames_applied
#3613	stop_recording reports the frames a strict=False recorder dropped. Partial loss is reported, total loss is refused by its real cause -- an all-frames-dropped episode had hit the empty-dataset refusal, which prescribed exactly the start_recording -> run_policy -> stop_recording recipe the caller had just followed
#3608	a declared image column that no frame fills is refused by name. The recorder stripped an undeclared camera to avoid LeRobot's Extra features error and assumed the symmetric Missing features branch away -- so a near miss on one of three camera names recorded zero frames, silently, and a total mismatch warned about a consequence that was not the one that happened
#3584	an action column present as None is refused, not recorded as 0.0 -- a placeholder is a reading nobody took

A shutdown that cannot complete does not keep the resource (new lane)

The #3519 -> #3532 lane asked what frees a thing whose start never finished. This asks the other half: what happens when the stop raises. The resource is released either way, because the caller who asked for the teardown has no second verb to reach it with.

PR	What
#3603	a node that will not be destroyed no longer holds the rclpy context -- the shutdown is reported, and the context released, rather than left initialized for the life of the process
#3602	a ROS 2 bridge that will not shut down cannot hold the devices, so a hardware Robot teardown completes even when the bridge's own stop fails
#3590	the ros2_bridge flag is graded before the rclpy probe, so an unusable posture is refused by the posture rather than diagnosed as a missing distro
#3596	the test half: four cells asserting an rclpy ImportError assumed rclpy absent instead of establishing it, so on any host with a distro sourced they ran past the probe into rclpy.init() and create_node() and left the context live for the rest of the session

An offline snapshot names what the live registry registers (new lane)

The dependency-floor lane's twin. That one asked whether a declared bound was high enough; this asks whether a copy of the library's own vocabulary is still true. training/lerobot keeps _FALLBACK snapshots for hosts where lerobot is absent, and a snapshot cannot tell you it has gone stale.

PR	What
#3616	the offline policy-type snapshot names every type lerobot registers -- it was nine types behind, and its own test graded the constant against itself, so no drift could ever fail it; the pin is now snapshot-against-live plus a per-capability subset invariant that needs no lerobot installed
#3614	sample_weighting accepts the fields its config declares -- one field list written down three times, whose consumers disagreed: build_config refused an unlisted key and build_command silently dropped it, so a typo either raised past the preflight or trained with the default under status="success"
#3618	a reward type that does not resolve gets no claim about its fields -- the gate refused a non-native type and then graded the caller's keys against the offline fallback for the type it had just said does not exist, naming three fields where the real one has 27
#3623	the third member of the lane, and the same misdirection shape one file over: a policy type that does not resolve gets no claim about its capabilities

A door refuses a collaborator it cannot use (new lane)

PR	What
#3617	a robot_name the host cannot route to is refused at the door -- the loop read it every tick while the session had already reported itself started, and the refusal sits ahead of the teardown of any stream already registered under that key
#3619	every door that accepts a teleoperator refuses one it cannot poll -- one surface already checked get_action, two did not, so an unpollable device got a topic, a running status and a loop burning AttributeError at 200 Hz

A criterion already satisfied at reset is reported, not read as a result (new lane -- the band's dominant fix shape)

The third and sharpest reading of one question. #3492 -> #3517 refused a value that could never be honored at the door; #3584 -> #3620 refused a result the run never produced. These five ask it one step earlier: what does a surface publish when the scene itself already satisfies the thing being measured, before the policy has acted? Every one of them is well-formed input, past every guard, producing a number the actor did not earn.

PR	What
#3629	the sharpest, and the one with the widest consequence: measure_agreement's two fractions are accuracies over a class-imbalanced single-label column, so a judge that answered a constant scored the holdout's own class balance -- 0.95 with one episode tagged in twenty, 1.00 with none -- and nothing in the payload distinguished it from a judge that had looked. The figure that decides whether a judge may filter training data was unreadable off the figure; it now reports the baseline its fractions must beat, computed over the compared population rather than the whole input
#3633	an episode already successful at reset is counted, not silently averaged in -- a rollout that commanded nothing scored success_rate: 1.0 with avg_steps: 1, the mirror of the success_measured guard that already warned about the hard 0.0
#3637	a stop_when clause already true at reset is reported, not read as an achievement -- the rollout stopped on step 1 with nothing moved and reported stopped_reason="predicate", indistinguishable from a goal reached
#3642	the same defect on a benchmark failure clause, which is the mirror case: a clause the spawn already trips marks every episode failed at step 1, and because the loop samples is_failure before is_success it discards a success the run had already earned
#3639	the RL side: evaluate() says whether its success_rate measured the policy at all, rather than publishing a figure over episodes whose outcome was never determined

A plan that commands nothing is a failure, not a plan

The #3584 -> #3620 lane read at the producer rather than the rollout. A planner returned a trajectory, the runtime accepted it, and the arm was sent nothing.

PR	What
#3652	a MoveIt2 plan that commands nothing is a planning failure -- a positionless waypoint row is refused where it had been unpacked into an empty command
#3657	a cuRobo plan whose waypoints do not all describe the same joints is refused. The joint-key list was resolved once, from the first waypoint, and then paired with every later one under a non-strict zip, so a narrower row downstream was filled from the wrong names -- and which behaviour you got depended on action_horizon, a streaming parameter that is no part of the plan. Graded at cache time, not at serve time, so the refusal lands before the first chunk runs
#3655	an evaluation whose every action commands nothing is refused rather than scored
#3643	run_policy stops on an episode the recorder could not flush instead of continuing over it

A ROS 2 name must be able to name a topic, and a refused call leaves no node behind (new lane)

Two halves of one surface, and the oldest interop lane in the package. The naming half is the #3557 -> #3570 alphabet question asked of the ROS 2 graph; the lifecycle half is the #3519 -> #3532 release question asked of a node.

PR	What
#3636	command_robot_name is graded against the topic grammar it is interpolated into, on both hardware bridges
#3638	a publisher is cached under its topic, not under the caller's spelling of it, so two spellings of one topic cannot hold two publishers
#3648	a published JointState names the joints its positions belong to
#3628	joint_limits is answered before the node exists, rather than requiring one to be built first
#3635	a refused mesh= leaves no ROS 2 node behind
#3651	list_topics reports every topic rosapi named, not the subset one call returned

A reader grades the size its own payload declares (new lane)

The sizes-a-read question the Modbus and DDS codecs already answered, asked of a file format and a dataset file.

PR	What
#3658	an .spz header count must account for its payload. num_points is a signed int32 that sizes all five body blocks and was graded nowhere, while the magic and version beside it were graded on the very next lines; a count short of the payload read each block's declared prefix, so every later block started at the wrong offset and the whole scene decoded to garbage and was reported as loaded. NumPy treats any negative count as "the rest of the buffer", so the under-declaring direction degraded silently rather than raising. The count now reconciles with the payload's exact byte length, validated against the shipped corpus scene the library itself references
#3650	a Robotiq reply is read by a length the codec has graded, not by the length the frame claims
#3627	a dataset metadata file that holds no JSON object is answered, not raised through the caller

An unreadable reply names the peer, not the codec (new lane -- the band's dominant fix shape)

Four clients, one shape: the peer's response was decoded inside a try whose except reported a codec failure, so a malformed, truncated or empty reply from the other end of the wire was answered as a local decoding bug. The caller is then sent to the wrong side of the connection to debug it, which on a remote-inference path means reading your own msgpack while the server is the thing that is broken.

PR	What
#3670	a frame the inference client cannot read names the peer, not the codec
#3672	the same on the Cosmos3 client
#3676	the same on the MoveIt2 client
#3677	the same on the GR00T service client

A configuration entry that is not a string is refused (new lane)

The posture-flag lane read on the value side rather than the boolean side. A credential, a name-list entry or a response body arrives as None, a number or a nested object, passes every existing guard because only its presence was checked, is stored, and is then dropped by whatever finally consumes it -- so the surface reports a configuration it is not running.

PR	What
#3666	a DDS Security credential must be a string the participant can carry -- the one with a tracker item behind it: _validate_dds_security_config accepted a None credential that the participant QoS then dropped, so a run believed it was authenticated when it was not (closes #3664)
#3667	a dashboard settings list entry is a string or the list is refused
#3674	a mesh ACL cert_common_names entry is a non-empty string or the ACL is refused
#3679	a Reachy daemon body that is not an object is refused, not read as one

A motion that did not reach names what stopped it (new lane)

The verified-halt lane asked whether a stop had actually stopped. This asks the same question of a move: the servo was commanded, it did not arrive, and the surface reported the pose it was asked for rather than the one it reached.

PR	What
#3733	a move_to that did not reach names the contact or the joint limit that stopped the servo
#3705	set_gripper says what the fingers closed on
#3703	attach_bodies reports whether the bodies actually touch
#3727	the feature underneath all three: get_robot_state says which way the arm extends from its base

#3732 is the instructive part of this lane. It was the feat-shaped version of the same idea -- "a not-reached move_to or rotate_wrist names what stopped the servo" -- and it was closed unmerged in favour of the narrower fix #3733 plus a tracker item, #3734, for the rotate_wrist half. That is a descope recorded rather than dropped, and it is why the rotate_wrist row is an open issue rather than a silent gap.

A camera name answers for the camera it names

The surface that has produced more naming defects than any other, four more in one band: a name the sim will list must be a name the sim will accept, once each, at every door.

PR	What
#3683	every camera name the sim lists can be rendered through it
#3693	a camera enumeration offers each camera once, not once per name
#3691	add_camera / remove_camera accept camera_name for name, as every sibling door does
#3730	start_recording names the cameras by a name that answers for them

The 0000- fragment placeholder defect is closed (sequencing item 6, retired)

Recorded on this page for three syncs as unactioned, and fixed as a class rather than as four filenames. The old guard permitted the placeholder because it only rejected a number at or above 9000.

PR	What
#3715	a 0000 fragment is a placeholder and is refused
#3721	the placeholder-number rule has one owner, so every surface refuses it -- the ## One owner answers one question shape applied to the repository's own bookkeeping
#3722	the one fragment that named no pull request now names the one that carries it

changelog.d/ holds zero 0000- fragments for the first time since they appeared, and --check reports OK (1600 pending).

The dashboard gets a server

strands_robots/dashboard/ had spent nine bands accumulating rails -- auth, access, consent, safety_state, settings, log_redaction, validate_scope, agent_hitl, task_timeout, ttl_cache, agent_motion -- with nothing serving them. This is the slice the #2848 decomposition argument was for.

PR	What
#3695	the server core, wired to the landed rails, fails closed (19 files, +1,609 / -74) -- server.py, access.py, routes_auth.py, cli.py and a served static/ bundle, taking the package eleven -> fifteen modules
#3710	a caller-supplied value cannot forge a log entry
#3738	a credential store that parses but cannot serve as one is recovered, not answered with a 500

#3122 is still open and still names what the server slice owes: both enrollment routes skipping the auth middleware, and two directory reads validating existence rather than containment. #3695 landing a server makes that issue sharper, not softer.

Seven features end a nine-band featureless run

Six of the seven are one theme -- the library says more about the thing it just handed you.

PR	What
#3695	the dashboard server core (above)
#3728	tool_frame declares the tool point a site-less model lacks
#3724	step feeds an open dataset recording, so a hand-stepped scene records like a rollout
#3694	SO-arm joints can be addressed by label rather than by index
#3727	get_robot_state says which way the arm extends from its base
#3731	Robot(tool_name=) names the tool, so two robots of one type can share one Agent
#3697	render_dir on the constructor moves the render sandbox per Simulation

A surface offers a next step the thing it just made can take

PR	What
#3686	add_robot offers a next step the model it added can take -- it had advertised run_policy for a model with zero actuators (closes #3684)
#3687	the sim tool description describes the world the agent is joining, not a generic one
#3719	a paused real-robot call tells the script how to resume it
#3742	an unreachable GR00T server fails fast and names the remedy
#3741	run_policy and eval_policy are refused while another rollout already drives the robot
#3680	the #3584 -> #3620 lane at the rollout entry point itself: a rollout that never commanded the robot is an error

A refusal applies nothing from its own call

PR	What
#3645	a refused resize applies nothing from its own call. set_geom_properties persisted colour, friction and size up front, then re-derived the body inertial from a compile of the persisted spec; a shrink below the inertia floor does not compile, and the rollback restored only the property the refusal named -- so a call returning status="error" left a colour and a friction the caller never asked to keep, against a docstring promising the model is "left untouched". Fixed by ordering rather than by undoing: every model write now sits below the last thing that can refuse
#3622	removing a robot no longer drops a loaded scene
#3625	a cameras= no native driver will open is refused, not dropped

The largest test-pin band on this page (11 pins)

Eleven of the band's 39 commits are test, the most any band here has carried, and each grades a door an existing suite reached only through a stand-in -- so the code under them was exercised without its answers being pinned. None owes a changelog fragment, which is why 39 PRs produced 28.

PR	What
#3641 #3647 #3632	one wire field, every door that writes it: the Feetech domain, the Dynamixel two-byte frame-field boundaries from both sides, and every Modbus field the Robotiq codec refuses
#3644	the IDL-class resolver both Unitree drivers report an SDK drift through -- a function two sibling drivers carry identically, whose reason nobody had ever produced, with a cell asserting the twins answer the same
#3660	every door on the Franka shutdown path -- the teardown-absorb arms a fake that fails only on close() reaches
#3646 #3659	the MuJoCo rebuild state whose handle does not survive, and every reason a disagreeing spec is refused
#3634	every agent verb the Robotiq gripper schema declares
#3621 #3626	a raising UR controller is reported, never propagated; a Go2 frame that never reached the wire is named
#3656	the twin of every answer a Feetech cell already reads

An evaluation reports how close it came and how often it holds

PR	What
#3521	the band's one feature: pass_hat_k and max_step_reward on the evaluate payload (+380 / -3 across 3 files) -- a run reports the probability that a k-attempt sample succeeds, by the subset form C(c,k)/C(n,k), rather than only a mean that cannot separate a consistent policy from an erratic one
#3528	the same figure on the other route: evaluate early-returns to _evaluate_with_spec for a spec and runs the success_fn path itself, so only the spec payload carried the new keys and eval_policy(success_fn=...) reported a success rate with no reliability figure beside it
#3520	the docs half -- the notebook series can score the policy it just trained (+808 / -0)

The docs state each contract once (new lane)

Seven docs PRs, every one net-negative, together -3,575 lines. Three of them are the same defect this page records everywhere else -- a page describing a surface the code does not have.

PR	What
#3469	README is 120 lines; configuration, the numbers and the whole-loop example move into docs/ (-1,459)
#3472	remove the dashboard pages that document a command the package does not ship (-1,012)
#3463	validate_command states its contract by refusal shape, not rule by rule (-900)
#3471	the lerobot-local and protomotions pages state each contract once (-861)
#3440	four policy pages keep their contract facts and lose the essays (-668)
#3470	the robot-family catalog is generated from robots.json, not typed
#3473	drop a documented RuntimeError stream_dataset cannot raise

One rule, one owner, across the test tree (new lane)

The whole-tree grader audit (#3105 -> #3126) established that a rule sweeping the repository must resolve the whole tree. This band asked the next question: how many copies of the rule are there?

PR	What
#3417	"a public module, class, method or function has a docstring" was asserted by fourteen copies of one AST walk under tests/, one per package; one linter rule replaces them (-1,663)
#3468	one grader enforces the module-not-filename citation rule across the package (-709)
#3483	the user-registry suite has one owner, and an inert alias pin is deleted (-486)
#3444 #3455 #3457 #3460 #3464 #3465 #3475 #3482	the import-hygiene run beside it -- a module that reaches an optional extra at import time skips rather than errors, a fixture that blocks a dependency restores what it displaced, MUJOCO_GL has one owner under tests/, and the camera pose domain is graded in CI rather than only where newton is installed

A telemetry field the frame does not carry is not a reading (new lane)

A vendor IDL declares a field the wire may leave unset, and the decoder defaulted it to a typed zero. So an absent measurement and a real one read identically: a level IMU and a missing IMU are both rpy=[0,0,0], a missing layout id is 0, a missing lidar fault code is 0.

PR	What
#3376	the Unitree drivers agree on which telemetry fields are readings
#3379	a G1 IMU field the message does not carry is not a level reading
#3380	the lowstate layout id is a reading, not an int() cast
#3382	a lidar field the message does not carry is a None, not a fault code
#3394	rt/lowstate.motor_state is a reading, and the humanoid takes it
#3381 #3383	the Booster T1 snapshot reports an absent IMU field, and an absent frame field, as absent

A stop reports its outcome, and now so do the drivers, the sessions and the inference server

The lane opened in the #2750 -> #2836 band (eight sites returning success because a stop had been issued) is the wave's longest-running fix shape, and it has now reached the driver programme, the rollout entry point, the training-session registry and the remote inference server.

PR	What
#3060	a cooperative stop issued before a rollout's first frame halts it, instead of being discarded (closes #2833)
#3091	a UR halt reported as done is a halt the rollout thread has actually left
#3092	a camera recording reported as stopped is one whose recorder thread has exited
#3100	a remote stop reports which rollouts it halted
#3104 #3110	a session reported stopped is one whose process is confirmed gone; a session whose process cannot be inspected keeps its record rather than being dropped
#3108	a driver stop() that could not halt the robot says so
#3114	a stopped PolicyServer stops serving the clients it already has, not only refuses new ones
#3224	stop_teleoperate reached on the teleop loop's own thread still releases the devices -- the lane's original defect wearing re-entrancy as a disguise
#3341	a recorded Isaac or Newton rollout leaves the robot idle when it ends -- the release the MuJoCo backend already performed, now on the other two backends
#3338	a UR halt reaches a setpoint already in flight inside send_action
#3326	a Crazyflie flight command the radio would not carry is a refusal in an envelope, not a raise past the tool boundary
#3374	a caller-stopped G1 or Go2 rollout names what stopped it
#3390	a Robotiq halt that never reached the gripper is reported
#3392	a Reachy motion_stopped names the halt that still holds, not the last one asked for
#3588	the last Go2 sport-mode release is verified, not assumed to have failed
#3609	the Go2 write gate shuts on a mode reading that says the robot is not free
#3597	a silent policy server is reported as silent, not as absent

The whole-tree graders grade the whole tree (new lane)

A guard that sweeps the repository is only as wide as the tree it resolves. Seven PRs in one day found four spellings of the same defect: the sweep read a narrower tree than the rule it enforces, so it reported a clean result over code it never opened. Each is a false pass, which is the direction that costs the most.

PR	What
#3107	the whole-tree grader preflight derives its roster from the tree instead of a hand-maintained list (closes #3105)
#3113 #3118 #3120 #3126	the derivation resolves a grader's walk root held in a loop variable, passed as a parameter, bound inside the function, and named by a symbol or a method
#3112	the pacing sweep grades every ticker, not every ticker spelled Ticker (closes #3106)
#3119	a whole-tree guard locates the tree it grades rather than the working directory the run started in
#3274	a walk root derived from a symbol imported through a re-exporting module is refused instead of silently resolving to nothing (closes #3273)
#3291	the render-gating rule reads a camera image out of an observation, not only out of render() -- the one such assertion in the tree failed as a bare KeyError on a headless host (closes #3286)
#3288 #3289	the GL-free half of a construct-then-spend case keeps running; the GL gate on a render test has one owner
#3313 #3293	every environment variable the package reads is documented, and a grader says so; a shared numeric domain that counts its own families enumerates that many
#3324	that same env-var grader follows a module constant to the names it holds, and named three training variables it had missed
#3316	check_merge_blockers names whether auto-merge will perform the merge it is waiting on

Refuse a span that resolves to no work (new lane)

Seven rate, period and horizon domains, one shape: a value that is positive, finite, inside every existing guard -- and still produces nothing. The sharpest case is a rollout that ran zero control steps and reported success.

PR	What
#3130	a rollout duration that resolves to no control step is refused
#3123	run_multi_policy honours the step count it resolved instead of re-deriving it from the float
#3116	the Isaac preview cadence refuses an infinite period
#3121	a lerobot_train launch topology that cannot be honored is refused
#3103	a rollout runs for the wall clock it was asked for (a deadline, not a delay)
#3124 #3129	the rate guards judge every fps spelling their own domain accepts; an on-disk recording rate beyond the float range is unreadable, not an exception
#3134 #3148	a lerobot_camera capture span its rate records no frame in is refused; stop_recording accepts no destination it will not write

A cache is valid only while its source still holds its bytes (new lane)

Four surfaces in four hours, one shape at each: a parse was memoised against something weaker than the bytes it was parsed from -- a directory's mtime, one file of several, or nothing but the process's lifetime -- so a file that changed underneath the entry kept being answered from the stale copy. Each is a false hit, which is the direction that costs the most.

PR	What
#3141	the dashboard auth store's cache serves a hit only while the file still holds its bytes
#3142	a cached mesh ACL is served only while its file still holds its bytes
#3145	a cached robot registry is served only while every file it was parsed from holds its bytes
#3147	an asset mesh scan is memoised for one pass, not for the process (closes #3144)
#3173	a cached render backdrop belongs to the camera it was rendered for, not to four of its six fields
#3177	a cached qpos-address map belongs to the robot it was built for
#3220	the VERA IK bridge a rollout solves on is a bridge for the model that is bound -- the slot was keyed on nothing, and the simulation rebinds a freshly compiled model on every rollout

#3144 named the mechanism precisely, and it is worth quoting because it is the reason this lane is a defect class rather than a tidy-up: the mesh cache keyed on the top directory's mtime while answering for the whole subtree, so a download into the declared meshdir was discarded as "no meshes available" -- the cache reporting an absence it had itself created. The shape has since moved on. The first four were "the source changed underneath the entry"; #3173, #3177 and #3220 are "the key does not name the thing the entry is about" -- a backdrop keyed on four of CameraParams' six fields, an address map keyed past its robot, and a MinkIKBridge keyed on nothing at all while the engine hands the policy a newly compiled MjModel on every bind. Every one of the seven is a false hit, which is the direction that costs the most: #3220's unchanged-DOF case solves against a model that is no longer in the scene and reports an action dict shaped exactly like a good one.

An undeclared agent verb is refused, not run as the halt (new lane)

A driver's tool_spec declares the verbs an agent may plan against, and nothing between the model and the tool validated the action it sent against that schema. Eleven of the twelve shipped native drivers ended their dispatcher in a bare else whose comment named the halt, so every value the enum does not cover ran the halt and answered as though the caller's own verb had been dispatched.

PR

What

#3213

an agent action a driver does not declare is refused instead of running the halt -- a typo ("sensor"), a verb borrowed from a sibling ("sensors" on a URDriver, whose read verb is spelled state; "stop" on a CrazyflieDriver, whose halt is spelled land) and a non-string all stopped the robot under a success-shaped answer. drivers.base gains declared_verbs + undeclared_verb_error as the single owner of that list and that refusal, read back off the schema the agent planned against; the refusal names every declared verb so an agent can correct itself in the same turn (17 files, +378 / -41)

The non-uniform vocabulary across the family is what made the confusion likely rather than hypothetical -- land where a sibling says stop, state where a sibling says sensors. FeetechDriver had already refused, and its own test docstring recorded the fallthrough as the defect; this closed the other eleven against that precedent rather than proposing a new rule.

A verdict names the thing it was written for (new lane -- the band's dominant fix shape)

Six PRs, one shape: a surface answered a question truthfully about some other subject. A pid the kernel had reissued, a stand-in the code under test never consults, a pull request that has already merged, a motion call adjacent to the one a human approved -- in each case the answer was well-formed and about the wrong thing, which is why none of them looked like a failure.

PR	What
#3219	a training session's running verdict names the process the record was written for
#3237	its residue: the store's own test docstring and _NotRunning stub still described the pid-existence verdict #3110 replaced, so a finished run whose pid was reissued read as live (closes #3227)
#3250	a stand-in the prune never consults left the verdict to whichever host the suite ran on
#3233	an already-merged pull request is not a mergeability GitHub is still computing, so the report stops prescribing a re-read that can never settle (closes #3231)
#3248	an unreviewed pull request is owed by whoever can actually clear the rule, not by whoever is nearest
#3254	a one-shot HITL motion grant names the call the human approved, not a near neighbour of it
#3255	a redacted log record is still renderable by the formatter it was written for
#3268	a standing request for changes is a blocker owed by the account that made it, not by whoever is nearest

One owner answers one question (new lane)

A question with a named owner acquires a second reader that answers it differently, and the two disagree in both directions. Each of these is the same repair: delete the second reading, not reconcile it.

PR	What
#3203	"are this model's meshes on disk?" is answered by its owner -- resolve_model_path was a third reader walking the model directory for files with a mesh extension, where _mjcf_missing_meshes resolves each file= against the model's <compiler meshdir> the way MuJoCo does, across every <include>d fragment
#3208	a ROS 2 topic name is judged by one rule, and it is the mapping's own
#3211	the Device Connect state RPC reads the device that owns the bus, rather than whichever wrapper or driver was nearest
#3216	a Cosmos3 action_mapping is a rename, so two columns cannot become one -- the key side was validated and the target side was not, so a mapping could silently deliver one joint's value under another joint's name
#3209	each robot's joint_states reach that robot's own topic on the hardware RTPS bridge
#3214	Isaac's remove_robot prune is bounded at /, USD's path separator, so a robot whose name merely extends the removed one keeps its prim (with arm and arm_left both live, removing arm had dropped /World/Robots/arm_left, leaving a live robot with no prim for destroy to release)
#3252	the episode count meta/info.json declares gets one owner, so the reader and the verifier cannot disagree about it
#3221	a registry alias is keyed by the fold every lookup applies to its query, so the index and the lookup agree on what a name is
#3230	a VERA port means the same thing whichever spelling named it
#3215	the Device Connect allowlist advisory follows the transport, not one of its two sources
#3265	a recorded length of zero is a frame count, not a missing column -- so a reading is not reported as an absence
#3300	the parquet-truth counts a rollout reports are graded by their one owner
#3323	a failed dataset-stack import is diagnosed by one owner, so the refusal names the install that fixes it rather than the module that happened to raise
#3308	a robot_mesh command body is decoded by one owner, so every refusal reaches the tool envelope rather than escaping past it
#3375	every simulation backend reports the rollouts it is driving, so a peer's status and its state agree (closes #3359)
#3378	list_policies_running answers on every backend, not only the one that implemented it
#3399	describe() advertises the rollout reader its own entries name
#3404	every rollout loop acquires chunks through one pipeline and one query helper
#3600	a backend loader is a callable returning the backend class, graded once
#3604	one owner for the off-policy warmup batch relation
#3606	a composite joint group is exclusive whichever child declares it
#3610	evaluate_benchmark resolves robot_name the way its siblings do
#3624	replay_episode resolves robot_name the way its siblings do -- the same surface, the last entry point that did not

A packaging declaration reaches the install that needs it (new lane)

A module imported a distribution that no extra declared, so the documented install resolved cleanly and then failed at import -- the same class of gap v0.5.1 was cut for (#1930), read one layer down: not a floor too low, but a requirement never written.

PR	What
#3310	the session tools' process dependency is declared
#3315	the rover driver's HTTP transport is declared where an install reaches it
#3318	the UR driver's RTDE client is declared where an install reaches it

A byte that cannot be decoded is reported, not raised (new lane)

Four surfaces, one shape: text was decoded on whatever the process happened to be configured for, or strictly enough that one bad byte took the caller's call down instead of being reported.

PR	What
#3307	on-disk text states encoding="utf-8" instead of taking the process locale -- 29 files, so a file this library wrote is a file it can read on a host configured differently
#3311	a child process's stream is decoded with substitution, not strictly
#3306	a dataset metadata file that cannot be decoded is reported, not raised
#3308	a robot_mesh command body is decoded by one owner, so every refusal reaches the envelope

A store is written whole, or the previous one is still there (new lane -- the band's dominant fix shape)

Five surfaces, one defect: a document was serialized straight over the file that held the previous one, so a write that failed part way left neither the old state nor the new. The store is not merely stale afterwards -- it is gone, and on two of these five the thing lost cannot be recomputed.

PR	What
#3322	a user-registry write stores the whole overlay or leaves it untouched
#3328	a pose library is stored whole, and the tool reports that it was not rather than claiming it was
#3330	an Isaac camera recording is encoded whole, or kept to be encoded -- a failed encode no longer discards the frames
#3345	a session store is written whole, or the previous one is still there -- the store is the only place a detached process's pid is written down
#3348	a calibration is written whole, or the previous measurement is still there -- and a calibration is a measurement of a physical robot, so a lost one has to be taken again

A dataset root resolves to the directory its repo_id recorded to (new lane)

The repo_id a caller records under and the directory a reader opens had drifted apart, so a dataset written under one name was read back from another. Four entry points, closed in one band, plus the recipe that documented the wrong read order.

PR	What
#3332	a recorded dataset lands in the directory create prepared for it
#3339	an append reopens the dataset its repo_id created
#3342	a read reopens the directory the repo_id recorded to
#3346	streaming a dataset back reopens the directory the repo_id recorded to
#3344	the streaming recipe documented as chronological read frames the reader shuffles (docs)

A seeded replay reads the forwarded seed, not the process-global RNG (new lane)

A test-quality lane, and a false pass in the direction that costs the most: a cell measuring random's process-global stream is graded by whatever else in the suite advanced it, not by the seed under test.

PR	What
#3331	a seeded eval's replay is measured off a private fork of the forwarded seed (closes #3329)
#3334	the seeded-replay policy reads the forwarded seed
#3337	two more seeded-replay cells read the forwarded seed (closes #3333)
#3336	a re-imported module is put back on its parent package, not only in sys.modules -- otherwise a later import of the same name gets a different module than the one the suite patched (answers #3335)

A selector that names nothing is refused, not widened (new lane)

The caller narrowed the request and the surface widened it -- the one direction a selector must never fail in. Two surfaces found the same defect in one band.

PR	What
#3305	a Reachy Mini motor selector that names no motor is refused, not widened to every motor
#3309	a lerobot_camera batch selector that names no camera is refused, not widened to the default cameras

A path stays inside the thing that contains it

The #3214 prim-prune shape asked of a filesystem join rather than a USD path, plus the log sink beside it.

PR	What
#3296	a checkpoint's norm-stats filename stays inside the checkpoint
#3298	a stored model path is contained wherever it is joined
#3297	the seed a WBC reset renders into its log line is escaped at the sink
#3325	a symlink inside a downloaded robot tree is never followed into the asset cache

An external call states the bound its caller is waiting on

PR	What
#3312	a docker query the VERA readiness wait makes states a bound
#3314	the shared docker query bound is one a healthy daemon can meet -- a guard against the fix becoming its own failure
#3317	a preset scene download bounds every socket operation, not only the connect

A dialled address has two halves (new lane)

ws://<host>:<port> is one value in two parts, and every transport that dials one validated the port end to end while taking the host as free text -- so the half that decides where a connection goes had no domain on any of them.

PR	What
#3272	the host is the other half of a dialled address, and takes the port's own domain (the policy providers)
#3271	the same for the VERA server URI
#3279	the rosbridge host half is graded before a pattern is offered it
#3287	the host half of the Reachy Mini daemon address is graded beside its port
#3299	the EarthRover SDK base URL addresses the host the caller wrote

Read as the "one owner answers one question" shape asked of a URI rather than of a registry: one owner for what a legal address is, rather than one per component.

A recording keeps the dataset it was about to replace (new lane)

PR	What
#3196	a refused start_recording keeps the dataset it was replacing -- overwrite=True is the one posture that deletes a real LeRobotDataset without asking, and it is what the run_policy agent tool records with. Every other camera-shaped refusal was already ahead of that deletion; a single unknown name in cameras= was refused from inside the schema-declaration block, roughly a hundred lines after the wipe, so the call reported an error having already removed the dataset
#3198	a reset() closes the open dataset episode on every backend, not only the one that already did
#3245	the same ordering question one directory over: a fresh lerobot_train start clears an empty output_dir, not a directory that holds a checkpoint -- so a refused run no longer costs the caller the checkpoints it was replacing
#3200	every span lerobot_camera reports is measured on the clock that cannot step (the #2404 -> #2425 monotonic lane, one more surface)
#3199 #3207 #3210 #3218	four more shared value domains -- an RL evaluation's episode count, a served policy's advertised metadata, every RTDE read held to the axis count the command path enforces, and a WBC substep count held to the count the caller named
#3189 #3193	a served policy's required_bodies reach the host that supplies them (the wire seam #3187 did not reach); a link written as raw HTML is graded like a Markdown one (the whole-tree-grader lane, one more false pass)

A run that stopped early states how much of it never ran (new lane)

The required suite runs under -x, so a red check names one failure of an unknown number -- #3164 measured a run that aborted at 74% of 46,583 items and hid a second real failure four lines below the first. Three PRs answered it, and the third exists because the first two each derived the same number independently.

PR	What
#3167	a session that stops early states how much of the suite never ran -- computed in process from the session's own item list
#3166	a red required check says how much of the suite it never reached -- parsed from the log into the job summary and an annotation (closes #3164)
#3172	how much of the suite ran is counted once, by the session that ran it -- one owner for the subtraction, text arithmetic as the fallback
#3149	perf, the wave's only one: collect coverage on the PEP 669 core rather than the C trace function -- +8.2% against the default's +76.6%, same reading per file, statement and exclusion

#3169 is the process record of how the pair happened, and it is the rare tracker item that files a negative result: widening the duplicate-claim key to simultaneous cross-references selects about 27 pairs across 9 issues for 2 genuine ones and fires on a third of the queue, so it is boilerplate rather than a gate. The documented spelling for a deliberate second implementation (Refs #N) is exactly the spelling that hides an unintended one.

A measurement that could not be taken is reported as no reading

PR	What
#3128	an unreadable episode count is reported as no reading, not as zero (closes #3125)
#3109	an evaluation that lost a recording episode does not report a success rate over it

A Reachy Mini verb surface

PR	What
#3082	14 Reachy Mini verbs in the consolidated shape -- the arrangement #3037/#3070 arrived at, adopted on arrival
#3094 -> #3097	filed against it the same morning (the head-body yaw coupling limit was enforced only where one action carried both values, and two of the new verbs send one) and closed by #3097
#3095	a verb states the coupling limit only where its own action can reach it

Every DDS endpoint is constructed under the shared lock (new lane)

#3062 was filed against the Go2 driver the day it landed: a sport-mode gate built an RPC client outside _DDS_INIT_LOCK, racing one driver's own subscriber set into a segfault. The fix and its graders closed the whole class rather than the one site.

PR	What
#3064	open the motion-switcher client under the shared DDS lock (closes #3062)
#3067	construct the Booster T1's DDS endpoints under the shared lock
#3071	derive the DDS-lock vocabulary from every lock-holder rather than a hand-written list
#3072	grade every endpoint construction against the shared lock
#3073	grade the DDS endpoint lifetime on both bring-up and teardown

The dashboard: from a decomposition argument to the wave's feature lane

#2977 asked for the 481-file draft #2848 to be decomposed before another review round. Two slices merged, the draft itself was closed 2026-09-02T02:24Z, and the extracted package then became the place the wave spends its features. strands_robots/dashboard/ now carries ten modules on main -- agent_hitl, agent_motion, auth, log_redaction, safety_state, settings, task_timeout, ttl_cache, validate_scope and the package __init__ -- and the newest band's three features are all here.

PR	What
#3050	operator auth rail, with the extra that supplies it
#3051	settings, log redaction, task timeouts and a bounded cache
#3078	a first enrollment is limited to the machine, not just the damaged-store case
#3080	a passkey ceremony is verified against the connection, not the caller's claim
#3184	a HITL gate pauses an agent tool call that would move real hardware (+1,058 / -31 across 13 files) -- and raises the strands-agents floor to >=1.13.0 for the interrupt / cancel_tool hook members it calls
#3185	a Lockout value type states what a dashboard may say about an e-stop
#3186	a policy preflight reports what it actually verified, not just "ok"
#3579	a renewed session keeps its provenance marker
#3585	a lockout verdict is ordered on this process's own clock
#3586	a settings section given a scalar is reported, not dropped in silence
#3594	a WebAuthn expectation needs a Host header that actually arrived

#3079 (expected_origin taken from the request's own Origin header, so the WebAuthn origin check cannot fail) closed 2026-09-02T00:21Z with no closing PR, exactly as the previous sync predicted once #3080 landed that remedy. #3115 then moved the challenge TTL onto a clock a correction cannot move, and the 481-file draft #2848 was closed 2026-09-02T02:24Z -- one minute after #3122 was filed naming what it still carried. Every module this section extracted has now taken at least one correctness pass -- #3131 made the log filter redact every part a formatter renders, #3141 held the auth store's cache to the bytes it was parsed from, and #3163 stopped a setting that is not a boolean opening the remote-code gate. #3122 is still open, and still names what the draft carried: both enrollment routes skipping the auth middleware, and two directory reads validating existence rather than containment.

An EarthRover verb surface

PR	What
#3081	native EarthRover Mini+ driver over the earth-rovers-sdk HTTP API
#3084	six rover verbs in one table-driven module (+562 / -1) -- the consolidated shape #3037/#3070 arrived at, adopted from the start rather than retrofitted

The conventions that catch a dead cross-reference

#2940 recorded the concrete failure: two consecutive g1 verb ports shipped dead cross-reference roles behind a green local run, because the whole-tree graders are exactly the ones a diff-scoped selector cannot see. It closed against #3087.

PR	What
#3090	report a docstring role naming a module the other side of a merge deletes
#3087	a qualified role must not name a module that lives in an open PR (closes #2940)
#3083	convention 7's rule is recoverability, and name= is one of three shapes
#3076	a test module name must state the behaviour, not the tracker item
#3077	an absent optional dependency names the module in ImportError.name

Mesh and Feetech correctness (the #3037 -> #3091 band)

PR	What
#3066	bound each teleop joint in the unit its follower declares -- closes #2935, where a degree shoulder and a percent gripper shared one scalar bound
#3086	the teleop slew bound is sized on recorded teleop, and full scale does not size a speed
#3085	validate the command envelope's routing fields, which are key-expression segments
#3074	the two-byte word order is an STS/SMS property, so say so and decide it once
#3063	refuse a misspelled backend option instead of dropping it (closes #2895)

doctor reports a build that cannot run on this host (new lane)

Each of these previously surfaced as a failure somewhere else -- a CUDA error inside a rollout, a GL error inside a render -- rather than as a diagnosis naming the build.

PR	What
#2893	report a Warp build that cannot target this GPU's architecture
#2908	report a torch build that carries no code for this GPU
#2911	the MUJOCO_GL remedy names a backend this host can load
#2896	release the sim smoke test's renderer instead of leaving it to the finalizer

Microduck, the band after it landed

The 14-DOF Pollen biped went end to end in #2856/#2862 last band. This one is its hardening pass, plus two real additions.

PR	What
#2906	episodic behaviors in MicroduckPolicyBundle auto-return on their own timer
#2909	route observation slot two through the ONNX gravity_source flag (with #2913, raw_accel slot two is a gravity direction, not the reading)
#2882 #2883 #2887 #2890 #2898 #2905 #2919	refuse what cannot be honored -- an unchecked graph action, a scale or threshold that would silently change the command, an unreadable floating-base block, a non-finite observation component, a gate key naming no held skill, an override the method cannot honor, a goal the velocity gate cannot read
#2902 #2903	the velocity gate arbitrates between its pair instead of selecting into it; the command width follows the graph, not command_names
#2884 #2920	hold the transport knobs to the shared numeric domains; grade the degraded driver surface, not just the happy path
#2900 #2926 #2927 #2922	name the scene each shipped skill needs, the ball scene's actual contents, the stance every weight was trained in; --scene reaches the scene a skill was trained in

Environment variables are named where an operator looks (new lane)

A page can document the extra that installs a feature and the credential a link carries, and still not name the variable that switches it on. Last band opened this shape on STRANDS_MESH_BACKEND; this band finished the mesh matrix.

PR	What
#2985 #2990 #3003	STRANDS_MESH_NAMESPACE, STRANDS_MESH_POLICY_TYPE_ALLOW and STRANDS_MESH_ACCEPT_PERMISSIVE_ACL are named on the README matrix and the security page
#2979 #2945	name the audit-log family's tamper-evidence and rotation knobs; name the three TLS material paths the default posture requires
#2952 #3038	the pages describing [all] name the extras it leaves opt-in; the driver="strands" refusal example names a robot that has one

Other named surfaces

PR	What
#2095	advertise set_obs_noise so an agent can configure sensor noise
#2116	apply the pose rule once in add_camera, not twice
#2023	delete the inert VERA n_action_steps knob
#2258	state the reason, not the review thread, in 43 docstring markers
#2055 #2057	document parameters with an enforced domain but no Args: entry; root the completeness guard at the package
#2190 #2210	mesh discovery prose + camera-offload docstring stop calling multicast the default
#2844 #2840	fill the undriven state columns of a synchronized rollout too; say that the MuJoCo grow path carries the joint-force buffer it carries

The remaining ~220 PRs are the sustained fix/test-pin sweep -- now including a validate() preflight trio that refuses a val_episodes whose episode count cannot be read (#2346), a dataset format version lerobot cannot read (#2351), and streaming alongside a validation split (#2352) rather than dropping one silently; and a policy/registry resolution run (GR00T data_config -> registry robot #2350, a provider's declared host default surviving build_policy_kwargs #2356, a non-lowercase policy URL scheme #2374, processor_overrides routed to the pipeline declaring the step #2375, an exclusive CompositePolicy joint group #2385) -- -- dominated by the sim facade refusing values it cannot honor (add_object mass/color/size vectors, control_substeps, rollout duration / policy_config, create_world timestep/gravity, plain-MP4 frame and pixel counts, MJPEG stream options, dataset-recording frame rates), MuJoCo version-drift resilience (the sparse-inertia mass matrix, camera intrinsics, the API floor), and matching contract pins. All of it lands in the same open [Unreleased] section.

v0.4.1 -- shipped 2026-07-01 (published wave)

Everything below merged to main after the v0.4.0 tag and shipped in v0.4.1 (now under the dated ## [0.4.1] heading in the CHANGELOG after #977). Grouped by theme.

Security -- hardening the LLM-callable filesystem + dependency surface

PR	What
#929	sandbox + harden render(output_path=...) against .. traversal, backslash separators, shell metacharacters, symlinked targets, oversize, and partial-write corruption (STRANDS_ROBOTS_RENDER_ROOT, atomic write)
#930	extend the same output-path hardening to the sibling video sinks run_policy(video={"path": ...}) and start_cameras_recording via a shared simulation.safe_output helper (STRANDS_ROBOTS_VIDEO_ROOT)
#931	remove the unregistered mimicgen dependency (dependency-confusion RCE, CVE-pending)
#950	remove the git-only vera extra that breaks PyPI upload (release-enablement)

ROS 2 / RTPS interop (new capability)

PR	What
#742	RosBridgedRobot -- present a ROS 2 robot as a first-class strands Robot
#744	use_rtps + RtpsRobot -- act as a ROS 2 robot over pure RTPS (no rclpy)
#747	publish simulation telemetry on a ROS 2 domain (ros2_bridge flag)
#755	publish a real robot on a ROS 2 domain (ros2_bridge flag)
#748 #750	use_ros showcase + README/docs ROS 2 overview; hero-example-first README
#767	unify rclpy + RTPS hardware bridges under RosTelemetryBase (HardwareRtpsBridge) -- byte-identical topics by construction
#875	DDS Security gate + joint-limit bounds on the hardware bridges

Training -- SDK now trains AND scores end to end (incl. from-scratch RL)

PR	What
#667	stream a Hub dataset into the LeRobot trainer (no full download)
#734	wire RA-BC sample weighting into the lerobot trainer
#735	relative-action training flag
#736	DAgger correction-collection action via lerobot-rollout
#749	SARM reward-model training + the RA-BC production loop (closes the produce half)
#777	dynamic RewardModelConfig parity with lerobot -- reaches every registered reward type (SARM-only 1/4 -> 4/4)
#656	create_trainer auto-discovery fallback
#877	from-scratch RL trainers with a PPO backend (on-policy)
#886	off-policy FastSAC RL trainer
#915	RL parity remainder -- staged_reward, gym adapter, vectorized PPO, render output_path + precise env union
#918	VecSimEnv -- N independent SimEnv presented as one (N, D)-batched env (parallel trajectory collection on the CPU/MuJoCo backend)
#920	BaseRLAlgo.evaluate() -- deterministic eval peer of train() (mean action, frozen normalizer, genuine-terminal success_rate)

New policy providers

PR	What
#663	VERA video-to-action policy provider (closes #78)
#466 (closed)	WBCPolicy (wbc/sonic) -- GR00T Whole-Body-Control (SONIC) for G1

Humanoid-locomotion policies (v0.5.0-scoped, now landing on main)

PR	What
#847	WBC: drop the wrong-family default checkpoint; actionable error + SONIC-stack detection
#848	auto-install WBC torque control on the run_policy default path
#852	WBCGaitPolicy gait-clock variant -- 95-dim obs + bipedal phase clock
#854	accept canonical GR00T-WholeBodyControl ONNX filenames without rename
#856	MotionBricks generative-motion provider for the Unitree G1
#876	MotionBricks consumes locomotion_style to steer the gait
#849	CompositePolicy -- stack locomotion + manipulation on one robot

Policy execution correctness (RTC / chunking / eval)

PR	What
#703	async_rtc chunk pipeline overlaps inference with execution in run_policy
#707	honor policy chunk length in the synchronized multi-robot loop (shared ChunkedPolicy)
#715	execution_horizon contract so RTC actually re-queries mid-chunk
#714 #725	RTC inference-delay uses real control frequency; made deterministic via caller step count
#731	honor ACT temporal ensembling instead of silently dropping it
#728	legacy success_fn path no longer drops the action chunk / checks success on stale obs
#772	auto-enable async_rtc for chunk-emitting policies + harden the seam
#776	honour the RTC contract in the real-robot (hardware) control loop
#779	re-anchor the RTC chunk-seam prefix for relative-action policies (was blended in a stale frame)
#780	forward inference_delay to lerobot's RTC denoiser (fixes a 2nd-chunk TypeError)
#784	pin the chunk re-query interval floor (>= 1) invariant
#788	RTC inference telemetry + opt-in async-RTC overlap in eval_policy
#790	detect image features by FeatureType.VISUAL, not a name substring
#794	policy resolution no longer shadows lerobot.policies with a partial stub
#833	fire on_frame on the success_fn eval path + expose it on eval_policy
#857	forward policy_kwargs through eval_policy / evaluate_benchmark
#905	eval_policy resolves robot_name like run_policy (auto-select the sole robot)

lerobot_local hardening (MolmoAct2 / SO-arm)

PR	What
#693 #737	SO-arm degrees<->radians + gripper RANGE_0_100; mid-center conversion like motors_bus
#753	accept HWC/uint8 camera frames (canonicalize layout before preprocess)
#704	surface silent zero-fill / near-zero actions that freeze the arm
#686	cache loaded model across policy instances
#773	surface lerobot_local model-load telemetry end-to-end
#653 #654 #680	MolmoAct2 transformers pin + report all missing aux deps in one error
#818	honor camera_key_map on the preprocessor remap path
#827	fail fast on camera / obs_rename mismatch before the model download
#834	catch ValueError when resolving unknown LeRobot policy types (clean ImportError)
#850	pin HF Hub revision for lerobot_local checkpoints
#883	reconcile a device-pinned processor step onto the host device
#892	canonicalize bare camera frames on the declarative embodiment path
#894	drive a single-camera SO-101 ACT on the declarative path
#895	surface routing-degradation telemetry (positional_fallback_used / generic_state_keys_used) in run_policy/eval_policy
#897	make state-key mismatch loud instead of a silent open-loop rollout
#898	surface non-ImportError module failures as a clean ImportError in resolve_policy_class_by_name
#900	bind bare declared image keys by name on the VLA observation path
#917	batch declarative obs when the pipeline has no AddBatchDimension step
#935	empty instruction no longer crashes the language-VLA preprocessor with a cryptic KeyError

Performance

PR	What
#782	PersistentPolicy + cache controls + policy_resident_rss_mb telemetry -- synchronous persistent worker eliminates per-episode model-reload in multi-rollout loops

Recording / dataset integrity

PR	What
#666	save_episode() for explicit episode boundaries
#691	reset() during recording flushes the buffered rollout as an episode
#716	per-episode recorder.save_episode wired in policy_runner (#708)
#681	refuse to push an empty LeRobotDataset to the Hub
#687	surface + remap camera-key mismatch instead of silently dropping frames
#684	replay reads actions from the column store, never decodes discarded video
#770	episode-count contract for run_policy + verify_dataset_episodes
#795	verify_dataset_episodes cross-checks meta/info.json against the parquet
#796	strands-robots verify-dataset CLI for episode-count integrity
#797 #805	scope recorded cameras via start_recording(cameras=...) / run_policy(dataset_cameras=...)
#798	record a per-step observation across action chunks (no stale-obs frames)
#800	verify per-episode video files in verify_dataset
#807	don't negatively cache lerobot-availability probes (recording recovers when lerobot resolves)
#813	accept an existing empty root dir in start_recording
#888	default to OpenCV-decodable H.264; route vcodec onto the LeRobot 0.5.2 encoder config
#906	verify-dataset flags dead (all-zero) action/state columns
#924	warn when start_recording sweeps the implicit default camera into a dataset

GPU-render reliability (EGL)

PR	What
#909	warn when MuJoCo EGL silently falls back to a CPU software rasterizer (was silent)
#921	auto-register the NVIDIA EGL vendor ICD so MUJOCO_GL=egl renders on the GPU

Simulation facade -- discovery, ergonomics + input validation

PR	What
#692	first-class multi-episode run_policy(n_episodes=N)
#665	agent-consumable JSON block returned from run_policy
#720	run_policy @tool wrapper (closes #708 fabrication vector)
#837	forward video config through the run_policy tool
#655	list_bodies to discover camera mount points
#758	zero-config robot discovery from robot_descriptions (long-tail Menagerie robots)
#761	surface joint velocity + base IMU in observation; WBC-driven recording
#764	roll back the spec mutation when an object/camera injection recompile fails
#771	surface [Benchmark] actions in the tool_spec description + enum/description parity
#809	render() honors per-camera configured resolution
#820	return a viewable depth PNG from render_depth
#891	metric depth passthrough + capture the fd-2 depth warning + silence export_xml attach noise
#823	validate add_camera fov/width/height eagerly
#829	accept ordered action vectors in send_action
#831	validate replay_episode robot_name + reject a negative episode index
#839	reject non-positive action_horizon in run_policy/eval_policy/start_policy
#844	reject non-positive episode/step counts across eval_policy + evaluate_benchmark
#845	complete control_frequency validation across all entry points/paths
#855	source the MuJoCo "no world" guard message from one shared constant
#879	validate replay() speed is positive (no ZeroDivisionError / silent full-speed)
#896	fail-fast on 100% unresolved action keys + per-actuator resolution stats
#903	resolve vertical (top-down/up) camera look-at orientation
#881	route Robot(backend=...) through create_simulation for unified backend selection
#841 #885	list_policy_types() discovery surface for lerobot_local + re-export as a peer of list_providers
#812 #815 #825	list render_depth/render_all, eval_policy/replay_episode, and scene-construction methods in the describe() discovery surface
#902	key policies by actuators, not joints, so tendon-gripper / passive-joint robots actually move
#926	fail-fast on an action-vector length mismatch in send_action
#927	randomize() recolors unnamed + material-backed geoms and reports the true recolored count
#933	scale body_inertia with body_mass in MuJoCo randomize_physics (mass-only scaling left bodies physically inconsistent)
#936	correct the MuJoCoSimEngine.__init__ **kwargs docstring to match runtime behavior
#939	add_object rejects unknown kwargs instead of silently dropping them
#940	mj_forward after mj_resetData so reset() leaves render-ready state
#945	mj_forward after randomize() so light-position jitter is not a silent render no-op

Newton GPU backend (post-tag maturation of #470, toward MuJoCo parity)

PR	What
#769	honour configured gravity in the Newton backend; add set_gravity / set_timestep
#774	Newton scene-discovery + per-joint state parity with the MuJoCo backend
#785	wire DatasetRecorder into the Newton backend (recording parity)
#786	domain randomization + sensor-noise hooks
#789	named per-robot cameras (MuJoCo parity)
#791	live interactive viewer
#792	mesh-object support in add_object (MuJoCo parity)
#793	load robot_descriptions URDF robots
#802	get_observation renders cameras while recording (parity with MuJoCo)

Scenes / registry / assets

PR	What
#901	add_object material/texture/reflectance for MuJoCo scenes
#904	make LeKiwi simulatable (Robot("lekiwi", mode="sim")) via a GitHub asset source
#781	re-export all @tools at the package root
#783	re-export list_providers + register_policy at the package root
#775	conform registry hardware entries to the full LeRobot catalog
#799	add reBot B601-DM single + bimanual hardware entries

Humanoid / G1 + hygiene

PR	What
#662	end-to-end VLA-on-G1 workflow (closes #471)
#766	embed the G1 walk rollout visual in the WBC policy docs
#664	harden device_connect + tools against auth bypass and path traversal
#673 #730 #733	strip emoji from runtime strings + widen no-emoji guard; ASCII-hyphen sweep
#824	reject numpy booleans in teleop input frames (mesh)

Added-then-removed in this window (net: neither ships)

PR	What
#853 -> #889	planning package + run_policy(planner=) added (#853), then removed (#889) -- locomotion intent is expressed only through policy_kwargs; MotionBricks keeps a plain-string locomotion_style contract (no planner import)
#878 -> #890	sim_managers config-driven RL-env framework added (#878), then removed (#890) as off-mission with no consumers

Plus a large test-coverage hardening sweep (#787 #801 #803 #804 #806 #808 #810 #811 #814 #816 #817 #819 #821 #826 #828 #830 #832 #835 #838 #840 #842 #846 #851 #882 #893 #907 #908 #910 #911 #912 #913 #914 #916 #919 #922 #923 #925 #928 #932 #934 #937 #938 #941 #942 #943 #948 #951, with #822 fixing the molmoact2 sim-pickplace example embodiment) across policy_runner/RTC, verify-dataset, mesh ACL + safety-resume + audit-log + wire-validation + e-stop + Zenoh transport, VERA view-key/handshake, MotionBricks + CompositePolicy, Newton domain-randomization, gr00t dispatch, discovery, libero adapter, the EGL vendor-ICD scan, and the from-source lerobot conformance invariant. Examples/docs also grew a CPU-only getting-started notebook series (#843), a use_lerobot discovery example (#874), the add_object size-is-full-extent docs (#899), and a #944 CHANGELOG backfill of 12 missing wave entries. Release-enabled by #947 (--cov-fail-under=80) + #950 (PyPI vera-extra fix).

v0.4.1 release status

• Bump version + tag v0.4.1 + GitHub release -- published 2026-07-01 (tag at #947).
• Collapse [Unreleased] into a dated ## [0.4.1] heading -- landed via #977.
• Open a fresh [Unreleased] for the post-v0.4.1 wave -- landed via #977.

Post-v0.4.1 -- shipped as v0.5.0 (2026-08-04) / v0.5.1 (2026-08-06); no v0.4.2 was ever cut

Merged to main after the v0.4.1 tag (#946, #949, #952 -> #1530; +516 PRs). Mostly consolidation -- the CHANGELOG heading collapse, a breaking-change CI gate, a sustained run of sim-facade correctness fixes, new sim capabilities, a lerobot 0.6 dependency bump, a remote policy inference client/server split (plus a lerobot gRPC async provider), a floating-base state lane, the absorbed Isaac Sim backend, the first locomotion reward-term DSL primitives now composed into a runnable go2_walk_forward benchmark, a terrain curriculum for locomotion (create_world heightfield kinds + a difficulty knob), a broad describe() discovery-surface expansion, ROS 2 action support, training fixes, lerobot_local normalization/embodiment-integrity fixes, example/packaging fixes, and a broad contract-pinning test sweep. Grouped by theme.

CHANGELOG housekeeping

PR	What
#977	collapse [Unreleased] into a dated ## [0.4.1] heading + open a fresh [Unreleased] for the post-tag wave
#1188	normalize two off-convention [Unreleased] CHANGELOG entries

Public-API breaking-change CI gate

PR	What
#946	informational breaking-change check for the public API
#949	detect breaking changes to the AgentTool action contract
#963	make the breaking-change check fork-safe

Dependencies

PR	What
#1114	track lerobot 0.6 -- require lerobot>=0.6.0 and drop the 0.5.1-era torch/torchcodec overrides

Isaac Sim backend (absorbed from robots-sim, epic #1144)

PR	What
#1156	absorb the Isaac Sim backend from robots-sim into this repo (closes the #1144 epic)
#1151	builtin backend wins over a stale entry-point plugin of the same name (closes #1150)
#1155	add the [isaac] optional-dependency extra (closes #1147)
#1154	port the Isaac Sim GPU integration tests into tests_integ/ (closes #1148)
#1153	document the Isaac Sim backend (closes #1149)
#1205	correct create_simulation() isaac-backend examples to the builtin backend (closes #1146)
#1354	Isaac create_world accepts terrain/difficulty and rejects terrain with an actionable error (no heightfield support; points at backend="mujoco")
#1357	Isaac add_object accepts + loudly rejects the base mesh_path/material contract (was silently swallowing them via **kwargs)
#1360	Isaac add_robot accepts + loudly rejects the base keyframe contract (was a bare TypeError)
#1372	Newton / Isaac create_world(difficulty=...) rejects a non-default difficulty with no terrain (was silently ignored; points at backend="mujoco")

Reward / predicate DSL + terrain curriculum -- locomotion groundwork now runnable (toward #873)

PR	What
#1198	base_velocity velocity-tracking reward term for the predicate/reward DSL
#1203	base_height locomotion-regularizer reward term
#1208	base_orientation flat-orientation reward term
#1215	base_lin_vel_z / base_ang_vel_xy locomotion-regularizer reward terms
#1239	base_velocity_tracking exponential-kernel velocity-tracking reward term
#1230	base_tipped fall-over termination predicate (failure-clause counterpart of base_orientation)
#1233	base_below_z height-collapse termination predicate
#1255	base_beyond_x forward-progress success predicate
#1259	register_builtin_benchmarks() + the canonical go2_walk_forward locomotion benchmark -- first runnable benchmark composing the DSL primitives (Unitree Go2 velocity-tracking; discoverable via list_benchmarks(), runnable via evaluate_benchmark())
#1287	g1_walk_forward humanoid locomotion benchmark (Unitree G1) -- the #873 G1 embodiment
#1288	t1_walk_forward humanoid locomotion benchmark (Booster T1) -- the #873 T1 embodiment
#1321	base_beyond_y lateral-progress success predicate + go2_strafe_left benchmark (vy command axis)
#1324	base_yaw_beyond turn-progress success predicate + go2_turn_left benchmark (wz command axis) -- completes the Go2 vx/vy/wz velocity-tracking vocabulary
#1223	reject a reward term used where a bool predicate is required in the benchmark DSL
#1236	registry name overrides the spec-internal name in register_benchmark_from_file
#1336	create_world(terrain="rough") -- rough-ground heightfield for locomotion (first terrain kind; #873 terrain-curriculum groundwork)
#1338	create_world(terrain="stairs") -- stepped-ground heightfield for locomotion
#1339	create_world(terrain="pyramid") -- concentric stepped pyramid for omnidirectional locomotion
#1340	create_world(terrain="slope") -- inclined-ramp heightfield for locomotion
#1344	create_world(terrain=..., difficulty=...) -- terrain elevation-curriculum knob (scales the heightfield peak so a curriculum ramps terrain magnitude across resets)
#1350	advertise create_world(difficulty=...) in the sim tool_spec + keep the terrain-kind hint in sync with SUPPORTED_TERRAINS
#1346	declarative benchmark spec accepts a natural-language instruction
#1364	base_below_z measures clearance above the local terrain surface, not absolute z (terrain-relative fall predicate)
#1368	base_height reward measures clearance above the local terrain surface, not absolute z
#1386	a floating-base robot spawns seated on the terrain surface (feet on the heightfield) at spawn/reset instead of buried below it
#1388	get_ground_height(x, y) -- public query for the local terrain surface height (facade + agent-dispatch action; place objects/cameras/goals on terrain correctly)

Remote policy inference (new capability)

PR	What
#1142	remote policy inference client/server split (strands_robots.inference) -- PolicyServer wraps any Policy; RemotePolicy is a drop-in Policy forwarding observations over a WS-JSON protocol; wired as the remote provider (create_policy("ws://gpu-box:8765")), RTC contract preserved end to end
#1152	decode observations into writable arrays (remote path parity -- in-place VLA preprocessors + torch.from_numpy no longer hit read-only arrays)
#1158	lift the server WebSocket frame limit so camera-image observations are not rejected
#1190	warn on a mistyped RemotePolicy endpoint instead of silently defaulting to localhost
#1212	lerobot gRPC async inference provider -- remote inference over lerobot's async gRPC transport (a real provider, distinct from the WS-JSON RemotePolicy)
#1250	PolicyServer publishes its port before marking the server bound (no race reading the port on startup)
#1398	lerobot_async forwards a rename_map to the remote PolicyServer for observation-key remapping
#1458	lerobot_async sources its server-supported policy set live from lerobot instead of a hardcoded copy

Simulation facade -- correctness, discovery + backend parity

PR	What
#955	move_object repositions static objects instead of silently no-op'ing (#966 pins the failure contract)
#962	bind replay / send_action action vectors to actuators, not joints
#968	flag eval_policy success_rate as unmeasured when no success criterion is given
#970	accept cameras= scope in Newton start_recording (dataset_cameras parity with MuJoCo)
#981	render_depth honors a camera's configured resolution
#991	enforce the Strands tool-result contract (move top-level extras into the content JSON)
#992	set_obs_noise on the MuJoCo backend (sensor-noise parity with Newton)
#997	list_cameras() on the MuJoCo backend (camera discovery at backend parity)
#1002	inverse_dynamics solves for zero desired acceleration -- gravity/bias compensation (was ~0 torques regardless of pose, stale qacc)
#1003	register mesh assets on a runtime add_object(shape="mesh") injection
#1006	get_jacobian / get_body_state reflect the current qpos/qvel (were computed against a stale configuration)
#1012	replay steps a full control period per recorded frame (was one physics dt -> under-integrated playback)
#1015	set_body_properties scales inertia with mass (mass-only left bodies physically inconsistent)
#1032	recompute geom_rbound / geom_aabb when set_geom_properties resizes a geom (stale broadphase bounds)
#1038	raycast / multi_raycast refresh geom poses and lock before mj_ray (was reading a stale configuration)
#1044	eval_policy(video=) records one rollout MP4 per episode
#1050	get_energy recomputes the forward pipeline before reading (was the energy of a stale pose after a direct qpos/qvel write)
#1051	evaluate_benchmark honors policy_object + control_frequency/control_substeps (parity with run_policy/eval_policy)
#1056	drop the _set_eval_seed back-compat alias for the public set_eval_seed
#1059	cap rollout-video fps at control_frequency for real-time playback
#1062	describe() advertises the physics-introspection / grounding surface (get_body_state, get_contacts, sensors, ...)
#1065	grasped predicate matches the LIBERO <body>_g<idx> geom-naming convention
#1069	honor CooperativeStop from an on_frame hook in eval_policy / evaluate_benchmark
#1076	evaluate_benchmark records one rollout MP4 per episode (video=)
#1078	verify-dataset flags truncated / partial-encode videos (a present, non-empty MP4 with fewer frames than recorded)
#1083	move_object places a dynamic object at rest at the new pose (drops retained velocity)
#1090	surface unresolvable body/joint names in the predicate & reward DSL (was a silent no-op returning a constant)
#1100	warn when an action value is clamped by a ctrl-limited actuator
#1112	add_robot(keyframe=...) spawns a robot in its canonical home pose
#1134	surface base state for a mobile base whose free joint is unnamed (was silently dropped for LeKiwi-style unnamed free joints)
#1137	get_robot_state surfaces a floating base as a structured base entry (not a flattened joint row)
#1139	Newton joint index maps no longer shift every joint after a floating base
#1159	make NewtonSimEngine teardown safe when __init__ fails part-way
#1164	Newton get_observation/get_robot_state surface a floating base as a structured pose (MuJoCo parity)
#1169	warn when a floating base's orientation + angular velocity are dropped from a recorded dataset
#1172	preserve a floating base's orientation + angular velocity in recorded datasets
#1176	surface a floating base's full position + linear velocity in the observation
#1179	reject a non-scalar send_action dict value instead of crashing
#1183	preserve a floating base's state in Newton-recorded datasets
#1187	report a floating base's angular velocity in the body frame on Newton
#1189	reject robot-setup kwargs in backend constructors instead of silently dropping them
#1191	surface cached-XML refresh failures in scene mutations instead of swallowing them
#1252	warn on an out-of-ctrlrange clamp for joint-name-addressed actions (sibling of #1100)
#1249	describe() advertises load_scene + the object-manipulation siblings
#1254	advertise get_features in describe() (the method run_policy recommends)
#1256	describe() advertises the benchmark scoring family (evaluate_benchmark / list_benchmarks / register_benchmark_from_file)
#1257	describe() advertises the sim-state checkpoint + pose-setting family
#1261	describe() advertises the background-policy stop/list lifecycle
#1262	describe() advertises the physics-tuning / domain-perturbation family (set_gravity / set_timestep / set_body_properties / set_geom_properties / apply_force)
#1264	describe() advertises the plain-MP4 camera-recording family (start_cameras_recording / stop_cameras_recording / get_cameras_recording_status)
#1266	describe() advertises the robot-registry + remove_robot surface (list_urdfs / register_urdf / remove_robot)
#1267	describe() advertises the world-lifecycle + MJCF-editing family (create_world / destroy / patch_scene_mjcf / replace_scene_mjcf / export_xml)
#1277	describe() advertises the teleoperation family (attach_teleop / teleoperate / stop_teleoperate / ...)
#1281	describe() advertises the interactive-viewer family (open_viewer / close_viewer)
#1286	describe() advertises the multi-robot rollout family
#1291	get_contact_forces reflects the current pose (mj_forward before reading)
#1298	base describe() advertises remove_robot + the scene-variation facades
#1299	add_robot reports an actionable error for an unknown model (close-match + list_urdfs)
#1303	actionable not found errors for move_object / remove_object / remove_camera
#1306	actionable Robot 'X' not found across the facade (close-match + list_robots)
#1308	actionable robot not found on run_policy / eval_policy / evaluate_benchmark (sibling of #1306)
#1318	get_observation no longer emits a floating base's free joint as a degenerate scalar (drops the junk floating_base_joint column from recorded floating-base datasets; both backends)
#1275	point warp/mjwarp install hints at strands-robots[sim-newton] (closes #1268)
#1348	base describe() advertises create_world / destroy
#1352	unknown-camera error hints the canonical list_cameras action
#1365	run_policy single-robot policy binding is fail-soft like the multi-robot path
#1370	object/camera facade guards catch the partial-world state (empty world / no robot)
#1383	add_robot(keyframe=...) keeps earlier keyframe-spawned robots posed when adding more robots (was silently collapsed to the zero pose)
#1405	DatasetRecorder.create(overwrite=...) for honest re-recording into an existing repo_id
#1407	strip a stray leading space from the render / render_depth no-GL error messages
#1416	get_ground_height(x, y) accepts NumPy scalar coordinates
#1417	add_camera(fov=...) accepts NumPy scalar angles
#1418	set_gravity accepts a NumPy real scalar z-component
#1419	sim action router accepts NumPy scalar components in vector params
#1420	run_policy/eval_policy/evaluate_benchmark accept a NumPy scalar control_frequency (+ reject non-finite)
#1423	replay_episode accepts a NumPy scalar speed + rejects non-finite
#1424	set_body_properties rejects a non-finite mass (nan/+inf slipped the positivity guard)
#1425	set_geom_properties rejects non-finite/invalid color, friction, size
#1426	set_joint_positions/set_joint_velocities reject non-finite/non-numeric values
#1427	move_object validates the pose is finite numbers before mutating state
#1429	add_camera rejects non-finite/non-numeric position/target components
#1430	add_object/add_camera reject a non-finite/malformed numeric vector before baking it into the MJCF
#1433	add_robot validates its base pose (finite, numeric, correct length) before baking it into the scene
#1435	raycast/multi_raycast validate origin/direction are finite numbers
#1436	physics/introspection lookups give an actionable not-found for an unknown body/site/geom/sensor

Hardware

PR	What
#1026	run_policy(policy_object=...) on the hardware Robot (sim parity for a pre-built in-process policy, no server on a port)

Training

PR	What
#1022	honor TrainSpec.learning_rate in the lerobot_local trainer (was silently dropped -> policy.optimizer_lr)
#1047	discover LeRobot policy types from lerobot's live registry (drops the hardcoded allowlist)
#1116	Gr00tTrainer.build_command augmentation parity with build_finetune_config
#1126	build_command emits --peft.lora_alpha for LoRA parity with build_config
#1085	HIL consent gate blocking dangerous lerobot_train extra_flags passthrough (security; a long-open PR, merged out of order)
#1438	build_command emits --reward_model.pretrained_path for warm-started reward-model runs
#1447	forward repo_type through StreamingDatasetReader.open (stream a Hub dataset of a given repo type into training)

ROS 2 interop

PR	What
#995	ROS 2 action support in use_ros (list_actions, action_send_goal) + goal-level navigate_to on RosBridgedRobot

lerobot_local / teleop / registry / mesh / tools fixes

PR	What
#979	derive teleoperate() status from frame/error counters (a dead teleop reported "success")
#980	reject non-finite STRANDS_MESH_CAMERA_HZ (inf busy-spins the camera loop)
#985	warn when a checkpoint's normalization pipeline is present but inert
#986	registry hot-reload honors the user-overlay file's mtime
#1057	anchor mesh identifier allowlists with \Z so a trailing newline is rejected
#1067	lerobot_teleoperate / lerobot_train status returns telemetry in a {"json": ...} block instead of **spread top-level keys the runtime drops
#1073	surface an embodiment-config failure instead of silently discarding the pipeline behind a misleading diagnostic
#1080	download_assets status marks each robot available or missing (was ambiguous)
#1082	ZMQ REQ inference-client sets LINGER=0 so teardown never blocks
#1096	reconstruct OLD-FORMAT in-model normalization dropped on a checkpoint load
#1097	degrade to passthrough on a corrupt model.safetensors instead of crashing
#1099	keep observation.state index-aligned on a partial robot_state_keys mismatch
#1107	zero-fill a missing embodiment state_key in place; use aloha's 14 actuator keys
#1178	resolve the hf CLI from the interpreter env in sync_to_bucket (recorder Hub upload)
#1195	decode cosmos3 msgpack arrays into writable, owning buffers
#1202	remove phantom lerobot_async / dreamgen provider references from the registry
#1204	raise on an unrecognized norm-stats mode instead of silent passthrough
#1207	run observation-history policies (Diffusion/VQBeT) via select_action
#1210	raise on an unknown list_robots(mode=...) instead of a silent unfiltered list
#1422	is_molmoact2 respects an explicit policy_type without a Hub probe
#1409	format_robot_table renders every robot regardless of category
#1446	strip orphan whitespace from the lerobot_calibrate header lines
#1450	catch zenoh.ZError in the mesh best-effort lifecycle cleanup
#1451	reject a policy alias/shorthand that shadows another provider's canonical name (registry)
#1457	screen the use_lerobot method blocklist against the module path

Examples / packaging

PR	What
#973	platform-aware MUJOCO_GL default (cgl is macOS-only; examples crashed on Linux)
#975	[ollama] extra so the local-LLM agent path installs out of the box
#989	examples/train_ppo_reach.py passes its own spec validation (num_mini_batches 4 -> 5)
#1331	add the streaming data-loop getting-started notebook (05)
#586	collect-train-run MolmoAct2 example (examples/lerobot/) -- long-open PR, merged out of order
#1390	10_evaluate_benchmark.py -- benchmark-evaluation getting-started example
#1391	11_author_a_benchmark.py -- declarative-benchmark authoring example
#1393	12_domain_randomization.py -- domain-randomization example
#1394	13_physics_introspection.py -- physics-introspection example
#1395	14_save_state_and_perturb.py -- save-state + perturb example
#1396	15_robot_catalog.py -- robot-catalog discovery example
#1402	platform-aware MUJOCO_GL in the notebooks + lerobot e2e fixture (fixes #1017; supersedes #1018)
#1434	consolidate loose top-level examples into topic subfolders

Docs (README + docstrings)

PR	What
#957	list the Cosmos 3 provider in the summary tables
#958	correct the simulation action count (64 -> 65)
#960	document the create_trainer training providers
#1013	correct stale lerobot reference paths in the norm_stats docstring
#1016	document action_horizon as a lower bound (clamped up to execution_horizon), not a maximum
#1054	correct the async_rtc seam-blending claim for non-RTC chunk policies
#1066	make the dataset IO docstrings self-contained
#1071	cross-reference sibling cosmos3 modules with :mod: roles, not source filenames
#1079	cross-reference MuJoCo backend siblings by module, not source filename
#1120	correct post-lerobot-0.6 VLA dependency guidance
#1129	correct stale from-source install hints for the >=0.6.0 PyPI floor
#1141	document steerable dataset annotation with lerobot-annotate
#1193	correct stale list_providers() example output in api-reference + policies overview
#1201	list all 11 providers in the policies overview + a regression-test guard
#1225	enumerate every benchmark reward term in the predicate-registry docstring
#1293	describe() advertises the humanoid builtin benchmarks (docs)
#1322	cross-reference MockPolicy by module path, not source filename
#1327	cross-reference MockTrainer by module path, not source filename
#1330	cross-reference sibling mesh modules by module path, not source filename
#1337	cross-reference simulation modules by module path, not source filename
#1374	correct the README simulation action count (65 -> 67)
#1375	list notebook 05 in the notebook indexes
#1376	point the quickstart at the notebook series
#1378	mention procedural terrain in the README simulation-tool feature
#1379	add the remote policy provider to the README providers table
#1387	add the procedural terrain example (examples/.../09_procedural_terrain.py)
#1401	document the public RL trainer/env methods + properties
#1410	correct residual pre-0.6 lerobot install guidance (architecture / troubleshooting / molmoact2)
#1441	update the Isaac backend docs to the in-tree strands-robots[sim-isaac] (closes #1440)
#1452	document the concrete Trainer provider public methods
#1456	point the streamed-training instructions at the renamed lerobot.scripts.lerobot_train module

CI

PR	What
#1167	retry apt-get update to survive transient third-party mirror failures
#1240	silence a pyarrow import-untyped mypy break after it dropped its py.typed marker (dependency drift)
#1284	remove a duplicate auto-generated codeql-advanced workflow

Coverage / contract-pinning sweep

PR	What
#952	dataset_recorder sync_to_bucket hf-CLI failure + delete-flag contract
#953	mesh/audit seq-counter restore/lock fail-soft contract
#954	lerobot_local ProcessorBridge optional-dep degradation contracts
#956	mesh/bridge idempotent-teardown fail-soft contracts
#959	lerobot_local best-effort CUDA allocator release in clear_model_cache
#961	render_all namespaced camera short-name resolution
#964	bidirectional name/robot_name router alias + None-vector passthrough
#965	mesh CA trust hardening (operator-pin allowlist + download timeout)
#967	RL FastSAC truncation-bootstrap contract (time-out vs terminal)
#969	MotionBricks control-signal builder edge cases (invalid mode/range/direction)
#971	LerobotTrainer builder-side escape-hatch validation

Plus the v0.4.1-era contract-pinning sweep, which ran continuously from #976 to #1455 across sim, mesh, lerobot_local, dataset_recorder, gr00t, MoveIt2, use_ros, CompositePolicy, teleop, training, hardware, the RL backends, LIBERO, VERA, WBC, the Newton and Isaac backends, the recorders, the predicate and reward DSL, the inference server, the registry and the docstring-cross-reference guards. Its per-batch prose was condensed on 2026-09-13 -- every pull request it enumerated is kept here as a roster, 238 in all: #976 #982 #983 #984 #987 #990 #993 #996 #998 #999 #1001 #1004 #1005 #1007 #1008 #1009 #1010 #1011 #1014 #1019 #1023 #1024 #1027 #1028 #1029 #1030 #1031 #1033 #1036 #1037 #1039 #1040 #1041 #1042 #1043 #1045 #1046 #1048 #1049 #1052 #1053 #1055 #1058 #1060 #1061 #1063 #1064 #1068 #1070 #1072 #1074 #1075 #1077 #1081 #1088 #1089 #1091 #1092 #1093 #1094 #1095 #1098 #1101 #1102 #1103 #1104 #1105 #1106 #1108 #1111 #1113 #1115 #1117 #1118 #1119 #1121 #1122 #1123 #1124 #1125 #1127 #1128 #1130 #1131 #1132 #1133 #1134 #1135 #1136 #1138 #1140 #1143 #1157 #1160 #1161 #1162 #1163 #1165 #1166 #1168 #1170 #1171 #1173 #1174 #1175 #1177 #1180 #1181 #1182 #1184 #1185 #1186 #1192 #1194 #1196 #1197 #1199 #1200 #1206 #1209 #1211 #1213 #1214 #1216 #1217 #1218 #1219 #1220 #1221 #1222 #1224 #1226 #1227 #1228 #1229 #1231 #1232 #1234 #1235 #1237 #1238 #1241 #1242 #1243 #1244 #1245 #1246 #1247 #1248 #1253 #1258 #1260 #1263 #1283 #1285 #1289 #1290 #1292 #1294 #1295 #1296 #1297 #1300 #1301 #1302 #1304 #1305 #1307 #1309 #1310 #1311 #1312 #1313 #1314 #1315 #1316 #1317 #1319 #1320 #1323 #1325 #1326 #1328 #1329 #1332 #1333 #1334 #1335 #1341 #1342 #1343 #1345 #1349 #1351 #1353 #1355 #1356 #1358 #1359 #1361 #1362 #1363 #1366 #1367 #1369 #1371 #1373 #1377 #1380 #1381 #1382 #1384 #1385 #1389 #1392 #1397 #1399 #1400 #1403 #1408 #1411 #1413 #1414 #1415 #1428 #1431 #1432 #1437 #1439 #1442 #1443 #1444 #1445 #1448 #1449 #1453 #1454 #1455.

v0.5.0 -- Humanoid locomotion + GPU sim + hardware (milestone drained, tag shipped)

Milestone 0.5.0: 11 closed / 0 open issues, 100% on the bar -- unchanged for a nineteenth sync, having moved all the way on 2026-08-24. The two issues that contested its substantiation are now closed, both NOT_PLANNED, both on 2026-09-10 inside the same second range. #2720 -- the issue reporting that 57 of 60 issues closed COMPLETED in one 10-minute window with no linked pull request, and that three of epic #858's deliverables were absent from the tree -- closed at 04:21:37Z on the grounds that "the board state it describes is history; closing with #858, whose absent deliverables are now out of scope." #858, the last unscheduled epic, closed one second later. So the milestone's reading is unchanged and the contest is settled by decision rather than by delivery: #3206 (FastTD3 as a native off-policy RL provider, +2,199 / -95 across 29 files, naming both issues in its title) is the one deliverable of the three that reached main, and it declared no closing reference, so the other two were dispositioned rather than shipped. That is a legitimate call and it is recorded as one -- NOT_PLANNED says "we are not doing this", which is what the previous twelve syncs asked for in place of a COMPLETED claim. It also means no issue is open against this milestone's substantiation any more, and this page is the only remaining record of the reservation. (The API's closed_issues: 14 counts the three closed pull requests also assigned to the milestone -- #163, #317, #319 -- so 11 is the issue figure and both are correct at different scopes.) The final four (#469, #472, #473, #474) all closed on 2026-08-24T09:30Z, within six seconds of each other, each with stateReason: COMPLETED. Together with #467/#468/#475 (closed 2026-08-14) and #466/#470/#471 (closed in June), every issue this milestone ever carried is closed, so the v0.5.0 and v0.5.1 tags -- both cut while it was still open -- are no longer ahead of it.

The code-level reservations this page recorded are unretired, and now unrepresented. All four 2026-08-24 closes fall inside the window #2720 described, and none was retired by a merged pull request: #469's namesake kinematic-constraints API still has no constraint module under strands_robots/policies/kimodo/; #472's registry expansion and #474's mobile-base velocity contract were last recorded as unstarted, though #2689's AckermannRosRobot and #2696's MobileBaseRobot drive contract are the closest thing on main to #474's scope; #473's HW-in-the-loop lane has its sim prerequisite (LeKiwi #904) and the floating-base state lane behind it, but no eval lane. The milestone is drained; the tracker and the tree agree less than the bar suggests, and with #2720 closed there is no longer an issue saying so.

Closed

Issue	What	Via
#466	WBCPolicy -- GR00T Whole-Body-Control (SONIC)	merged (closed 2026-06-18)
#470	NewtonSimulation -- GPU-native backend (Warp / MuJoCo-Warp)	merged (closed 2026-06-19)
#471	End-to-end VLA-on-G1 hero example	#662 (closed 2026-06-25)
#467	MotionBricksPolicy -- generative G1 locomotion	provider #856/#876; issue closed 2026-08-14
#468	CompositePolicy -- stack policies across DOF groups	#849, integration-tested #1009; issue closed 2026-08-14
#475	QwenVLAPolicy -- Qwen-VL-based manipulation provider	closed 2026-08-14
#469	KimodoPolicy -- text-to-motion + kinematic constraints	provider #2270/#2279/#2271 merged; closed COMPLETED 2026-08-24 with no linked PR -- the constraints API is not in the tree
#472	Expand robots.json (mobile-manip + expressive gaps)	partial on main (reBot #799, LeKiwi sim #904); closed COMPLETED 2026-08-24 with no linked PR
#473	HW-in-the-loop eval lane (LeKiwi, EarthRover)	prerequisites landed (#904, floating-base state); closed COMPLETED 2026-08-24 with no linked PR
#474	Mobile-base action contract (velocity cmds) for lerobot_local	nearest code on main is #2689/#2696; closed COMPLETED 2026-08-24 with no linked PR
#94	this roadmap	closed 2026-08-14 as completed

Open

None -- the milestone is drained. Four of the eleven closes are unverified against a merged PR; see #2720.

Suggested sequencing

1. Cut the tag. Nothing is holding it, and the cost of waiting keeps growing. 0.5.0 is drained, twelve driver buses are in, FastTD3 and the observer lane are in, an RL checkpoint is deployable, an evaluation reports a reliability figure, and the dashboard now has a server (#3695) -- and there is still no ## [0.5.0] or ## [0.5.1] heading, [Unreleased] is frozen at 168 entries, and 1,600 fragments are pending in changelog.d/ (through #3750) with --check clean. 1,451 unreleased pull requests and 50 days since the CHANGELOG was last edited (#1690) are both the largest this page has recorded. The concrete cost is unchanged and unaddressed: vera, lerobot_calibrate, motionbricks and strands_robots.transforms are gone from main and still present in the last published tag, so an installer following the published docs gets four surfaces main no longer has. Milestones v0.4.0 (due 2026-07-31) and v0.3.9 (due 2026-05-30) are also still OPEN and fully drained -- closing them is free.

2. The extra-approval rule is no longer the story; the queue's size is. main requires one approval and an extra approval for unattributed changes, so one account's approval can never carry a pull request -- and that remedy is now routine, with seven heads reading APPROVED (#3729, #3743, #3746, #3749, #3751, #3752, #3753). What has not kept pace is the rest: ten heads are green, armed, thread-resolved and still REVIEW_REQUIRED -- #3205, #3343, #3427, #3496, #3515, #3577, #3615, #3714, #3726, #3737. At 21 open this is the largest queue this page has recorded, and it grew by ten in a day.

3. Decide #3343. A new outside contributor's Isaac backend: 108 files, green, all four review threads resolved, armed since 2026-09-09T13:42Z, no approving or change-requesting review ever submitted -- only comments -- and unchanged for a seventh sync. The author has done everything an author can do. It is a maintainer call on review scale, and it has been owed longer than anything else here.

4. #3205 is the oldest head on the board at eleven days -- MERGEABLE, green, armed, thread-resolved (+2,223 / -7 across 12 files), and still the mesh_bridge decomposition #2977 asked for. #3695 landing a dashboard server makes it more mergeable in principle and no more reviewed in practice; this page is the only remaining record of that request.

5. Three heads are owed by their author, and one of them is red. #3400 (3 files, CHANGES_REQUESTED, unarmed, six days old), #3640 (4 files, CHANGES_REQUESTED, one unresolved thread) and the new #3673 (3 files, CHANGES_REQUESTED, one unresolved thread, checks failing) -- all shipitfast's, all small. #3739 is a fourth kind: 22 files, two unresolved threads and no review decision yet.

6. Fixed -- the fragment naming defect is closed. changelog.d/ holds zero 0000- placeholders, where it held four for three syncs, and the guard no longer permits the number at any surface: #3715 refuses it, #3721 gives that rule one owner so every surface refuses it, and #3722 renamed the one fragment that named no pull request. Naming a fragment for the issue it answers remains the established convention and needs no change. This item is retired.

7. The duplicate-implementation rate went from zero to four, so the procedural mitigation is not holding. Two bands took none; this one took four, every one a second implementation of a fix that merged: #3665 against #3666, #3671 against #3672, #3736 against #3731, and #3732 descoped into #3733 plus #3734 -- instances sixteen to nineteen of the shape #3169 measured. #3169's negative result stands (widening the duplicate-claim key fires on a third of the queue for two genuine pairs), so the remedy is still procedural: check the issue for an open head before opening a second one. At a 21-head queue that check matters more, not less.

8. The five open issues are one safety gap, one vendor answer, and three things the tree can answer. #3122 (dashboard, P2) is still the highest-priority item on the board and #3695 sharpens it: both enrollment routes skip the auth middleware and two directory reads validate existence rather than containment, on a package that now serves and that gates hardware-moving tool calls (#3184). #3744 (policies, P2, new) is the interesting one -- the relative-action roster cannot hold across the pinned lerobot range, which is the offline-snapshot lane (#3616/#3618/#3623) read as a range rather than a version. #3099 (drivers, P3) still needs a vendor answer or a measurement on a physical pack. #3734 (simulation, P3, new) is the rotate_wrist half of #3733, deliberately deferred. #3578 (simulation, P3) overlaps #3343's diff and is best decided with that review.

9. The tracker bookkeeping is holding -- keep it. Both issues filed inside this band that could be closed from the tree were closed against the commit that answered them: #3664 via #3666 and #3684 via #3686, each inside two hours of filing. That is a fifth consecutive clean pass, after a band with no tracker movement at all. Set against 2026-08-24 (sixty closures, three reconciling) and 2026-09-10 (thirteen, one closer), it remains the standard this page asked for.

10. The removals still need a hand-written section, and five bands running have added nothing to it. The wave's five removed surfaces are vera and lerobot_calibrate (both marked !), and motionbricks, strands_robots.transforms and the MolmoAct2 normalization port (none marked). Any tooling that derives a release's breaking changes from commit metadata under-reports the wave by three. The tag is a minor at least, its note has to name all five plus the two providers #3075 still lists as going, and the forward fix is a convention: a removal that deletes a public surface carries the marker.

Other open issues (no milestone)

Five open issues, 3 -> 5, after a band in which nothing moved in either direction. Two were filed and closed inside the band against the commit that answered them (#3664 via #3666, #3684 via #3686), which extends the clean reconciliation run to a fifth pass; two more were filed and are still open. All five carry an area label and a priority. The two long survivors are now on this table for a fourteenth and fifteenth sync, and both remain blocked on something other than review.

Issue	Title	Note
#3122	dashboard: path-containment and enrollment gates owed by the server slice	filed 2026-09-02T02:22Z, dashboard / P2, one minute before the #2848 draft was closed. Both enrollment routes skip the auth middleware and two directory reads validate existence rather than containment. Still the highest-priority item on the board, and #3695 sharpens it: the package it names now ships a server, on top of gating hardware-moving tool calls (#3184) and building its consent request from a refusal code (#3403). #3509, #3594, #3710 and #3738 have each hardened a neighbouring surface without touching either route this issue names -- the fifth band running that a fix has landed beside it
#3744	training: the relative-action roster cannot hold across the pinned lerobot range	filed 2026-09-16T05:04Z, policies / P2, the newest and the most structurally interesting item on the board. The offline-snapshot lane (#3616, #3618, #3623) asked whether a copy of lerobot's vocabulary is still true; this asks whether one copy can be true, when the pinned range admits releases that disagree about the roster. #3698 documented the roster the gate accepts one commit earlier, which is what made the contradiction visible
#3099	drivers/g1: BmsState_.current has no declared unit	filed 2026-09-01, drivers / P3. The vendor IDL declares no unit and the g1_battery verb asserted amps, so the verb published one the wire does not carry. Needs a vendor answer or a measurement on a physical pack -- unchanged for thirteen syncs, and still the only item on this page that cannot be closed from the tree
#3734	simulation: rotate_wrist that did not reach should name the obstruction	filed 2026-09-16T03:53Z, simulation / P3, in the same second that #3732 closed. It is the deliberately-deferred half of the motion-that-did-not-reach lane: #3733 shipped the move_to case as a fix and this carries the rotate_wrist case, so the descope is recorded rather than dropped
#3578	simulation/isaac: SO101_RECORD_CONVERGE is read into state nothing consults	filed 2026-09-13T02:20Z, simulation / P3. A documented operator environment knob is parsed into an attribute with no reader, so setting it changes nothing -- the write-only-private-state shape #3580 and #3583 swept the package for, caught at a surface that is also an operator-facing contract. Whether the fix is to wire it up or to delete it is a design call, and it lands inside the file #3343 rewrites, so it is best decided with that review

Open PRs

21 open (11 at the previous sync) -- the largest review queue this page has recorded, and it grew by ten in a day. #3661 and #3662 merged off the previous table; twelve heads opened (#3673, #3714, #3726, #3729, #3737, #3739, #3743, #3746, #3749, #3751, #3752, #3753) and nine carried over. shipitfast has seven of the twenty-one.

The extra-approval remedy is now the board's normal path, and the bottleneck has moved from the rule to the volume. main carries require_extra_approval_for_unattributed_changes: true beside required_approving_review_count: 1 (and require_last_push_approval: true, required_review_thread_resolution: true, one required check, squash only), so one account's approval can never carry a pull request. Seven heads now read APPROVED -- #3729, #3743, #3746, #3749, #3751, #3752, #3753 -- where two did a day ago. What has not moved is the ten that satisfied every reviewable condition and wait on a second account: #3205, #3343, #3427, #3496, #3515, #3577, #3615, #3714, #3726, #3737, all MERGEABLE, green, armed, every thread resolved, all REVIEW_REQUIRED. Three are owed by their author (#3400, #3640, and the new #3673, whose checks are failing), and #3739 is its own case at 22 files with two unresolved threads and no review decision yet.

PR	Title	Status
#3205	feat(dashboard): mesh_bridge slice of #2848 -- the signed rail honours the kill switch and an e-stop reports what the peers said	opened 2026-09-05T01:34Z by an outside contributor, MERGEABLE, checks green, armed 2026-09-05T11:04Z, REVIEW_REQUIRED, its one thread resolved (+2,223 / -7 across 12 files). The oldest thing on the board at eleven days, and the decomposition #2977 asked for -- waiting on a reviewer, not on its author
#3343	fix(simulation/isaac): Isaac Sim backend	opened 2026-09-08T22:24Z by a new outside contributor, MERGEABLE, checks green, armed 2026-09-09T13:42Z, REVIEW_REQUIRED, all four review threads resolved, and no approving or change-requesting review ever submitted -- only comments (+15,087 / -914 across 108 files). Claims MJCF-importer joint parity with MuJoCo, floating-base observation and dataset coverage, runtime-vs-requested status reporting, and new get_contacts / apply_force / raycast / randomize / set_obs_noise / move_to surfaces, verified 26/26 on an A10G. Unchanged for a seventh sync and still the decision that is actually owed
#3400	fix(drivers/booster): read_mode passes GetMode its out-parameter	opened 2026-09-10T05:05Z by shipitfast, MERGEABLE, checks green, unarmed, CHANGES_REQUESTED since 05:34Z, no open threads (+41 / -4 across 3 files) -- the smallest diff on the board, and it has owed its author a round for six days
#3427	fix(drivers/crazyflie): the telemetry block fits the packet cflib allows	opened 2026-09-10T10:34Z by shipitfast, MERGEABLE, checks green, its thread resolved, an approving review recorded and armed 2026-09-13T02:08Z (+53 / -4 across 4 files) -- waiting only on the extra-approval rule
#3496	fix(policies/lerobot_async): rename_map renames the camera it declares	opened 2026-09-11T10:45Z by shipitfast, MERGEABLE, checks green, armed 17:21Z, both threads resolved, two approving reviews from one account -- and still REVIEW_REQUIRED (+81 / -10 across 4 files). The clearest thing on the board, and the one a second reviewer would close in a click
#3515	fix(docs): sim teleop fence passes robot_name to teleoperate only	opened 2026-09-11T20:15Z by shipitfast, MERGEABLE, checks green, both threads resolved, armed 2026-09-13T04:09Z (+121 / -2 across 3 files) -- same state as #3427
#3577	fix(policies/groot): groot_version is graded in service mode too	opened 2026-09-13T01:58Z by shipitfast, MERGEABLE, checks green, armed 2026-09-13T07:50Z, no open threads (+84 / -28 across 4 files)
#3615	fix(policies): lerobot_local names the constructor kwargs it ignores	opened 2026-09-14T02:20Z by shipitfast, MERGEABLE, checks green, armed 2026-09-14T08:31Z, its thread resolved (+56 / -1 across 3 files) -- cleared its CHANGES_REQUESTED and joined the extra-approval queue
#3640	fix(tools): download_assets grades unknown names and failed clones	opened 2026-09-14T17:26Z by shipitfast, MERGEABLE, checks green, unarmed, CHANGES_REQUESTED with one unresolved thread (+55 / -2 across 4 files) -- owed by its author for a second sync
#3673	fix(tools): lerobot_teleoperate action description names dagger	opened 2026-09-15T10:52Z by shipitfast, MERGEABLE, checks FAILING, unarmed, CHANGES_REQUESTED with one unresolved thread (+35 / -1 across 3 files) -- the only red head on the board, and owed by its author
#3714	feat(hardware_robot): a real arm can be read by an agent without running a policy on it	opened 2026-09-15T22:35Z, MERGEABLE, checks green, armed 22:36Z, both threads resolved, REVIEW_REQUIRED (+1,844 / -27 across 10 files)
#3726	feat(dashboard): the fleet, a simulated robot per thread, and an e-stop that reaches one still starting	opened 2026-09-16T01:45Z, MERGEABLE, checks green, armed 01:46Z, all fourteen review threads resolved, REVIEW_REQUIRED (+1,808 / -5 across 13 files) -- the follow-on to #3695's server
#3729	fix(docs): the Training data-loop fence opts into the remote-code gate	opened 2026-09-16T02:19Z by shipitfast, MERGEABLE, APPROVED, checks green, unarmed, no threads (+64 / -1 across 3 files)
#3737	fix(policies/lerobot_local): the camera pre-flight check reads the features the checkpoint declares	opened 2026-09-16T04:10Z, MERGEABLE, checks green, armed 04:11Z, its thread resolved, REVIEW_REQUIRED (+438 / -9 across 5 files)
#3739	feat(dashboard): a 3D twin drawn from the compiled model, with no physics in the browser	opened 2026-09-16T04:33Z, armed 04:35Z, checks pending, two unresolved threads, no review decision yet (+3,890 / -6 across 22 files) -- the largest head after #3343
#3743	fix(simulation): stop_policy answers once the robot is free	opened 2026-09-16T04:58Z, APPROVED, armed 04:58Z, no threads, checks FAILING (+410 / -10 across 8 files) -- approved and red, so it is owed a push
#3746	fix(simulation): a policy the constructor refuses is reported, not raised or lost	opened 2026-09-16T05:55Z, APPROVED, armed 06:22Z, checks pending, no threads (+371 / -10 across 5 files)
#3749	fix(docs,hardware_robot): the quickstart records a real arm with the tool that records, and the refusal names where each verb lives	opened 2026-09-16T06:31Z, MERGEABLE, APPROVED, armed 06:31Z, checks pending, no threads (+309 / -13 across 4 files)
#3751	fix(docs): the quickstart says what steps 4 and 5 need, not "everything else runs in sim"	opened 2026-09-16T07:20Z, APPROVED, armed 07:21Z, checks pending, no threads (+190 / -1 across 3 files)
#3752	fix(hardware_robot): a checkpoint provider with no checkpoint is refused before the arm is energized	opened 2026-09-16T07:35Z, APPROVED, armed 07:35Z, checks pending, no threads (+248 / -2 across 5 files)
#3753	fix(hardware_robot): a command the dispatcher would refuse is refused before the operator is asked	opened 2026-09-16T07:50Z, APPROVED, armed 07:50Z, checks pending, no threads (+198 / -2 across 5 files) -- the newest head, and the fourth hardware_robot one in a day

Seven of the twenty-one read APPROVED, ten wait on a second account, three are owed by their author and one (#3739) is owed a review decision. The two large contributor features remain asymmetric: #3205 is unblocked and #3343 still needs a decision. The new shape this band is a hardware-robot cluster -- #3714, #3749, #3752, #3753 -- which is where the queue's growth came from.

Closed / superseded notes

• The featureless run ends at nine bands, on volume rather than on one feature. 73 PRs, 58 fix + 7 feat + 4 test + 3 docs + 1 refactor, 277 files, +18,945 / -682 -- net +18,263 lines, the largest net-positive band this page has recorded. feat moves 144 -> 151 and correctness still climbs to 78.5%, because the fix count grew faster than the feature count even here. The one refactor is examples-scoped (#3709), so the net-negative refactor table holds at thirteen, the removals ledger (vera, lerobot_calibrate, motionbricks, strands_robots.transforms, the MolmoAct2 normalization port) is untouched for a fifth band, and the absence of a ! or a BREAKING CHANGE trailer on all 73 commits is correct rather than a gap.
• The dashboard finally has a server, and it fails closed. #3695 (19 files, +1,609 / -74) wires the core to the rails nine bands had accumulated -- auth, access, consent, safety_state, settings, log_redaction, validate_scope, agent_hitl -- taking strands_robots/dashboard/ eleven -> fifteen modules plus a served static/ bundle. It is the slice the #2848 decomposition argument was for, and #3726 and #3739 are already queued behind it. #3710 and #3738 are its correctness pass in the same band, and #3122 is still open naming what the slice owes.
• An unreadable reply names the peer, not the codec -- the band's dominant fix shape, and new in kind. #3670, #3672, #3676, #3677. Four clients decoded a peer's response inside a try reporting a codec failure, so a malformed or truncated reply from the far end was answered as a local decoding bug and sent the caller to the wrong side of the wire.
• A configuration entry that is not a string is refused. #3666 (closes #3664), #3667, #3674, #3679 -- the posture-flag lane read on the value side. #3666 is the consequential one: a DDS Security credential that was not a string was accepted and the participant QoS then dropped it, so a run believed it was authenticated.
• A motion that did not reach names what stopped it. #3703, #3705, #3727, #3733 -- the verified-halt question asked of a move rather than a stop. #3732 is the instructive part: the feat-shaped version of the same idea was closed unmerged in favour of the narrower fix #3733 plus a tracker item, #3734, for the rotate_wrist half -- a descope recorded rather than dropped.
• A camera name answers for the camera it names. #3683, #3691, #3693, #3730 -- a name the sim will list must be a name the sim will accept, once each, at every door.
• Seven features, six of them one theme: the library says more about the thing it just handed you. #3695, #3728, #3724, #3694, #3727, #3731, #3697.
• Sequencing item 6 is retired -- the 0000- fragment placeholder defect is fixed. #3715 refuses the number, #3721 gives that rule one owner so every surface refuses it, #3722 renames the one fragment that named no pull request. changelog.d/ holds zero placeholders, where it held four for three syncs, and --check reports OK (1600 pending).
• The tracker moved again, in both directions. #3664 and #3684 were filed and closed inside the band against the commits that answered them (#3666, #3686) -- a fifth clean pass -- while #3734 and #3744 were filed and remain open, so open issues went 3 -> 5. #3744 is the structurally interesting one: the offline-snapshot lane asked whether a copy of lerobot's vocabulary is still true; it asks whether one copy can be, across a pinned range whose releases disagree.
• The board grew by ten in a day to 21, its largest, and the constraint moved from the rule to the volume. Seven heads read APPROVED (#3729, #3743, #3746, #3749, #3751, #3752, #3753) where two did a day ago; ten are green, armed and thread-resolved on REVIEW_REQUIRED (#3205, #3343, #3427, #3496, #3515, #3577, #3615, #3714, #3726, #3737); three are owed by their author (#3400, #3640, #3673, the last of them red). #3343 is unchanged for a seventh sync.
• The clean sweep broke after two bands, and it broke on the duplicate shape. Four pull requests closed unmerged inside #3661 -> #3750, every one a second implementation of a fix that merged: #3665 against #3666, #3671 against #3672, #3736 against #3731, and #3732 descoped into #3733. Instances sixteen to nineteen of the shape #3169 measured, after two bands at zero. The counting streak also broke in the under-reporting direction: the tip is #3740 while #3750 is the highest merged number.
• The 2026-09-15 band bullets are now Finding/References rows below, and the 2026-08-26 -> 2026-09-14 ones were condensed on 2026-09-11, 09-13, 09-14 and 09-15. Every pull request named is tabled in full under ## Post-v0.5.1 and every band has a dated row in the log; the claims and cross-references are kept, the per-bullet prose is not.

Finding	References
A criterion already satisfied at reset is reported, not read as a result -- the third and sharpest reading of one question (2026-09-15). #3492 -> #3517 refused a value that could never be honored and #3584 -> #3620 refused a result the run never produced; these five ask what a surface publishes when the scene itself already satisfies what is being measured. #3629 is the widest in consequence -- measure_agreement's two fractions are accuracies over a class-imbalanced column, so a judge answering a constant scored the holdout's own class balance (0.95 at one-tagged-in-twenty, 1.00 at none), and the number deciding whether a judge may filter training data was unreadable off the number. #3637 and #3642 are the mirror pair: the same defect reads as an achievement on a stop_when clause and as a policy failure on a benchmark failure clause, both on step 1 with nothing moved	#3492, #3517, #3584, #3620, #3629, #3633, #3637, #3639, #3642
Four more lanes, and the largest test-pin band on the page (2026-09-15). A plan that commands nothing is a failure, not a plan (#3657's cuRobo joint-key list resolved once from the first waypoint, so the outcome depended on action_horizon); a ROS 2 name must be able to name a topic, and a refused call leaves no node behind; a reader grades the size its own payload declares (#3658's .spz num_points sized all five body blocks and was graded nowhere while the magic beside it was graded on the next line); a refusal applies nothing from its own call (#3645 rolled back only the property its refusal named). Eleven test commits, the most any band here has carried, each grading a door an existing suite reached only through a stand-in	#3621, #3622, #3625, #3626, #3627, #3628, #3632, #3634, #3635, #3636, #3638, #3641, #3643, #3644, #3645, #3646, #3647, #3648, #3650, #3651, #3652, #3655, #3656, #3657, #3658, #3659, #3660
The cleanest sweep the page had recorded, and a tracker that did not move at all (2026-09-15). Exactly one number inside #3621 -> #3660 was absent from main (the open #3640), none closed unmerged, and no issue was filed or closed -- a first for the wave. The extra-approval remedy entered routine use with #3661 and #3662 both APPROVED; rows also went to the offline-snapshot (#3623) and one-owner (#3624) tables, with #3630, #3631, #3649, #3653 and #3654 the rest	#3169, #3623, #3624, #3630, #3631, #3640, #3649, #3653, #3654, #3661, #3662
A run that commanded nothing is refused, not scored -- the outcome form of the lane #3492 -> #3517 opened (2026-09-14). Those refused a value that could never be honored at the door; these five refuse the result, where every argument passed and the run still applied nothing. #3620 is the sharpest -- success_rate, avg_reward and pass_hat_k published under status="success" for a rollout in which send_action was never called, with steps_advanced and actions_applied now separate numbers. #3608 is the one with an external premise: a source comment asserting that "LeRobot tolerates absent columns" where validate_frame refuses missing and extra features symmetrically, three lines apart	#3492, #3517, #3584, #3608, #3612, #3613, #3620
A shutdown that cannot complete does not keep the resource, and an offline snapshot names what the live registry registers (2026-09-14). The #3519 -> #3532 lane read from the teardown end -- that one asked what frees a thing whose start never finished, this asks what happens when the stop raises -- with #3596 the test half and the sharper finding: four cells asserting an rclpy ImportError assumed rclpy absent rather than establishing it, so on any host with a distro sourced they built a live node and left the context initialized. Beside it the dependency-floor lane's twin, where #3616's _FALLBACK policy-type snapshot was nine types behind and its own test graded the constant against itself, and #3618 is the misdirection case -- a gate refusing a non-native reward type and then describing the fields of the thing it had just said does not exist	#3519, #3532, #3590, #3596, #3602, #3603, #3614, #3616, #3618
A door refuses a collaborator it cannot use, and the board's remedy was proven (2026-09-14). #3617 (a robot_name the host cannot route to) and #3619 (a teleoperator the loop cannot poll, on the two surfaces that skipped the get_action check a third door already performed) each accepted, reported the session started, then failed every tick. And #3584 became the first head here to read APPROVED on a review from a second account and merged, so require_extra_approval_for_unattributed_changes is a queue rather than a deadlock	#3584, #3617, #3619
A posture flag is checked, not read by truthiness -- the lane opened 2026-08-20 finished across nine surfaces at once, and reached every policy provider (2026-09-13). On two of the nine the truthy branch is the permissive one: #3546 decides whether a Kimodo checkpoint executes its own code and #3574 whether WBC proceeds with models it could not find, so "false" had been selecting the unsafe posture. #3582 is the read-side twin -- a rover's sensor summary read a headlamp by truthiness, so a field the frame did not carry and a value the write door had already refused both rendered as off	#3538, #3543, #3546, #3548, #3549, #3553, #3554, #3574, #3575, #3582
Three lanes opened, each the next question after one already here (2026-09-13). An action naming a key the surface cannot carry is refused (#3534, #3537, #3539) sits one level below #3213: that fix refused an undeclared verb, these refuse an undeclared key inside a declared verb's payload, which had simply been dropped. A name is held to the alphabet its consumers can carry (#3557, #3558, #3559, #3565 closing #3562, #3570) is the containment question #3296/#3298 asked of a name -- a camera name reaches an MJCF attribute, a dataset column and an MQTT topic, each with its own grammar. A private name must have a reader (#3580, #3583) is a cross-cutting invariant with its own grader, and the shape #3578 was then filed against	#3213, #3296, #3298, #3534, #3537, #3539, #3557, #3558, #3559, #3562, #3565, #3570, #3578, #3580, #3583
The verified-halt lane reached robot_mesh in the form where the caller held the answer and did not read it (2026-09-13) -- every route reported success and audited ok=True over a delivered refusal, a timeout and a lockout alike. Beside it one documented camera-routing precedence became honoured on every router that implements it, #3555 the sharpest, an embodiment= caller's camera_key_map discarded outright so the model received no image input at all	#3545, #3550, #3555, #3560, #3561, #3566, #3568, #3571
The duplicate-implementation shape stopped being occasional and took five instances in one band (2026-09-13): #3536 against merged #3535, #3564 and #3569 both against merged #3565, #3567 against merged #3566, #3572 against merged #3571 -- instances eleven to fifteen of the shape #3169 measured, four inside 105 minutes on the evening of 2026-09-12, with issue #3562 drawing three implementations of one fix. #3169's negative result stands but was measured against roughly one instance per band. The same sync recorded the 0000- fragment placeholders going 2 -> 4 (#3534, #3537) instead of being renamed	#3169, #3534, #3535, #3536, #3537, #3562, #3564, #3565, #3566, #3567, #3569, #3571, #3572
A value that can never be honored is refused at the entry point (13 fixes, 2026-09-12's dominant shape) -- in domain, finite, correctly typed, past every existing guard, and still unable to produce what the caller asked for. #3502 is the sharpest: a benchmark clause naming an entity that cannot exist compiles, never fires, and the evaluation publishes success_rate: 0.0 beside success_measured: true; on a failure clause the same defect reads as 100% success. #3506's GR00T route accepted expert_only, reached no flag, and full-finetuned the backbone while reporting success	#3492, #3493, #3494, #3497, #3501, #3502, #3504, #3506, #3507, #3510, #3511, #3514, #3516, #3517
A resource a failed bring-up adopted is released -- the verified-halt lane's missing half, asking what frees a thing whose start never finished, because the success path is the only place it was adopted. Both init_device_connect_sync routes left a loop parked in run_forever on an unreachable thread; a Reachy teardown only stopped a loop it never closed; leaving the mesh returned while nine publishing loops were live. Every one is a leak no later verb can reach	#3519, #3522, #3526, #3530, #3532
The queue's constraint was the ruleset, not review latency, and this page had it wrong for eight syncs -- require_extra_approval_for_unattributed_changes beside required_approving_review_count: 1, so one account's approval could never carry a pull request; #3496 was green, armed, threads resolved, twice approved and still BLOCKED. Resolved in kind on 2026-09-13, when a second account approved #3584	#3496
The removals widened while the marking stopped (2026-09-11) -- motionbricks (58 files, -4,583), the augmentation-transforms package (32 files, -6,064, 1,943 lines with no importer) and the MolmoAct2 normalization port (-2,179) all left main with no ! and no BREAKING CHANGE trailer, where #3389/#3386 carried both one band earlier, so the tag's removals section has to be assembled by hand. motionbricks was the first of the four providers #3075's ruling named as going, with #3405 the counterpart -- the first GR00T change since the ruling. Whole band +24,826 / -26,659, net -1,833; #3485/#3488 were the same defect twice, list_objects and list_robots reporting where a thing was put rather than where it is	#3075, #3386, #3389, #3405, #3407, #3411, #3419, #3426, #3485, #3488
An operator is asked before a command reaches a real actuator, and the docs and the test tree lost 6,433 lines saying each thing once (2026-09-11) -- the HITL gate #3184 established was bypassed by four agent-callable surfaces that write to physical devices, and #3403 closes it from the other end by building the consent request from a refusal's code rather than its prose. Beside them #3417 replaced fourteen per-package copies of one docstring AST walk with a single linter rule (-1,663), #3469 cut README to 120 lines (-1,459), and seven docs pages plus three test consolidations went net-negative	#3184, #3403, #3417, #3418, #3420, #3430, #3431, #3440, #3463, #3468, #3469, #3470, #3471, #3472, #3483
The tracker drained to two and every closure reconciled against merged code (2026-09-11) -- #3398, the board's only P1 (a Feetech set_torque that never read the status packet a unicast WRITE returns, so a servo that had gone quiet read as one that complied) closed against merged #3407 about 27 hours after filing, and #3146 (break the loader/user_registry cycle or accept py/cyclic-import) against merged #3480, which took the third option and made the package import in a DAG. Survivors #3122 (P2) and #3099 (P3). The 2026-09-12 pass held it: #3531 filed 04:44Z, closed 06:18Z against merged #3532, with two fragments merged under the reserved 0000- placeholder (#3526, #3529) the one new bookkeeping defect	#3099, #3122, #3146, #3398, #3407, #3480, #3526, #3529, #3531, #3532
FastTD3 reached main (+2,199 / -95, 29 files), the first of epic #858's off-policy deliverables and the first code answer to #2720 -- closing neither, its two fragments named for the issues rather than the PR	#858, #2720, #3206
A verdict names the thing it was written for, new in kind: a session verdict against the process its record was written for, its pid-reissue residue, a stand-in the prune never consults, an already-merged PR read as a pending mergeability, an unreviewed PR owed to whoever was nearest, a HITL grant naming a neighbouring call	#3110, #3219, #3227, #3231, #3233, #3237, #3248, #3250, #3254
Nine more shared value domains, two of them a value refused on one field and accepted on its sibling, one a non-finite wire timestamp four mesh freshness gates accepted	#3238, #3239, #3241, #3242, #3243, #3244, #3247, #3249, #3251
A destructive step waits for the last refusal: a refused start_recording had already wiped the LeRobotDataset it was replacing; reset() flushes the open episode on every backend; the same ordering asked of a trainer's output_dir	#3196, #3198, #3245
The floor lane read from the other end -- a test pinned an np.cross behaviour numpy 2.5 removed, so the upper half of the declared range was red	#2335, #2342, #2345, #2349, #3229, #3234
Duplicate work, instances one to five, all measured and none caught by a documented key: two implementations of one README line, of one script fix, and of one -x subtraction that both merged -- #3169 is the negative result (widening the key selects ~27 pairs across 9 issues for 2 genuine ones)	#3132, #3135, #3136, #3164, #3166, #3167, #3169, #3170, #3171, #3172, #3189, #3192, #3194, #3195, #3220, #3231, #3232, #3233
The features moved off main into the queue -- four feat heads from one outside contributor in ninety minutes, including the mesh_bridge slice #2977 asked for and the mesh peer-retention head	#858, #2720, #2848, #2977, #3189, #3205, #3206, #3212, #3220
An undeclared agent verb ran the halt on eleven of the twelve shipped drivers -- a typo, a verb borrowed from a sibling, or a non-string stopped the robot under a success-shaped answer; drivers.base now owns declared_verbs and the refusal that lists them	#3213
The cache-validity lane, seven surfaces, shape moving from the source changed underneath the entry to the key does not name what the entry is about -- the mesh cache reporting an absence it had created, and a MinkIKBridge keyed on nothing at all	#3141, #3142, #3144, #3145, #3147, #3173, #3177, #3214, #3220
#2907's arithmetic: the "Update branch" button is a push, so its sole approval was spent twice; #3197 measured the eight review rounds a base absorption costs and why the resulting BLOCKED reads as latency. An off switch that still published one step closed beside it	#2907, #3173, #3175, #3176, #3178, #3180, #3190, #3191, #3197
The dashboard went from a decomposition argument to the wave's feature lane: #2878 restated as #2977, both extracted slices merged and hardened, the 481-file draft #2848 closed with no check run in its life one minute after #3122 named what it carried, then a HITL gate over hardware-moving tool calls, a Lockout type, an honest preflight, strands-agents>=1.13.0 and the written-bound guard	#2335, #2342, #2345, #2349, #2848, #2878, #2977, #3050, #3051, #3078, #3080, #3122, #3184, #3185, #3186, #3188
The wave's only perf commit: coverage on the PEP 669 core, +8.2% against +76.6%, identical reading per file, statement and exclusion -- buying room on #3143's bound rather than deciding it	#3143, #3149
The whole-tree graders were audited and found under-reading: a hand-maintained roster, four unlearned spellings of a walk root, a pacing sweep grading only tickers spelled Ticker, and guards grading the working directory -- every one a false pass. The band that followed had every PR number on main, its seven gaps all issues	#3092, #3094, #3095, #3099, #3105, #3106, #3107, #3111, #3112, #3113, #3118, #3119, #3120, #3122, #3125, #3126, #3128, #3130
Closures that reconciled -- five against a merged PR, and the two this page called closeable resolved as written	#2833, #2895, #2935, #2940, #2963, #2981, #3060, #3062, #3063, #3064, #3066, #3079, #3080, #3086, #3087
The review queue drained 17 -> 5 in a day and the native-driver programme doubled to twelve buses across 25 registry robot names, its verbs shipped in the consolidated shape from the start; what it kept carrying was two hardware items and two design ones	#2734, #2739, #2749, #2750, #2753, #2757, #2762, #2763, #2765, #2767, #2779, #2786, #2812, #2836, #2891, #2928, #3037, #3040, #3043, #3045, #3046, #3050, #3051, #3053, #3058, #3060, #3063, #3070, #3081, #3084
Every DDS endpoint is constructed under the shared lock -- filed against the Go2 driver the day it landed, then closed as a class rather than a site, with the lock vocabulary derived from every holder	#3039, #3062, #3064, #3067, #3071, #3072, #3073
A stop reports its outcome instead of asserting one -- the wave's longest-running idiom, opened at eight sites and reaching the drivers, the rollout entry point, the training sessions and the inference server	#2809, #2815, #2816, #2821, #2825, #2826, #2828, #2831, #2832, #2833, #2834, #2835, #3060, #3091, #3092, #3100, #3104, #3108, #3110, #3114
#2765 closed against merged code (the motion-switcher decoder plus the FSM-producer wiring answered its five rt/lowcmd decisions), promoting the read-cadence question; and the 81-PR neon-the-g1 verb port peaked at 79 modules before consolidating to 22	#2765, #2888, #2891, #2916, #2946, #2948, #2950, #2953, #3037, #3065, #3070, #3089
Three net-negative refactors merged and a fourth was refused -- the vendored LIBERO adapter gone, 92 g1 lookup verbs folded into one dispatcher, 13 execution modules into one table -- roughly 76,000 lines, with the GR00T removal stopped by an argument about public surface	#3037, #3047, #3056, #3070, #3075
The 2026-08-24 tracker pass: sixty issues closed in one 10-minute window, all COMPLETED, 57 with no linked PR -- the pass #2720 was filed against, which is why every count on this page is reported rather than endorsed. Three of the closes do reconcile against merged code	#858, #859, #873, #1537, #1905, #2062, #2145, #2179, #2221, #2315, #2330, #2337, #2459, #2460, #2507, #2685, #2720
Both long-stalled external PRs drained by re-landing rather than rebase, each original closed within minutes of its replacement	#1035, #1722, #1905, #2689, #2696
The duplicate/overlap merge gates narrowed three times in one band; the residue is the concrete failure that motivated them	#2762, #2769, #2774, #2791, #2797, #2830
The Isaac asset-reader sweep (22 PRs) -- MJCF and URDF attributes read the way the compiler resolves them, not literally at the site that declares them; LIBERO's box proxies replaced by real meshes; parity itself closed out	#2156, #2221, #2233, #2498, #2656, #2712
The dataset-transform / VLM-judge batch shipped in full within five days, plus its own correction pass and the continuable refusal that carries a code and its subject	#2480, #2486, #2508, #2610, #2612, #2616, #2617, #2618, #2623
#1087's scope shipped broadened -- a publish blocklist became a HITL gate over every verb that can command a robot -- and #1087 closed as superseded, leaving the halt-vs-drive contract tracked separately	#1087, #2313, #2315
A dependency-floor audit closed five floors in one window, each raised to the release that ships the imported capability, plus the one that cannot be floored because it is not on PyPI -- the #1930 class v0.5.1 was cut for	#1930, #2335, #2340, #2342, #2345, #2349
A monotonic-clock lane closed six published budgets measured by subtracting two time.time() readings, and filed its own remainder	#2404, #2406, #2408, #2412, #2413, #2414, #2425
The photoreal 3DGS + GS-lighting batch landed whole, closing four of its five tracking issues plus one more, with its correction pass inside the same wave and the gallery and parent closed later	#1537, #2321, #2322, #2323, #2324, #2325, #2326, #2328, #2331, #2332, #2336, #2337, #2441, #2443, #2446, #2452
Two lanes closed their last subtask -- the Isaac motion-primitive parity suite, and the fleet example epic at 5/5	#2156, #2179, #2181, #2198, #2233
ProtoMotionsPolicy took its post-merge hardening pass, no public-surface change	#2287, #2411, #2416, #2417, #2418
The six CodeQL py/uninitialized-local-variable alerts cleared, and the Newton set_timestep parameter-name fix landed	#2308, #2359
The removal direction reached shipped public surface, on a ruling rather than an argument (2026-09-10 band): two refactor! commits, vera deleted (132 files, +424 / -15,204; uv.lock 258 -> 237 packages, the five [tool.uv] conflicts rows gone, extras 34 -> 33) and lerobot_calibrate deleted (22 files, +227 / -2,313, 859 LOC that never calibrated anything -- LeRobot ships four console scripts that do). Neither amputated a shared piece: discover_ee_frame stayed in simulation.ik with its three-rung ladder pin moved to the shared home, three shared guards were re-anchored on the surviving providers, and both removals took their own changelog.d fragments with them as the LIBERO one had. What licensed them was #3075, closed NOT_PLANNED 2026-09-10T04:21:21Z with the decision in words -- "GR00T stays. Service mode and the container tool have no upstream replacement" -- keeping GR00T/MolmoAct and dropping lerobot_async, curobo, moveit2, motionbricks. So #3047's refusal produced a ruling rather than a stalemate. Whole band 327 files, +15,052 / -25,581, net -10,529; wave removals ~101,000 lines across seven refactors at that point	#3047, #3056, #3075, #3386, #3389, #3411
The tracker emptied a second time, and #2720 was closed by the pattern it was filed about: thirteen issues closed 04:21:02Z -> 05:16:27Z, seven NOT_PLANNED, five COMPLETED with no closer, one with a closer -- #2796 via merged #3395, the required check now evaluating the merge commit rather than the branch head, retiring the oldest CI item on the page. #2720 ("57 of 60 issues closed COMPLETED with no linked PR inside one 10-minute window") closed NOT_PLANNED, with no linked pull request, inside a ten-minute window. The distinction that matters: NOT_PLANNED is a decision and 08-24's COMPLETED was a claim, so seven of the thirteen are a different act from the pass #2720 described -- and five are not. #858, the last unscheduled epic, closed one second after it, with FastTD3 (#3206) the one deliverable of the three that reached main. Both hardware-blocked items went (#2753, #2812), so nothing on the page waits on a physical robot. Open issues 16 -> 4, and the survivors carried priority labels for the first time, including the new P1 #3398, whose remedy was already approved as #3407	#858, #2720, #2753, #2791, #2796, #2812, #2891, #2928, #2977, #3075, #3143, #3169, #3197, #3206, #3395, #3398, #3407
The clean sweep broke after two bands, and the shape that broke it is the one whose record closed the same morning: #3361 and #3370 closed unmerged inside #3352 -> #3406, each a second implementation of a fix that merged as #3360 and #3368 -- the seventh and eighth instances of the shape #3169 measured, on the day #3169 itself closed NOT_PLANNED. #3361 is the sharper: it added the boolean check to lerobot_calibrate's overwrite, and #3386 deleted the tool, so its subject no longer exists. That is also how #3356 finished -- six PRs checked the posture flags on the surfaces an agent calls, and the seventh row resolved by deletion	#3169, #3352, #3355, #3356, #3360, #3361, #3364, #3365, #3367, #3368, #3370, #3386, #3406
Both contributor features resolved exactly as the previous sync predicted: #2907 closed unmerged 2026-09-09T07:08Z with its observer lane already on main as #3351, and #3212 closed unmerged at 07:24Z once #3352 landed its payload cherry-picked and attributed -- one pull request whose work had shipped and one decision on two heads, both bookkeeping rather than review rounds. The queue that replaced them was thirteen heads, four APPROVED and armed, with #3343 promoted out of draft at 106 files with all four threads resolved	#2907, #3212, #3343, #3351, #3352

Pre-v0.5.0 release and cluster notes (condensed 2026-09-10, further 2026-09-14). Every claim below is carried by the ## Published releases table and the mermaid above, so what is kept here is the roster. Motion-planner epic #415 closed (cuRobo #442, MoveIt2 #305, both in v0.4.0). The v0.3.9 mesh-security scope shipped folded into v0.4.0; its nine consolidation umbrellas (#386-#394) and their follow-ups stay closed. v0.4.1 (2026-07-01) collapsed [Unreleased] into a dated heading and opened a fresh one via #977 -- a collapse never repeated, which is why one [Unreleased] section now spans two shipped tags. v0.5.0 (2026-08-04T01:32Z) shipped as a minor, the planned v0.4.2 patch never cut, while its milestone still carried 8 open issues, so the tag and milestone boundaries did not agree; v0.5.1 (2026-08-06T20:23:39Z) was a correctness-only patch answering a resolution gap -- stream_dataset(repo_type="bucket") needs the >=0.6.1 constructor that v0.5.0's >=0.6.0 floor did not guarantee (#1930) -- plus the [wbc] huggingface_hub>=1.5.0 floor (#1957). Inside those waves: remote policy inference as strands_robots.inference (#1142); the Isaac Sim migration epic #1144 closed via #1156, with #1151/#1153/#1154/#1155/#1205 closing #1146-#1150 and the vendor subtask #1145; the first runnable locomotion benchmark #1259 over #1239 and #1255, joined by #1287/#1288 and completed on all three command axes by #1321/#1324; the terrain curriculum's four heightfield kinds #1336/#1338/#1339/#1340 plus the difficulty knob #1344 (tool_spec-discoverable #1350) against the #873 groundwork; the security HIL gate #1085 merged out of order; and #1360 -> #1389 adding public get_ground_height (#1388), terrain-relative predicates (#1364/#1368), floating-base robots seated on terrain (#1386), Isaac/Newton base-contract rejects (#1360/#1372), the multi-robot keyframe fix (#1383), fail-soft facade guards (#1365/#1370) and the README/notebook docs run (#1374, #1375, #1376, #1378, #1379, #1387). The short-lived planning (#853) and sim_managers (#878) packages were both removed (#889/#890) inside their own wave, so neither added net surface. #586 merged 2026-07-15 and #1018 was superseded by #1402 (fixing #1017). This roadmap closed as completed 2026-08-14T02:17Z, in the grooming pass that also closed #467, #468 and #475 within 78 minutes and took 0.5.0 from 3 closed / 8 open to 7 / 4; the roadmap label carries no successor issue, so tracking moved to the milestone and the open-issue set. #1087 (superseded 2026-08-19) and #1035 (2026-08-24, after #2689 landed its scope) were the last long-open pair, while the robots-sim examples-migration cluster landed (#1278, #1280, #1282 merged; #1276, #1279 closed), the mesh IoT-transport test #1347 closed 2026-07-23, #1110 merged 2026-07-31 and #1000 closed unmerged 2026-08-02.

2026-07-09 -> 2026-07-22 (thirteen daily reconciliations, condensed 2026-09-06, rostered 2026-09-14). Thirteen runs against main while v0.4.1 was still the latest tag, folding the post-v0.4.1 [Unreleased] wave forward from +234 to +516 PRs, with milestone 0.5.0 reading 3 closed / 8 open throughout. Every pull request they enumerated is tabled in full under ## Post-v0.4.1 above, and the arc they describe -- the Isaac backend migration, the reward/predicate DSL becoming runnable, the humanoid and omnidirectional benchmarks, the terrain curriculum, the describe() expansion, the NumPy-scalar and finite-value validation sweeps, and the getting-started example series -- is carried by that section's own theme tables. Markers, in order: 07-09 @ #1209 (+234); 07-10 @ #1238 (+263); 07-11 @ #1263 (+288); 07-12 @ #1307 (+319); 07-13 @ #1334 (+346); 07-14 @ #1359 (+370); 07-15 @ #1389 (+400); 07-16 @ #1418 (+427); 07-17 @ #1442 (+447); 07-18 @ #1458 (+465); 07-19 @ #1478 (+485); 07-21 @ #1497 (+504); 07-22 @ #1530 (+516). Every number those entries named, kept as a roster: #586 #1017 #1018 #1085 #1144 #1146 #1150 #1151 #1153 #1154 #1155 #1156 #1183 #1187 #1198 #1203 #1205 #1208 #1209 #1212 #1215 #1223 #1230 #1233 #1236 #1238 #1239 #1249 #1254 #1255 #1256 #1257 #1259 #1261 #1262 #1263 #1264 #1266 #1267 #1268 #1270 #1275 #1277 #1278 #1280 #1281 #1286 #1287 #1288 #1291 #1298 #1299 #1303 #1306 #1307 #1308 #1318 #1321 #1322 #1324 #1327 #1330 #1331 #1334 #1336 #1337 #1338 #1339 #1340 #1344 #1346 #1348 #1350 #1352 #1354 #1357 #1359 #1360 #1364 #1365 #1368 #1370 #1372 #1374 #1375 #1376 #1378 #1379 #1383 #1386 #1387 #1388 #1389 #1390 #1391 #1393 #1394 #1395 #1396 #1398 #1401 #1402 #1405 #1407 #1409 #1410 #1416 #1417 #1418 #1419 #1420 #1421 #1422 #1423 #1424 #1425 #1426 #1427 #1428 #1429 #1430 #1431 #1432 #1433 #1434 #1435 #1436 #1437 #1438 #1439 #1440 #1441 #1442 #1443 #1444 #1445 #1446 #1447 #1448 #1449 #1450 #1451 #1452 #1453 #1454 #1455 #1456 #1457 #1458 #1459 #1460 #1461 #1462 #1463 #1464 #1465 #1466 #1467 #1468 #1469 #1470 #1471 #1472 #1473 #1474 #1475 #1476 #1477 #1478 #1479 #1480 #1481 #1482 #1483 #1484 #1485 #1487 #1488 #1489 #1490 #1491 #1492 #1493 #1494 #1495 #1496 #1497 #1498 #1499 #1503 #1505 #1508 #1509 #1511 #1512 #1513 #1515 #1517 #1518 #1519 #1520 #1521 #1522 #1523 #1524 #1525 #1526 #1527 #1528 #1529 #1530 #1531.

2026-08-04 / 08-07 / 08-08 (three release-state correction passes, condensed 2026-09-06). None was a full reconciliation, so Last synced stayed at @ #1932 throughout. 08-04 recorded the new v0.5.0 tag (published 2026-08-04T01:32Z, superseding the v0.4.2 patch that was never cut -- an earlier revision of that entry had said "no new tag", written ~9h after the release shipped), corrected the Open-PRs table to the three genuinely open (#1722, #1087, #1035) after #1596 and #1110 merged and #1559 / #1000 closed unmerged, and dropped #1145 (closed 2026-07-11) and #1109 from the no-milestone table. 08-07 found v0.5.1 published 2026-08-06T20:23:39Z and unmentioned anywhere on the page -- the header had asserted v0.5.0 was latest for ~14h -- so it corrected the header and added the missing v0.5.0 / v0.5.1 rows to the Published-releases table, which had stopped at v0.4.1. 08-08 fixed four statements that were false rather than stale: #1145 described as open in two places, nine of eleven items in a "still open" bullet already merged or closed, the robots-sim examples migration described as in flight, and the ## Post-v0.4.1 heading still targeting the uncut v0.4.2. Every defect drifted the same direction -- work was further along than the document admitted. All three deliberately left the dated blocks above untouched (true on their dates) and re-verified milestone 0.5.0's "3 closed / 8 open" against the bar's 6 (3 closed issues + 3 closed PRs = 6/14 = 42.857%, both correct at different scopes). Live drift they recorded for the next full sync: merged tip #2000, then #2034, i.e. 50 and 73 PRs beyond the marker (#2001 was open but unlisted at 08-07; #2028 and #2034 both opened and merged inside one morning, so neither was ever listable); and a known gap left unpatched -- the no-milestone table listed 3 issues while 32 were live.

2026-08-16 -> 2026-08-28 (eight daily reconciliations, condensed 2026-09-07, rostered 2026-09-14). Eight runs against main while v0.5.1 was still the latest tag, carrying the post-v0.5.1 wave from +226 to +690 PRs. 08-16 was the final reconciliation taken after this roadmap closed on 2026-08-14; 08-26 is the one that mattered -- milestone 0.5.0 went 7 closed / 4 open -> 11 / 0 (100%), open issues 78 -> 4, and #2720 was filed against exactly that pass, which is why every count on this page is reported rather than endorsed. The theme sections those entries opened -- the native hardware-driver seam, one drive contract for every mobile base, the Isaac asset-reader run, the shared-value-domain sweep, the dataset-transform and VLM-judge batch, a stop reporting its outcome, refuse-what-cannot-be-honored, and the docs/asset/remote-inference lanes -- are all tabled in full under ## Post-v0.5.1 above. Markers, in order: 08-16 @ #2307 (+226); 08-17 @ #2387 (+282); 08-18 @ #2426 (+318); 08-19 @ #2456 (+350); 08-20 @ #2521 (+381); 08-26 @ #2748 (+581); 08-27 @ #2836 (+650); 08-28 @ #2879 (+690). Every number those entries named, kept as a roster: #467 #468 #475 #1035 #1087 #2122 #2123 #2145 #2198 #2233 #2268 #2269 #2307 #2310 #2363 #2367 #2368 #2378 #2387 #2388 #2390 #2391 #2392 #2393 #2396 #2401 #2402 #2403 #2407 #2410 #2414 #2415 #2420 #2421 #2422 #2423 #2426 #2427 #2433 #2439 #2440 #2451 #2453 #2454 #2456 #2457 #2458 #2461 #2463 #2464 #2465 #2466 #2467 #2468 #2469 #2473 #2474 #2475 #2476 #2478 #2479 #2481 #2483 #2485 #2487 #2488 #2489 #2491 #2492 #2493 #2494 #2497 #2499 #2500 #2501 #2503 #2504 #2510 #2511 #2515 #2516 #2517 #2521 #2522 #2685 #2720 #2746 #2748 #2749 #2763 #2765 #2771 #2791 #2792 #2793 #2794 #2795 #2796 #2798 #2801 #2812 #2823 #2829 #2830 #2833 #2836 #2848 #2856 #2862 #2874 #2879.

2026-08-31 -> 2026-09-02 (three daily reconciliations, condensed 2026-09-07, rostered 2026-09-14). The three consecutive runs that took the wave from +841 to +910 PRs while v0.5.1 stayed the latest tag, and between them they hold the wave's turning point in both directions. 08-31 was the largest band this page recorded, 151 PRs, and it inverted the wave's character -- 84 feat against 48 fix took feat 36 -> 120 and the correctness share 81.7% -> 73.5%, but by concentration, since 81 of the 151 were one programme, the neon-the-g1 operator verbs at their 79-module peak; it also opened the net-negative refactor lane, the doctor lane and the environment-variable lane, and left open PRs at 17, the largest queue recorded. 09-01 drained it to 5, doubled the native-driver programme to twelve buses across 25 registry robot names, landed both g1 consolidations (79 modules -> 22) and refused #3047 one minute after #3075 was filed against its premise. 09-02 was the most correctness-dominated band to that point and closed the 481-file dashboard draft #2848 one minute after #3122 was filed naming what it still carried; it also opened the whole-tree-grader, refuse-a-span-that-resolves-to-no-work and no-reading lanes, and was the first band in which every pull-request number inside it was on main. Every pull request they enumerated is tabled in full under ## Post-v0.5.1 above. Markers, in order: 08-31 @ #3057 (+841); 09-01 @ #3091 (+877); 09-02 @ #3130 (+910). Every number those entries named, kept as a roster: #2765 #2833 #2848 #2888 #2891 #2893 #2896 #2906 #2907 #2908 #2909 #2911 #2916 #2945 #2952 #2963 #2979 #2981 #2985 #2990 #3003 #3009 #3025 #3037 #3038 #3039 #3040 #3042 #3043 #3045 #3046 #3047 #3050 #3051 #3053 #3054 #3055 #3056 #3057 #3058 #3060 #3062 #3063 #3064 #3066 #3067 #3070 #3071 #3072 #3073 #3074 #3075 #3076 #3077 #3078 #3079 #3080 #3081 #3082 #3083 #3084 #3085 #3086 #3087 #3090 #3091 #3092 #3094 #3095 #3097 #3099 #3100 #3103 #3104 #3105 #3106 #3107 #3108 #3109 #3110 #3112 #3113 #3114 #3116 #3118 #3119 #3120 #3121 #3122 #3123 #3124 #3125 #3126 #3128 #3129 #3130 #3131 #3132 #3133 #3134.

2026-09-03 -> 2026-09-05 (three daily reconciliations, condensed 2026-09-08, rostered 2026-09-15). The three runs that took the wave from +944 to +979 PRs while v0.5.1 stayed the latest tag. 09-03 was the third featureless band (34 PRs, one out-of-order merge #3009 closing #3007; share 73.2% -> 73.5%) and the band whose structural event was duplicate work measured on itself -- #3164 named the -x subtraction, #3167 and #3166 answered it 50 minutes apart, and #3169 filed the negative result that widening the claim key selects ~27 pairs across 9 issues for 2 genuine ones; it opened the cache-validity lane (#3144's mechanism) and the run-that-stopped-early lane (with #3149, the wave's only perf commit, +8.2% against +76.6%), and took open PRs to 2, the smallest queue recorded. 09-04 ended the featureless run with three dashboard features (#3184 the HITL gate, #3185 Lockout, #3186 the honest preflight, six modules -> ten), raised strands-agents to >=1.13.0, and held open issues at 16 in number and membership for the first time. 09-05 was the fourth featureless band and the one where the features moved off main into the review queue (open PRs 2 -> 9, four feat, three opened inside ninety minutes by one outside contributor), its most consequential merge the safety fix #3213, and it opened the undeclared-agent-verb, one-owner and keeps-the-dataset lanes. Every pull request they enumerated is tabled in full under ## Post-v0.5.1 above, and the themes those days opened each have a section there. Markers, in order: 09-03 @ #3172 (+944); 09-04 @ #3191 (+959); 09-05 @ #3220 (+979). Every number those entries named, kept as a roster: #858 #2335 #2342 #2345 #2349 #2720 #2848 #2907 #2977 #3007 #3009 #3132 #3134 #3135 #3141 #3142 #3143 #3145 #3147 #3148 #3164 #3166 #3167 #3170 #3171 #3173 #3175 #3176 #3177 #3178 #3180 #3183 #3187 #3188 #3189 #3190 #3192 #3193 #3194 #3195 #3196 #3197 #3198 #3199 #3200 #3202 #3203 #3204 #3205 #3206 #3207 #3208 #3209 #3210 #3211 #3212 #3214 #3216 #3218 #3220 #3222.

2026-09-06 -> 2026-09-09 (four daily reconciliations, condensed 2026-09-09/10, rostered 2026-09-15). The three runs that took the wave from +1,011 to +1,072 PRs while v0.5.1 stayed the latest tag. 09-06 folded in 32 PRs with four out-of-order merges and landed #3206, FastTD3 (+2,199 / -95 across 29 files) -- the first of epic #858's off-policy deliverables to reach main and the first of the three #2720 reported absent to become code rather than argument -- declaring no closing reference, so the bookkeeping stayed owed; it opened a verdict names the thing it was written for, and its two fragments were numbered for the issues they answer, a first. 09-07 was the fifth featureless band (26 PRs; correctness past three-quarters at 75.0%), almost end to end one idiom, whose sharpest form is two owners of one value, and it opened a dialled address has two halves. 09-08 was the sixth and largest featureless band (35 PRs; 75.7%), its idiom the shared domain read one step earlier -- a value is graded before it is converted -- and it opened three lanes including a packaging declaration reaches the install that needs it; its one non-reconciling closure (#2853, closed with no linked pull request 42 minutes before #3297 merged the nearest code to it) is recorded against #2720, and the clean sweep returned. Every pull request they enumerated is tabled in full under ## Post-v0.5.1 above, and the themes those days opened each have a section there. 09-09 followed, 27 PRs with no out-of-order merge, ending the featureless run at six bands on a single feature that landed by rebase rather than by review -- #3351 carrying #2907's observer lane byte-identically with authorship preserved, which made #2907 closeable as superseded and #3212 the same offer as #3352; it opened a store is written whole, a dataset root resolves to the directory its repo_id recorded to and a seeded replay reads the forwarded seed, and #3343 arrived as the largest head on the board since #2848 closed. Markers, in order: 09-06 @ #3254 (+1,011); 09-07 @ #3284 (+1,037); 09-08 @ #3319 (+1,072); 09-09 @ #3351 (+1,099). Every number those entries named, kept as a roster: #3320 #3321 #3327 #3340 #858 #1930 #2720 #2853 #2907 #3007 #3054 #3136 #3169 #3170 #3171 #3183 #3188 #3194 #3196 #3197 #3202 #3204 #3205 #3206 #3212 #3215 #3217 #3219 #3221 #3222 #3223 #3224 #3226 #3227 #3229 #3230 #3231 #3232 #3233 #3234 #3235 #3236 #3237 #3238 #3239 #3241 #3242 #3243 #3244 #3245 #3247 #3248 #3249 #3250 #3251 #3252 #3253 #3254 #3255 #3256 #3257 #3258 #3259 #3260 #3261 #3262 #3263 #3264 #3265 #3266 #3267 #3268 #3269 #3270 #3271 #3272 #3273 #3274 #3275 #3276 #3277 #3278 #3279 #3280 #3281 #3282 #3283 #3284 #3285 #3286 #3287 #3288 #3289 #3290 #3291 #3292 #3293 #3294 #3295 #3296 #3297 #3298 #3299 #3300 #3301 #3302 #3303 #3304 #3305 #3306 #3307 #3308 #3309 #3310 #3311 #3312 #3313 #3314 #3315 #3316 #3317 #3318 #3319.

2026-09-10 -> 2026-09-14 (five daily reconciliations, condensed 2026-09-11 to 2026-09-15, rostered 2026-09-16). The five runs that took the wave from +1,144 to +1,339 PRs while v0.5.1 stayed the latest tag. 09-10 is where the removal direction first reached shipped public surface behind a written ruling -- two refactor! commits, #3389 dropping the vera provider (132 files, +424 / -15,204; uv.lock 258 -> 237 packages, extras 34 -> 33) and #3386 dropping lerobot_calibrate (-2,313, 859 LOC that never calibrated anything), licensed by #3075 closing NOT_PLANNED with the direction in words, "GR00T stays" -- and where the tracker drained 16 -> 4 in one 55-minute window: seven NOT_PLANNED, five COMPLETED with no closer, exactly one with a closer (#2796 via merged #3395), closing #2720 by the pattern it was filed about and #858, the last unscheduled epic, one second later. It also opened a telemetry field the frame does not carry is not a reading and a removed surface takes its own history with it. 09-11 was the largest band since 08-31 (80 PRs) and the one where the removals widened while the marking stopped -- motionbricks (#3419), the augmentation-transforms package (#3426, 1,943 lines with no importer) and the MolmoAct2 normalization port (#3411) all left main with no ! and no BREAKING CHANGE trailer, so the tag's removals section must be assembled by hand -- while the docs and the test tree lost 6,433 lines saying each thing once (#3417 replacing fourteen per-package copies of one AST walk, #3469 taking README to 120 lines) and, for the first time in the wave, every tracker closure reconciled against merged code (#3398 via #3407, #3146 via #3480). It opened an operator is asked before a command reaches a real actuator. 09-12 was the first band here that removed nothing (+10,234 / -483) and refused a value that can never be honored at the entry point at thirteen sites, #3502 the sharpest -- success_rate: 0.0 published beside success_measured: true. 09-13 finished the posture-flag lane across nine surfaces including every policy provider, and took the hardest clean-sweep break recorded: five PRs closed unmerged, each a second implementation of a fix that merged, four inside 105 minutes. 09-14 was featureless and refactor-free, refusing a run that commanded nothing (#3620, success_rate under status="success" for a rollout in which send_action was never called) and a shutdown that cannot complete (#3596 the test half). Every pull request they enumerated is tabled in full under ## Post-v0.5.1 above, and the lanes they opened each have a section there. Markers, in order: 09-10 @ #3406 (+1,144); 09-11 @ #3488 (+1,224); 09-12 @ #3533 (+1,264); 09-13 @ #3583 (+1,304); 09-14 @ #3620 (+1,339). Every number those entries named, kept as a roster: #858 #1048 #1690 #1987 #2720 #2753 #2791 #2796 #2801 #2812 #2853 #2891 #2907 #2928 #2977 #3075 #3099 #3122 #3126 #3143 #3146 #3169 #3197 #3205 #3212 #3213 #3296 #3298 #3343 #3349 #3352 #3353 #3355 #3356 #3359 #3360 #3361 #3362 #3364 #3365 #3367 #3368 #3369 #3370 #3371 #3373 #3374 #3375 #3376 #3378 #3379 #3380 #3381 #3382 #3383 #3386 #3387 #3389 #3390 #3392 #3393 #3394 #3395 #3396 #3397 #3398 #3399 #3400 #3401 #3403 #3404 #3405 #3406 #3407 #3408 #3409 #3410 #3411 #3412 #3414 #3417 #3418 #3419 #3420 #3423 #3426 #3427 #3429 #3430 #3431 #3432 #3433 #3434 #3435 #3436 #3438 #3439 #3440 #3441 #3442 #3443 #3444 #3445 #3446 #3449 #3450 #3451 #3452 #3453 #3455 #3457 #3459 #3460 #3461 #3462 #3463 #3464 #3465 #3466 #3467 #3468 #3469 #3470 #3471 #3472 #3473 #3475 #3476 #3478 #3479 #3480 #3482 #3483 #3485 #3488 #3489 #3490 #3491 #3492 #3493 #3494 #3495 #3496 #3497 #3498 #3499 #3500 #3501 #3502 #3503 #3504 #3505 #3506 #3507 #3508 #3509 #3510 #3511 #3512 #3513 #3514 #3515 #3516 #3517 #3518 #3519 #3520 #3521 #3522 #3523 #3524 #3525 #3526 #3527 #3528 #3529 #3530 #3531 #3532 #3533 #3534 #3535 #3536 #3537 #3538 #3539 #3540 #3541 #3542 #3543 #3544 #3545 #3546 #3547 #3548 #3549 #3550 #3551 #3552 #3553 #3554 #3555 #3556 #3557 #3558 #3559 #3560 #3561 #3562 #3563 #3564 #3565 #3566 #3567 #3568 #3569 #3570 #3571 #3572 #3573 #3574 #3575 #3576 #3577 #3578 #3579 #3580 #3581 #3582 #3583 #3584 #3585 #3586 #3587 #3588 #3589 #3590 #3591 #3592 #3593 #3594 #3595 #3596 #3597 #3598 #3599 #3600 #3601 #3602 #3603 #3604 #3605 #3606 #3607 #3608 #3609 #3610 #3611 #3612 #3613 #3614 #3615 #3616 #3617 #3618 #3619 #3620 #3621.

Date	Marker	Wave	What that day folded in
09-15	@ #3660	+1,378	39 PRs in a clean contiguous run #3621 -> #3660 with no out-of-order merge; 1,339 + 39 = 1,378 exactly, tip commit the highest merged number again after a one-band break. Featureless, the ninth here, and refactor-free for a second, so net-positive for a fourth: 27 fix + 11 test + 1 docs, 125 files, +8,791 / -303; fix 896 -> 923, test 143 -> 154, docs 108 -> 109, feat held at 144; correctness 77.6% -> 78.2%. Refactor table held at thirteen, removals ledger untouched for a fourth band, no ! or BREAKING CHANGE trailer -- correct rather than a gap. Eleven test pins, the most any band here has carried (#3621, #3626, #3632, #3634, #3641, #3644, #3646, #3647, #3656, #3659, #3660), each grading a door an existing suite reached only through a stand-in. Dominant fix shape new in kind and the third reading of one question -- a criterion already satisfied at reset is reported, not read as a result: #3629 the widest in consequence (measure_agreement's fractions are accuracies over a class-imbalanced column, so a constant answerer scored the holdout's own class balance), #3637 and #3642 the mirror pair (the same defect reading as an achievement on a stop_when clause and as a policy failure on a benchmark failure clause, both on step 1 with nothing moved), with #3633 and #3639. Four more lanes: a plan that commands nothing is a failure, not a plan (#3643, #3652, #3655, #3657, whose joint-key list was resolved once from the first waypoint); a ROS 2 name must be able to name a topic, and a refused call leaves no node behind (#3628, #3635, #3636, #3638, #3648, #3651); a reader grades the size its own payload declares (#3627, #3650, #3658, an .spz count sizing five body blocks graded nowhere while the magic beside it was graded on the next line); a refusal applies nothing from its own call (#3622, #3625, #3645). Rows also to the offline-snapshot (#3623) and one-owner (#3624) tables; the rest #3630, #3631, #3649, #3653, #3654. Milestone 0.5.0 unchanged at 11/0, eighteenth sync. Open issues held at 3 (#3099, #3122, #3578) and for the first time in the wave no issue was filed or closed. Open PRs 9 -> 11 (#3621 merged; #3640, #3661, #3662 opened), the extra-approval remedy in routine use with #3661 and #3662 both APPROVED, while #3205, #3343, #3427, #3496, #3515, #3577 and #3615 stayed green, armed, thread-resolved and REVIEW_REQUIRED; #3615 cleared its CHANGES_REQUESTED, #3640 took its place beside #3400, and #3343 was unchanged for a sixth sync. Fragments 1,530 (+28, none deleted -- the eleven test commits owe none), --check clean, CHANGELOG untouched 49 days (#1690), 0000- placeholders at four with sequencing item 6 unactioned for a third sync. The cleanest sweep the page had recorded: exactly one number in range absent from main (the open #3640), none closed unmerged, so #3169's shape took no sixteenth instance for a second band. One correction to the previous marker: it read 1,501 fragments where the tree held 1,502. Older caveats stood (cff130dc carries no PR reference; #2801's squash artifact)

2026-09-16: Reconciled against main @ #3750. Still no new tag -- v0.5.1 (2026-08-06) is 41 days old, now with 1,451 PRs (#1987 -> #3750) behind it. Folded in 73 PRs spanning #3661 -> #3750, with two out-of-order merges (#3661, #3662, both rows on the previous Open-PRs table): 1,378 + 73 = 1,451 reconciles exactly. Band shape 58 fix + 7 feat + 4 test + 3 docs + 1 refactor, so fix moved 923 -> 981, feat 144 -> 151, test 154 -> 158, docs 109 -> 112 and refactor 24 -> 25 with ci held at 17; correctness 78.2% -> 78.5%, a fourteenth sync, because the fix count grew faster than the feature count even in the band that broke the streak. The nine-band featureless run is over, and it ended on volume: 277 files, +18,945 / -682 -- net +18,263, the largest net-positive band this page has recorded, nearly double the previous high. The one refactor is examples-scoped (#3709), so the net-negative refactor table holds at thirteen and the removals ledger is untouched for a fifth band; no ! or BREAKING CHANGE trailer, correct rather than a gap. Drivers hold at twelve buses across 25 registry robot names and tools/g1 at 22, while dashboard/ moves eleven -> fifteen. The seven features are the news, and #3695 is the one that matters most: the dashboard has a server (19 files, +1,609 / -74), the core wired to the rails nine bands had accumulated and failing closed -- the slice the #2848 decomposition argument was for, with #3726 and #3739 already queued behind it. The other six say more about what the library just handed you: #3728 (tool_frame), #3727 (which way the arm extends), #3694 (SO-arm joints by label), #3724 (step feeds an open recording), #3731 (Robot(tool_name=)), #3697 (render_dir per Simulation). The dominant fix shape is new in kind: an unreadable reply names the peer, not the codec (#3670, #3672, #3676, #3677) -- four clients decoded a peer's response inside a try reporting a codec failure, so a malformed reply from the far end sent the caller to the wrong side of the wire. Three more lanes, each tabled above: a configuration entry that is not a string is refused (#3666 closing #3664, #3667, #3674, #3679 -- #3666 is the consequential one, a DDS Security credential the participant QoS then dropped, so a run believed it was authenticated); a motion that did not reach names what stopped it (#3703, #3705, #3727, #3733); a camera name answers for the camera it names (#3683, #3691, #3693, #3730). Rows also to the commanded-nothing lane at its entry point (#3680), a rollout made exclusive (#3741), and a surface offering a next step the thing it just made can take (#3686 closing #3684, #3687, #3719, #3742); the dashboard's own pass is #3710 and #3738; four test pins #3662, #3669, #3675, #3745. Milestone 0.5.0 did not move: 11 closed / 0 open (100%), nineteenth sync; v0.4.0 and v0.3.9 still OPEN and drained. Open issues 3 -> 5. Two were filed and closed in band against the commit that answered them (#3664 via #3666, #3684 via #3686), a fifth clean pass, and two remain open -- #3734 (the rotate_wrist half #3733 deliberately deferred) and #3744 (policies / P2, the offline-snapshot lane read as a range: the relative-action roster cannot hold across the pinned lerobot range, made visible by #3698 documenting it one commit earlier). Open PRs 11 -> 21, the largest queue recorded, up ten in a day -- and the constraint has moved from the ruleset to the volume: seven heads read APPROVED (#3729, #3743, #3746, #3749, #3751, #3752, #3753) where two did a day ago, while ten are green, armed, thread-resolved and still REVIEW_REQUIRED (#3205, #3343, #3427, #3496, #3515, #3577, #3615, #3714, #3726, #3737). Three are owed by their author (#3400, #3640, #3673 -- the last red, the only failing head on the board) and #3739 is owed a review decision. Sequencing item 3 (#3343) is unchanged for a seventh sync; #3205 is now the oldest head at eleven days. Sequencing item 6 is retired: the 0000- fragment placeholder defect is fixed (#3715 refuses the number, #3721 gives that rule one owner so every surface refuses it, #3722 renames the fragment that named no PR), so changelog.d/ holds zero placeholders where it held four for three syncs. Release hygiene is otherwise unchanged and still the only standing gap: no ## [0.5.0] or ## [0.5.1] heading (CHANGELOG.md untouched since #1690, now 50 days), [Unreleased] frozen at 168, changelog.d/ at 1,601 .md of which 1,600 are fragments, --check OK (1600 pending); the band added 70 and deleted none -- 73 PRs produced 70 because the four test commits owe none and one PR carried two. Caveats -- the counting streak broke again, in the direction that under-reports: the tip is #3740 while #3750 is the highest merged number, so a marker read off git log's newest subject is ten short. The clean sweep broke after two bands, and on the duplicate shape: seventeen numbers in #3661 -> #3750 are absent from main -- nine open PRs, two open issues, two issues closed in band, and four PRs closed unmerged, every one a second implementation (#3665 against #3666, #3671 against #3672, #3736 against #3731, #3732 descoped into #3733 plus #3734) -- instances sixteen to nineteen of #3169's shape after two bands at zero, which is why sequencing item 7 stands. One stray: #3702's body carries Closes #767, a pull request merged in the v0.4.1 wave rather than an issue, so it closed nothing. Older caveats stand (cff130dc carries no PR reference, so git log shows 1,452 commits for 1,451 PRs, measured exactly at this marker; #2801's squash artifact makes a naive (#NNNN) grep return 1,450). On size: 1.2 KB of headroom was nowhere near enough for the largest band on record, so three condensations were taken, each verified by cross-reference conservation before the edit -- the 2026-09-10 -> 2026-09-14 log rows rostered into one paragraph (the reduction the previous sync named as next, -18.3 KB), the 2026-09-15 entry folded into the log table, and its band bullets into three Finding/References rows. 0 of 1,833 cross-references lost. The next pass: the 09-15 and 09-16 log rows take the same treatment, after which the only reductions left are the v0.4.1-era theme tables, which are the historical record this page exists to hold.

[cagataycali]

📣 Roadmap update — May 22, 2026

Key changes from the April 21 version:

1. v0.3.9 is in security gate phase — all 25 feature PRs are merged. Only 3 security PRs remain (feat(mesh): security hardening — authentication, validation, rate limits, lockout, audit integrity #194, fix(deps): resolve 19 dependabot security vulnerabilities #153, improve: stricter input validation in gr00t_inference #90) before we tag.

2. Scope grew significantly — v0.3.9 now includes:

   • Robot Mesh (Zenoh + AWS IoT) via feat(mesh): Mesh, AWS IoT - session, presence, RPC, streams, wiring #101
   • Full LIBERO benchmark evaluation (15 PRs)
   • CI security baseline (CodeQL, dependency review, action pinning)

3. v0.4.0 rescoped — Newton (feat(newton): NVIDIA Warp/Newton simulation backend — GPU-native SimEngine #96) and Isaac Sim (feat(isaac): Isaac Sim simulation backend — USD + IsaacLab 3.0 #97) deferred to v0.4.1+ (milestone Jul 31). v0.4.0 is now docs + policies + polish, targeting June 1.

4. New mermaid diagrams — release flow, Gantt chart, and architecture diagram added for clarity.

---

What's blocking v0.3.9 release right now:

• feat(mesh): security hardening — authentication, validation, rate limits, lockout, audit integrity #194 — Mesh security hardening
• fix(deps): resolve 19 dependabot security vulnerabilities #153 — Dependabot security vulns
• improve(gr00t_inference): consolidated input validation + loopback security default #196 — improve(gr00t_inference): consolidated input validation + loopback security default

Once those three land, we tag.

[cagataycali]

Roadmap reconciliation -- 2026-06-14

The body was stale at the #395-#403 consolidation train. Reconciled against main @ #452. Key deltas:

Consolidation train fully landed. All 9 umbrellas #386-#394 are closed; their PRs merged (several under renumbered IDs: #387->#409, #386->#410, #394->#411, #388->#402, plus #397/#399/#400/#402). The ~63 deferred follow-ups are closed. Open-issue count dropped to 2 (#94 + #415).

v0.4.0 features are merged, not in-review. Since the last update: Device Connect #370, Cosmos 3 #317, CuroboPolicy (#306 closed -> #442 migrated to cuRobo main API, closes #421), strands-robots doctor CLI #419, API-ergonomics #431, 5 hero examples #432, LeRobot 0.5.2 recording #366, Thor/Jetson GPU torch #374, SimEngine.describe() #407, plus sim-render fidelity (#400/#428/#436) and the em-dash sweep #434 all landed.

Corrections to old body:

• feat(policies): CuroboPolicy under [curobo] extra (closes #301) #306 CuroboPolicy: now merged via the refactor(curobo): migrate to cuRobo main API (MotionPlanner / DeviceCfg / GoalToolPose) (closes #421) #442 cuRobo-main-API migration (old body said in-review).
• feat(policy): Add Qwen-VLA: unified VLA policy provider + 4-stage training pipeline #319 Qwen-VLA: closed unmerged, dropped from v0.4.0 (old body said in-review).
• security(codeql): suppress py/unsafe-cyclic-import on simulation triple (closes #215) #216 CodeQL suppression: closed-superseded (old body said in-review).
• feat: add Device Connect integration (on top of main) #370 Device Connect: merged (old body listed it as a v0.4.1 draft).

Only one v0.4.0 feature PR still open: #305 (MoveIt2). Plus in-flight coverage PR #453.

Release gap. v0.3.8 remains the only published tag; all v0.3.9 + v0.4.0 changes are stacked under Unreleased in the CHANGELOG. Next action is cutting the combined v0.4.0 tag once #305/#453 resolve and the CHANGELOG is collapsed under a dated heading.