Release v1.1.1

[nu-jliu]

Summary

Adds gravity + Coulomb friction compensation, sim-side gravity comp, a threaded mujoco control loop, and self-collision gating for the viser interface. Tuned per-arm grav_comp_kd and coulomb_friction for yam, yam_pro, yam_ultra, and big_yam after hardware validation.

Changes

• .gitignore: ignore *.mjb mujoco binary cache files.
• examples/control_with_mujoco/control_with_mujoco.py: split the script into a robot_worker + viewer_worker pair spawned via multiprocessing, with the main process hosting only the IPC primitives (SharedMemory-backed numpy state, command Queue, stop Event). Isolates the 250 Hz CAN control loop from viewer/IK/rendering GIL contention. ProxyRobot / ProxyMotorChain / ProxyJointState stubs forward the Robot surface MujocoControlInterface uses; the published loop_hz mirrors motor_chain.comm_freq (fallback to a 1 s RateRecorder) so --log reports the real driver rate.
• i2rt/robot_models/arm/big_yam/big_yam.xml: update base + link1–link5 rgba values to match the shared yam family palette (purple base, tan link1, blue link2, olive link3, gray-tan link4, pink link5) for visual consistency.
• i2rt/robot_models/gripper/crank_4310/crank_4310.xml: no net mass change in this PR (briefly bumped, then reverted to 0.60366 kg).
• i2rt/robots/config/yam.yml: add grav_comp_kd [0.1, 0.1, 0.1, 0.3, 0.05, 0.05] and coulomb_friction [0.3, 0.3, 0.3, 0.06, 0.06, 0.06]; retune gravity_comp_factor after friction-comp activation.
• i2rt/robots/config/yam_pro.yml: align gravity_comp_factor, grav_comp_kd, and coulomb_friction with yam/yam_ultra tuning.
• i2rt/robots/config/yam_ultra.yml: same tuning as yam_pro.
• i2rt/robots/config/big_yam.yml: grav_comp_kd [0.5, 0.5, 0.5, 0.3, 0.1, 0.1] and 0.1 Nm coulomb_friction on joints 0–1 (J4 weighting matches its larger wrist moment arm).
• i2rt/robots/config/{crank_4310,flexible_4310,linear_3507,linear_4310,no_gripper,yam_teaching_handle}.yml: drop multi-line last_joint_mount block comments.
• i2rt/robots/get_robot.py: load hw.grav_comp_kd and hw.coulomb_friction, append 0.0 for the gripper slot (gripper is position-commanded, not gravity-balanced), and pass both into MotorChainRobot. Pass unit gravity_comp_factor to SimRobot so sim uses exact model-based comp.
• i2rt/robots/motor_chain_robot.py: accept grav_comp_kd and coulomb_friction kwargs, store with length asserts, seed _commands.kd from grav_comp_kd only when zero_gravity_mode=True, apply coulomb_friction * sign(q_dot) alongside gravity comp in update(), expose both in get_robot_info(). Add enter_gravity_comp_idle() that resets _commands to zeros with kd=self._grav_comp_kd so CONTROL→VIS toggles re-enter idle damping cleanly without rewriting cached _kp/_kd.
• i2rt/robots/sim_robot.py: add a daemon physics thread that runs gravity compensation against the mujoco model, with enable_gravity_comp / disable_gravity_comp toggles, cmd ↔ mj_qpos mapping, wall-clock-relative sleep at 200 Hz, and clean shutdown in close().
• i2rt/robots/utils.py: add grav_comp_kd and coulomb_friction ndarray fields to _ArmHWConfig, parse with non-negativity asserts and logging in _load_arm_config.
• i2rt/utils/mujoco_control_interface.py: move sliders/mocap → IK → collision → command_joint_pos into a daemon control thread at control_dt cadence; guard shared MjData with viewer.lock() and only hold the lock for brief snapshot/write-back sections so the viewer never stalls commands. Start sim physics on run(), default to VIS in sim, toggle grav-comp on SPACE between VIS and CONTROL, and drop per-command joint command: log spam.
• i2rt/utils/viser_control_interface.py: add a scratch _check_data MjData buffer and _has_self_collision so IK and joint-slider modes refuse commands that would self-collide. Refactor _denormalize_slide_joints / _enforce_eq_constraints into _on(data) variants so the collision check writes to _check_data without corrupting render state. Add _enter_vis_grav_comp / _enter_control_grav_comp helpers (hasattr dispatch: SimRobot.enable/disable_gravity_comp, MotorChainRobot.enter_gravity_comp_idle) and detect VIS↔CONTROL transitions in the main loop instead of from the GUI callback (avoids races with in-flight command_joint_pos calls).

Test Plan

• Sync the branch and install the repo: git fetch origin && git checkout feat/grav-comp-friction-threaded-mujoco && uv sync.
• Sim grav-comp (no hardware needed):
  • Run uv run python examples/control_with_mujoco/control_with_mujoco.py --robot yam --sim and confirm the viewer opens in VIS mode by default.
  • Press SPACE to toggle CONTROL — sliders/mocap should drive the arm; SPACE again should release the arm and let it hold pose under sim gravity comp without drifting.
  • Verify --log prints a loop_hz near the expected control rate (≈250 Hz).
• Hardware grav-comp on yam:
  • With CAN connected, run uv run python examples/control_with_mujoco/control_with_mujoco.py --robot yam and confirm the arm enters gravity-comp idle on startup (you can hand-guide each joint with smooth settling, no chatter).
  • SPACE → CONTROL: drag a slider, confirm the arm tracks. SPACE → VIS: confirm the arm releases (no leftover PD clamp) and re-enters idle damping.
• Hardware grav-comp on yam_pro / yam_ultra / big_yam: repeat the previous step with --robot yam_pro, --robot yam_ultra, and --robot big_yam. Confirm joint settling is smooth with the new grav_comp_kd / coulomb_friction values.
• Self-collision gating: run the viser interface (uv run python examples/control_with_viser/control_with_viser.py --robot yam --sim). Drag the IK target into a self-colliding pose and confirm the arm refuses the command (joints clamp at the last safe pose).
• Edge case — friction comp polarity: in sim, command a slow motion in both directions; confirm the joint torque sign flips with q_dot sign and the arm doesn't lurch when reversing.
• Run any existing unit/integration tests: uv run pytest.

🤖 Generated with Claude Code