sim2real

A lightweight and modular sim2sim and sim2real deployment stack.

Chinese version: README_zh.md

Full documentation: https://egalahad.github.io/sim2real/

If you're looking for the HDMI deployment stack, go to hdmi tag.

Runtime Artifacts

Large runtime artifacts are not stored in git. Download the shared sim2real artifacts folder and place checkpoints/ and third_party/ at the repo root.

See Download Artifacts for the expected directory layout and onboard dependency notes.

Quick Start

uv sync --extra inference-cpu

For G1 onboard installation or repair, invoke the repository Codex skill $configure-g1-sim2real from .agents/skills/configure-g1-sim2real.

Run offline motion tracking (sim2sim):

uv run sim2real/sim_env/base_sim.py --robot g1
uv run sim2real/rl_policy/tracking.py --robot g1 \
  --policy_config checkpoints/mimic-lite/v1_1/policy.yaml \
  --motion_path hf://elijahgalahad/any4hdmi-g1-lafan/motions/walk1_subject1.npz

After both processes are up, press ] in the policy terminal to start. Open the mjviser URL printed by base_sim.py, then use the Elastic Band controls in the viewer UI to disable or tune the virtual gantry.

Migrating to sim2real

This repo includes a Codex skill for adapting policies trained in external codebases into sim2real:

.agents/skills/adapt-policy-to-sim2real

Converted checkpoints are distributed through the shared sim2real artifacts folder.

Currently supported adapted / distributed checkpoint families:

Policy family	Config path(s)	Notes
Mimic Lite v1.1	checkpoints/mimic-lite/v1_1/policy.yaml	8x8192 PPO-ROA student with student-only reference-motion noise.
MimicLite-ROA	checkpoints/mimic-lite/roa/policy.yaml	Latest 16x16384 PPO-ROA student release.
MimicLite-PPO	checkpoints/mimic-lite/ppo/policy.yaml	Latest 16x16384 Huge PPO release.
HEFT	checkpoints/heft	PMG and compliance variants.
HoloMotion v1.4.0	checkpoints/holomotion/v1_4_0/policy.yaml	Uses the official unmodified ONNX from HorizonRobotics/HoloMotion_models; place it at checkpoints/holomotion/v1_4_0/policy.onnx.
SONIC release	checkpoints/sonic/release	Release G1 and SMPL encoder variants.
SONIC low-latency	checkpoints/sonic/low_latency	Low-latency G1 and SMPL variants.
SONIC v1.1	checkpoints/sonic/v1_1/g1/policy.yaml	G1 policy with heading-normalized reference orientation.
GRIT v0.0.1	checkpoints/grit/v0_0_1/policy.yaml	Nine-frame reference context and ten-frame proprioceptive history.
ScaleBFM	checkpoints/scalebfm	ScaleBFM Humanoid Transformer XL and M ONNX exports from WeishuaiZeng/ScaleBFM.
BFM-Zero	checkpoints/bfm-zero/exp_lafan40-100style_update_z10/policy.yaml	Latent-conditioned motion tracker.
TeleopIT	checkpoints/teleopit/policy.yaml	TeleopIT policy wrapper.
Humanoid-GPT	checkpoints/humanoid-gpt/policy.yaml	Humanoid-GPT policy wrapper.
TWIST2	checkpoints/twist2/policy.yaml	TWIST2 policy wrapper.

Mimic Lite v1.1

Mimic Lite v1.1 keeps the teacher and reward target clean while adding reference-motion noise only to the student command. Deployed inputs remain clean. The full Train / Adapt / Finetune chain uses 8x8192 environments for 4000 / 1000 / 2000 updates and consumed 50.106 measured GPU hours.

MotionDecode metric	Locomotion-80	Manipulation-48	Ground-60	Dance-40
Progress	99.320%	93.619%	65.542%	55.837%
Tracking Return	1.9126	1.7905	0.9198	0.9811
Body position	25.11 mm	28.25 mm	144.66 mm	43.13 mm
Body orientation	0.08142 rad	0.09452 rad	0.55465 rad	0.16783 rad
Wrist position	22.15 mm	23.57 mm	n/a	n/a
Wrist orientation	0.07745 rad	0.09997 rad	n/a	n/a

Download checkpoints/mimic-lite/v1_1 from the shared sim2real artifacts folder. Full results are on the Motion Tracking Leaderboard.

For a fair comparison, we report the motion-lookahead latency required by each policy, defined by its furthest required future-reference frame. All values use the shared 50 Hz reference-motion contract.

Policy	Mimic Lite v1.1	MimicLite-ROA	MimicLite-PPO	HEFT	HoloMotion	SONIC	SONIC low-latency	SONIC v1.1	GRIT v0.0.1	ScaleBFM XL	ScaleBFM M	BFM-Zero	TeleopIT	Humanoid-GPT	TWIST2
Motion-lookahead latency	0.08 s	0.08 s	0.08 s	0.12 s	0.20 s	0.90 s	0.18 s	0.90 s	0.26 s	0.10 s	0.10 s	0.12 s	0.00 s	0.02 s	0.00 s

Real-robot Environments

Robot SDKs are kept out of the generic root environment. G1 inline deployment uses uv sync --extra inference-cpu --extra robot-g1. See Robot I/O Modes for setup and deployment commands.

Next Steps

• Docs Home
• Getting Started
• Offline Motion Tracking Tutorial
• Pico Teleoperation Tutorial

Citation

If you find sim2real useful in your research, please cite:

@misc{sim2real2026,
  author       = {{RoboParty Lab Team}},
  title        = {sim2real: A Lightweight and Modular Sim2sim and Sim2real Deployment Stack},
  year         = {2026},
  howpublished = {\url{https://github.com/EGalahad/sim2real}},
  note         = {Documentation: \url{https://egalahad.github.io/sim2real/}}
}