metadata
license: apache-2.0
tags:
- lerobot
- so101
- mujoco
- ppo
- distillation
- sim2real
SO101 MuJoCo sim2real: PPO teacher -> vision student
Checkpoints from lerobot.rl.so101_mujoco. Code, configs and full instructions:
github.com/suuu-u/Learn_SO101.
Download them to outputs/so101_mujoco/ of the lerobot checkout, where the configs expect them:
hf download suuu3/so101_mujoco_sim2real --local-dir outputs/so101_mujoco
| File | Kind | Description |
|---|---|---|
teacher_realscene.pt |
PPO teacher (privileged state: cube pose, contacts, ...) | Trained on the real-scene-matched MuJoCo scene (ppo_realscene/config.yaml), best checkpoint = iteration 7600, rolling success rate ~0.54 |
student_realcam.pt |
Vision student (front + wrist cameras + 18-dim joint state), used for deployment | DAgger distillation from teacher_realscene.pt (distill_realcam/config.yaml), best checkpoint = iteration 200, rolling success rate ~0.48 |
Success rates are rolling training statistics under domain randomization, not a separate evaluation. The released weights were trained in stages; training from scratch with these configs may give slightly different results.
ppo_realscene/ and distill_realcam/ contain the training config.yaml and metrics.jsonl.
# Evaluate in simulation / deploy (from the lerobot checkout)
python -m lerobot.rl.so101_mujoco.eval --checkpoint outputs/so101_mujoco/student_realcam.pt
python -m lerobot.rl.so101_mujoco.deploy --headless true --n_episodes 50