Part III: Running a trained policy · §43 of 43
# ---------- once per session ----------
conda activate lerobot
bash ~/Documents/openarm_lerobot/setup_openarm_lerobot.sh # 16/16 motors
candump -n 20 -T 500 can0; candump -n 20 -T 500 can1 # nothing = bus free
# ---------- every run ----------
python ~/Documents/openarm_lerobot/compare_mini_follower.py --once # arms at the training start pose?
bash ~/Documents/openarm_lerobot/run_policy.sh --interactive=true --duration=10 # /start, /reset, /stop
bash ~/Documents/openarm_lerobot/run_policy.sh # plain 30 s run, Ctrl+C ONCE to stop early
# ---------- tuning (one at a time) ----------
bash run_policy.sh --policy.n_action_steps=30 # re-plan every 1 s
bash run_policy.sh --policy.n_action_steps=1 --policy.temporal_ensemble_coeff=0.01 # smoothest (ACT)
bash run_policy.sh --interpolation_multiplier=2 # 60 Hz commands
# ---------- evaluation episodes (recorded) ----------
# --strategy.type=episodic --dataset.repo_id=<hf_user>/rollout_<name> --dataset.single_task="…"
# --dataset.num_episodes=10 --dataset.episode_time_s=40 --dataset.reset_time_s=20 --dataset.push_to_hub=false| Must match training | Value on zeus |
|---|---|
--robot.type | bi_openarm_follower (single arm: openarm_follower) |
--robot.id | my_bimanual_follower (single arm: my_follower) |
| Buses / sides | can0 = left, can1 = right |
--robot.cameras | copied from the recording command |
--fps | the dataset’s fps (30) |
--task | the recording’s --dataset.single_task |
Official docs: https://docs.openarm.dev · Discord: https://discord.gg/FsZaZ4z3We