Prandium Docs

39. Running a single-arm policy

1 min read

Part III: Running a trained policy · §39 of 43

A policy trained on one arm (observation.state [8], recorded with --robot.type=openarm_follower) runs with the single-arm robot type. On zeus the right arm is can1 and the left arm can0. The single-arm calibration file is my_follower:

lerobot-rollout \
  --strategy.type=base \
  --policy.path=<path>/pretrained_model \
  --robot.type=openarm_follower \
  --robot.id=my_follower \
  --robot.port=can1 \
  --robot.side=right \
  --robot.max_relative_target=10.0 \
  --robot.cameras='<same as recording>' \
  --task="<task>" --fps=30 --duration=30 --device=cuda --display_data=false

The other arm isn’t connected: it stays unpowered and hangs. A bimanual policy can’t run on one arm, and the reverse doesn’t work either: the number of joints must match.

Graph View

Backlinks