Part III: Running a trained policy · §42 of 43
Everything in §13 and §26 applies. Specific to running policies:
- A policy is not a program you can read. It can do things it never did in training, especially in situations it hasn’t seen. First runs: short
--duration, interactive mode, a person at the power switch. - No soft start. The first action goes out at full stiffness. Start from the training start pose.
- No collision checking, in the policy or in the return move. Keep the workspace clear of anything that isn’t part of the task, and keep people out of reach.
- Silent zeros: a motor that doesn’t answer reads 0°. If it’s in the start pose capture, the return move drives that joint to 0. Always confirm 16/16 motors before running.
- Torque off at the end: the arms fall from the start pose. Support them, or start from hanging so there’s nowhere to fall.
- Ctrl+C once. Twice can cut the return move short. Never Ctrl+Z.
- The right
--robot.id, or LeRobot re-zeroes the follower motors (and the ROS zero with them). - Never unattended.
sentryand--duration=0exist, but this robot has no external safety system.