Part III · §30–§43 · Running a trained policy with LeRobot, step by step
Who this part is for
You have never run a learned policy on a robot before. You have (or will soon have) a policy trained with
lerobot-trainon demonstrations recorded with the Minis (§24), and you want to let it drive the two OpenArm followers on zeus.Part III starts from zero: what a policy is, what files you need, every flag of the command, what happens second by second, a first run broken into small safe steps, how to tune it, how to measure it, and what to do when it misbehaves. Read §30 to §34 once before the first run. After that, §34 and the cheat-sheet in §43 are enough day to day.
Status (2026-10-09)
None of Part III has been run on the zeus robot yet. No policy has been trained so far. Everything below was checked against the LeRobot source installed on zeus (
~/Documents/lerobot, commitb863b42, v0.6.2): flag names, defaults, error messages and log lines. Behaviour that can only be confirmed on the hardware is marked ⚠ verify. When something in this part turns out to be different on the real robot, write it in the day’s log and correct it in the next guide version.
Everything in Part III runs on the host, in the lerobot conda env, with no ROS launch running on the follower buses (§25). The safety rules of §26 all still apply.