Version archive

A guide’s content is never edited after the fact. Every content change becomes a new version (minor for additions and corrections, major for rewrites), so links into an older version keep meaning what they meant.

What changed in V3.1

• New: Part III (§30–§43), a complete, beginner-level guide to running a trained policy with lerobot-rollout: concepts, checking the model files, every flag, what happens when, a first run in small steps, tuning, evaluation, DAgger, troubleshooting. It replaces the old §24.4–§24.11, which were short. §24.5 “Which policy” is now §24.4.
• Correction: lerobot-rollout refuses recording datasets whose name doesn’t start with rollout_. The V2.2 examples (eval_…, hil_…) would have failed at start-up. Fixed in §37 and §38.
• Correction: lerobot-record appends a date-time tag to the dataset name unless --dataset.no_stamp=true, so the V2.2 training command couldn’t find the dataset under the name it was recorded with. §24.1 now uses no_stamp.
• New pitfall: zeus logs in to a Wayland session, where the LeRobot keyboard controls (episodic, DAgger) probably don’t work (§37.3).

What changed in V3

• New: Part III (§30–§43), a complete, beginner-level guide to running a trained policy with lerobot-rollout: concepts, checking the model files, every flag, what happens when, a first run in small steps, tuning, evaluation, DAgger, troubleshooting. It replaces the old §24.4–§24.11, which were short. §24.5 “Which policy” is now §24.4.
• Correction: lerobot-rollout refuses recording datasets whose name doesn’t start with rollout_. The V2.2 examples (eval_…, hil_…) would have failed at start-up. Fixed in §37 and §38.
• Correction: lerobot-record appends a date-time tag to the dataset name unless --dataset.no_stamp=true, so the V2.2 training command couldn’t find the dataset under the name it was recorded with. §24.1 now uses no_stamp.
• New pitfall: zeus logs in to a Wayland session, where the LeRobot keyboard controls (episodic, DAgger) probably don’t work (§37.3).

What changed in V2.2

• Correction: V2.1 said that with interpolation_method: none the trajectory controller “ramps linearly” to the first point over time_from_start. That was wrong. It holds the current position for the whole time_from_start and then jumps to the target (confirmed in sim, see log 2026-10-09). On real motors that’s a full-speed step. Fixed in §4, §8.1, §12.3 and §16.
• New: how to switch the bringup controllers to interpolation_method: splines, and how to check it (§4). This is now a pre-flight item for the real robot (§12.1).
• Correction: the §12.1 checklist said can0 = right arm. On zeus it’s the reverse (left = can0, §22).