demodoctor-act-cleaned-s2

ACT policy on PushT from the DemoDoctor grid: condition cleaned, seed 2, 60,000 steps, batch size 32. Trained on the 175 episodes left after dropping stall- and jitter-flagged episodes from the corrupted set.

This policy was retrained locally on 2026-10-05 (LeRobot 0.6.1) after the original pod checkpoint was lost. Evaluation (50 rollouts, lerobot-eval, batch size 1): success 0/50, mean max reward 0.331.

Study, code and results (a null result for data corruption; the grid is underpowered): https://github.com/raihan-js/demodoctor