← work

ACT gradient project

Finding the exact point where imitation learning fails

GitHub profile

This started with a question: where exactly does an imitation-learned policy start to fail? I trained ACT on 50 demonstrations across 3 MimicGen tasks with 100-action chunks and instrumented the gradients during training. It turns out that opposing gradients across the action horizon mark precisely where compounding error takes over. Restarting learning from stored boundary observations raised task success from 60% to 69% over baseline ACT, with 12% faster convergence - across 3 seeds, in MuJoCo and on a physical SO-101 arm.