Phase-conditioned ACT block stacking, 15k steps (SO-101, wrist camera)
ACT (chunk_size 100) with an extra input, observation.environment_state: a one-hot of the current sub-task
(pick/place × orange/green/purple). See hrl-imerse/doug-stack-phases for how the labels are made.
- Trained on the 84 clean episodes (75 train / 9 val), 15k steps, colour jitter plus RandomAffine augmentation.
- Validation loss: 5k 0.570, 10k 0.551, 15k 0.560.
Running it: at run time, the phase is tracked live from the gripper reading and fed to the model, with temporal ensembling (coeff 0.01). If the gripper closes on nothing while placing, the run drops back to picking the same block. The runner run_phased.py is in the GitHub repo msong33/hrl_imerse, branch doug/stack-phases, folder assignment2/task2_phases.
Real-arm test (2026-11-17): stacked orange perfectly after a few grasp attempts, then hesitated and did not head for green. The phase input clearly steers the arm: given the same image, "pick" inputs point it toward the blocks and "place" inputs toward the tower. Grasp precision is still the bottleneck.
- Downloads last month
- 23