my_act_pusht

ACT reimplemented from scratch in PyTorch (see scripts/act.py + scripts/my_train.py in the repo): 157.1 avg_sum_imputed_reward, 61.1% success at 100k steps (n=5000) - inside the reference implementation seed distribution. Raw state_dicts every 20k steps + loss curves.

Part of a 6-week robot-learning training block (https://github.com/JaHeRoth/robot-learning). Evaluated on gym-pusht at n_action_steps=16; avg_sum_imputed_reward imputes 0.95/step to horizon 300 after success.