act_so100_reach

Reference LeRobot ACT trained on jaheroth/so100_reach (scripted-expert reach task on a simulated SO-100, MuJoCo). chunk_size=50, n_action_steps=16, n_decoder_layers=7, lr 2e-5, batch 64. Stopped at 43k steps (loss plateau ~0.014, ~325 epochs); checkpoints at 20k and 40k. Part of https://github.com/JaHeRoth/robot-learning