ACT checkpoints — ISR vs uniform vs full data on ALOHA transfer-cube (human demos)
LeRobot ACT pretrained_model folders, one per <job>/<step>. Jobs: aloha_transfer_cube_{raw,uni3,isr,isr_lam0,isr_k6} (seed 1000, steps 020000/040000/060000) and *_s2000 (seed 2000, 020000/040000).
Datasets are in degrees — at eval convert actions deg→rad and observations rad→deg (see sitecustomize.py in the experiment repo), and rate-match playback for compressed arms (action-repeat 2 for isr/isr_lam0, 3 for uni3/isr_k6).
Results, data, code: Kavin60606/isr-aloha-transfer-cube-experiment.