Autobrik/ecse552-act-stochastic

Original ACT (tonyzhaozh/act, unmodified) trained on Autobrik/ecse552-ocbench-block at revision 6b5e02b6ec6b7a1083479ca57411b3c5295a6021, for the McGill ECSE 552 project Learning to Manipulate a New Object from Action-Free Videos (code, config configs/act_stochastic.toml, adapter src/action_free/act_original.py). Differences from ACT: images only (robot state is zeroed), our 7 action dimensions zero-padded to upstream's 14, variable-length episodes, chunk 20 at 10 Hz (2 s, as ACT's 100 at 50 Hz). Actions are absolute joint targets. Runs on several collections (oracle (d), pseudo-labelled targets for (b)/(c)) list them in act_run.json (data.collections).

Each n<demos>/seed<k>/ holds one run: policy_best.ckpt (lowest validation loss), policy_last.ckpt, selected.ckpt and selection.json (the better of the two on source validation rollouts, the one evaluated on test), dataset_stats.pkl, act_run.json (arguments, split, data), history.json (losses per epoch) and provenance.json. Load with action_free.act_original.ActPolicy("n300/seed0/selected.ckpt", device).