Autobrik/ecse552-act-baseline-a

Image-only ACT policies (no joint-position inputs) from the McGill ECSE 552 project Learning to Manipulate a New Object from Action-Free Videos (code, config configs/act_baseline_a.toml), trained on Autobrik/ecse552-ocbench-block at revision aa4e3eecc9b7266f561923be09f848c63cbe3088.

Each n<demos>/seed<k>/ holds one training run:

File	Content
best.pt	lowest L1 on the source validation episodes (80/20 split by episode)
last.pt	weights after the full number of steps
selected.pt	the better of the two on validation rollouts (selection.json)
config.json	resolved config, action statistics, data splits, training initial states
metrics.jsonl, summary.json	training and validation losses
provenance.json	code commits, dataset revision, GPU

import torch
from action_free.policy import ACTPolicy
policy = ACTPolicy.from_checkpoint(torch.load("n100/seed0/selected.pt", weights_only=False))
chunk = policy.eval().predict(images)  # uint8 (B, 3 cameras, 224, 224, 3) -> (B, 10, 7)