ACT
ACT (Action Chunking Transformer) is MetaFine's lightest VLA reference. It predicts chunks of future actions in one shot — a simple, cheap baseline that's easy to overfit on small demonstration sets and a useful sanity check before reaching for π0 or OpenVLA.
What it is
ACT (Action Chunking Transformer) — a transformer policy that predicts short action chunks from image + proprioceptive observations. Vendored under core/policies/act.
Data
ACT consumes a LeRobot dataset. Run the full record → merge → replay → convert flow, converting with utils/convert_to_lerobot.py.
Install & run
Each backbone pins its own (often conflicting) deps — install it in isolation:
$ pip install -e core/policies/act
ACT's exact training and evaluation flags are documented in its own in-repo guide:
$ cat core/policies/act/README.md
Status: vendored. The end-to-end ACT path inside MetaFine has not been validated; treat core/policies/act/README.md as the source of truth.