Action Chunking Transformer policy
Cup nesting on the SO-101
One arm, six joints, two cameras. This page records what went into the policy, what came out of the training run, and what has not been measured yet.
Screening
Every recorded episode was reviewed before training. Episodes that failed a check were removed rather than repaired.
Dataset commit—
Policy commit—
Report generated—
Training run
A single uninterrupted run on one L4. The eight defective episodes were the only source of instability; removing them was enough to make the run clean.
—
—
—
—
Configuration
Read from train_config.json in the model repository, so these are the values the run actually used rather than framework defaults.
Action space
Six joint targets issued at 30 Hz. Ranges below are the 5th, 50th and 95th percentiles across the verified episodes.
Evaluation
Rollout results on physical hardware are not published for this checkpoint. The stages below are the protocol, not results.
Access and licensing
The policy and its dataset are released under CC BY-NC 4.0 for research and evaluation. Commercial use, redistribution, or deployment in commercial products requires a separate licence, and the dataset terms carry through to the weights.
The public dataset repository is an evaluation sample. This checkpoint was trained on the full verified tranche. Pilot tranches, production tranches, and collection on other embodiments are available on request.
info@decisionfacts.io