ACT (Action Chunking Transformer)
The policy behind ALOHA: a CVAE transformer predicting chunks of ~100 future actions, learning millimetre-precision bimanual skills from only 50 demonstrations per task.
- 90%
- 79K
- 50 Hz
- 1.2K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
Battery slot insertion | ALOHA real | 90% |
Cup opening / candy wrapper | ALOHA real | 84% |
Sim transfer cube | ALOHA sim | 86% |
Demos required | ALOHA | 50 |
Training
CVAE transformer with action chunking
80M
1x RTX 4090 / A100
40 accelerator-hours
12 h
1.2K
Train from scratch per task, 2000 epochs, temporal ensembling of overlapping chunks at inference.
Chunk length and temporal ensembling are the two decisive hyperparameters; more cameras beats more data.
Linked datasets, robots & papers
Resolving related models, datasets and papers…