Octo-Base
Open generalist policy trained on 800k Open X-Embodiment trajectories with a diffusion action head; supports new sensors and action spaces by adding lightweight input/output heads.
- 68%
- 71K
- 10 Hz
- 800K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
WidowX manipulation | Octo eval | 68% |
New-robot fine-tune | 4 institutions, 9 robots | 24 |
Goal-image conditioning | Bridge V2 | 63% |
Training
Diffusion-head transformer policy
93M
128x TPU v4
8.4K accelerator-hours
15K h
800K
300k steps of masked-goal conditioning so one checkpoint accepts either language or goal images; DDPM head over 4-step action chunks.
Adding a new robot needs only a new tokenizer/head; full backbone fine-tuning converges in <2 h on 1 node.
Linked datasets, robots & papers
Resolving related models, datasets and papers…