Model Family
Octo
Open-source transformer policy for robot control trained on Open X-Embodiment.
Definition
Octo is an open generalist policy from UC Berkeley trained on 800k robot demonstrations across 9 embodiments. Smaller than OpenVLA, it targets fast fine-tuning and inference on a single GPU.
Common use cases
- Imitation learning
- Multi-embodiment policy
- Sim-to-real