Octo is an open-source generalist robot policy: a transformer-based model trained on a diverse mix of 800,000 robot trajectories. The project has 1,808 stars on GitHub and was last updated in 2024. Developed by researchers at UC Berkeley and partner institutions, Octo is one of the reference open models for robot learning with foundation models.
Instead of training a separate controller for every robot and task, Octo learns a single diffusion policy from the Open X-Embodiment dataset, the largest robot data collection assembled at the time. Out of the box it accepts multiple RGB camera inputs, can drive various robot arms, and follows either language commands or goal images. Its modular attention structure means the same pretrained model can be fine-tuned to new sensory inputs, action spaces, and robot morphologies using only a small target-domain dataset and modest compute. The codebase includes training and fine-tuning pipelines, and the pretrained checkpoints can be loaded in a few lines of Python. The project page collects demonstration videos, the paper, and links to the training datasets, making it a practical starting point rather than just a model release.
Octo is free and open source under the MIT license, including the code and the pretrained model weights. You only need your own GPU and robot hardware to run it.
Octo is for robotics researchers and engineers who want a strong open baseline for generalist robot control, and for labs fine-tuning foundation policies onto their own robots instead of training from scratch.