Octo

Open-Source Generalist Robot Policy — transformer-based diffusion policy for manipulation.

Overview

Octo is a transformer-based diffusion policy trained on 800K robot trajectories from Open X-Embodiment. It supports language commands or goal images, multiple RGB cameras, and various robot arms. Octo-Small (27M) and Octo-Base (93M) can be finetuned in hours on consumer GPUs.

Architecture & Scale

Octo-Small 27M params, Octo-Base 93M params
800K trajectories from Open X-Embodiment
Language or goal-image conditioning
Multi-camera, multi-robot out-of-the-box

Official Links

octo-models.github.io — Project site
github.com/octo-models/octo — Code (MIT)
Hugging Face: rail-berkeley/octo-models — Weights

Citation

See the project site for BibTeX and paper references.