Octo

Octo is an open-source, widely applicable generalist robot policy developed by the Octo Model Team, a group of researchers from UC Berkeley, Stanford University, Carnegie Mellon University and Google DeepMind. The model is built on a transformer backbone with a diffusion action head (diffusion policy), letting it model multi-modal manipulator action distributions.

Octo was pretrained on 800k trajectories drawn from a mixture of 25 datasets in the Open X-Embodiment collection — the largest robot manipulation dataset to date. It can be instructed via natural-language commands or goal images, and its modular attention structure enables fast fine-tuning to new robots, sensors and action spaces within a few hours on standard consumer GPUs.

Two variants are available: Octo-Small (27M parameters) and Octo-Base (93M parameters). The code is released under the MIT license (JAX framework) and the weights are published on Hugging Face (rail-berkeley). Across experiments on 9 robotic platforms, Octo serves as a versatile policy initialization that can be effectively finetuned to new observation and action spaces.