Octo Model Team

Octo Model Team is the research collaboration behind Octo, an open-source transformer-based diffusion policy pretrained on 800k robot episodes from the Open X-Embodiment dataset. Octo ships in Small and Base sizes, supports language and goal-image conditioning, and finetunes efficiently to new observations, action spaces, and embodiments.

Evaluated across nine real robot setups at four institutions, Octo matches much larger closed policies out of the box and beats training-from-scratch baselines by a wide margin after finetuning. The team releases code and a shared finetuning recipe openly, making Octo a flexible default generalist policy for downstream robotics research and applications.