Octo

Octo is an open-source generalist policy for robotic manipulation from UC Berkeley, Stanford, CMU, and Google DeepMind, published in May 2024. This transformer-based policy with a diffusion action head was trained on 800,000 trajectories from the Open X-Embodiment dataset, the largest robot manipulation dataset at the time.

Octo can be instructed via language commands or goal images and fine-tuned to new observation and action spaces within hours on consumer GPUs. At just 93M parameters (Octo-Base), it achieved performance comparable to the 55B closed RT-2-X, making it one of the first practical open foundations for vision-language-action robotics.