🐙 Octo: An Open-Source Generalist Robot Policy
AI summary of the linked articleOcto is an open-source, transformer-based diffusion policy for robotic manipulation, pretrained on 800k robot episodes from the Open X-Embodiment dataset and designed to be quickly finetuned to new robot setups. The project introduces two initial versions, Octo-Small (27M parameters) and Octo-Base (93M parameters). Out of the box, Octo outperforms RT-1-X when tasks are specified in natural language and performs similarly to the 55-billion-parameter RT-2-X. On the WidowX tasks, goal image conditioning achieved 25% higher performance on average. Across six finetuning evaluation setups, Octo outperforms the next best baseline by 52%, using about 100 target demonstrations per task. |
Recommended by 2 curators |
Comments No comments yetWhat caught your eye? Start the discussion. |
