Octo Model Team offers Octo
Product
Octo is an open-source transformer-based diffusion policy pretrained on 800k robot episodes for generalist manipulation.
- Currently Offered
Octo is an open-source transformer-based diffusion policy pretrained on 800k robot episodes for generalist manipulation.
Octo is an open-source transformer-based diffusion policy for generalist robotic manipulation, pretrained on 800k episodes from 25 Open X-Embodiment datasets. Offered as Octo-Small with 27M parameters and Octo-Base with 93M parameters, it accepts language instructions or goal images plus observation histories and decodes multimodal actions through diffusion.
Out of the box Octo controls multiple robots from its pretraining mix and outperforms RT-1-X while matching the 55B-parameter RT-2-X. With about one hundred target demonstrations and one shared finetuning recipe, it adapts to new observations, action spaces, and embodiments, outperforming from-scratch training by 52 percent on average across six setups and anchoring open generalist-policy research.
OpenVLA is a 7B-parameter open-source vision-language-action model for generalist robot manipulation.
Gemini Robotics 2 is Google DeepMind's most advanced vision-language-action model for controlling any robot embodiment.
OpenVLA is a 7B-parameter open-source vision-language-action model for generalist robot manipulation.
Gemini Robotics 2 is Google DeepMind's most advanced vision-language-action model for controlling any robot embodiment.
22-degree-of-freedom bimanual mobile robot with a 440 lb base, 180 lb lift, 22 in-house actuators and a video-pretrained model.
EgoVerse is an ecosystem for curating human egocentric data and learning from it for robot learning.
Causal world model platform for Physical AI and robotics.
