Octo Model Team Octo

2024-05-20 · Publisher : Octo Model Team (Berkeley, Stanford, CMU, Google DeepMind) · Domain : Robotics

An academic team led by Berkeley releases Octo, an open generalist robot policy trained on 800,000 Open X-Embodiment trajectories and fine-tunable in a few hours on a consumer GPU.

### Background On May 20, 2024, the Octo Model Team, bringing together researchers from UC Berkeley, Stanford, Carnegie Mellon and Google DeepMind, released Octo, an open-source generalist robot policy. The goal is a shared starting point that any lab can adapt to its own robot. ### Specifications - **Architecture**: a large transformer taking images plus either a language command or a goal image. - **Data**: 800,000 trajectories from the Open X-Embodiment dataset, the largest robot manipulation corpus available at the time. - **Adaptation**: fine-tuning to new sensors and action spaces within a few hours on consumer GPUs. - **Evaluation**: experiments across 9 robot platforms, according to the paper. - **License**: code and weights released under the MIT license. ### Why it matters Octo shows that a policy pretrained on heterogeneous data can initialize very different robots. It paved the way for the open models that followed, such as OpenVLA a few weeks later. ### Key takeaway Octo is one of the first open building blocks of generalist robotics: a free, documented, adaptable starting model rather than one specialist model per task.

Tags : Politique généraliste, Open Source, Open X-Embodiment, Manipulation