Boston Dynamics and TRI: Large Behavior Model for Atlas

2025-08-20 · Publisher : Boston Dynamics et Toyota Research Institute · Domain : Robotics

Boston Dynamics and the Toyota Research Institute present a 450-million-parameter, language-conditioned end-to-end policy controlling the full Atlas humanoid at 30 Hz for long manipulation tasks.

### Background On August 20, 2025, Boston Dynamics shared the first results of its collaboration with the Toyota Research Institute (TRI): Large Behavior Models (LBMs) enabling the electric Atlas humanoid to perform long-horizon manipulation from a natural-language prompt. ### Specifications - **Policy**: a 450-million-parameter diffusion transformer model trained with a flow-matching objective. - **Inputs and outputs**: images, proprioception and a text prompt in; actions for the full Atlas body out, at 30 Hz. - **Whole body**: the policy uses the full humanoid form, including stepping, precise foot placement and crouching. - **Data**: teleoperation on Atlas, using the same control interface as the policy, which allows reuse of architectures developed for manipulation-only settings. - **Lineage**: builds on TRI's LBMs, which scale Diffusion Policy-like architectures. ### Why it matters Boston Dynamics, long known for hand-engineered controllers, adopts end-to-end learning for its humanoid. The TRI alliance combines mechanical know-how with learning expertise. ### Key takeaway Atlas moves from scripted acrobatics to learned behaviors, a necessary step before the product version shown at CES 2026.

Tags : Large Behavior Models, Humanoïdes, Atlas, Boston Dynamics, TRI