Large Behavior Model

Toyota Research Institute's term for a general-purpose, multi-task robot policy pretrained on large amounts of demonstration data.

Large Behavior Model is Toyota Research Institute's (TRI) name, by analogy with large language models, for a general-purpose manipulation policy pretrained on a large amount of multi-task robot data and used to output actions directly. TRI first raised the idea of building an LBM when it published its diffusion-policy skill-learning results in September 2023; a July 2025 paper then gave a concrete implementation: a diffusion Transformer that takes wrist and scene camera images, proprioceptive state, and a language instruction, and predicts 16 steps of action at once. It was trained on about 1,700 hours of data, roughly 468 hours of which came from TRI's own bimanual teleoperation. The paper found that after pretraining, learning a new task needs 3-5x less data, and performance scales smoothly with pretraining data size. TRI has partnered with Boston Dynamics since October 2024, and in August 2025 the two demonstrated a single LBM controlling a full electric Atlas, continuously walking while carrying parts and sorting them into bins.

ExampleIn the August 2025 demo, a single LBM controls Atlas's arms and legs together: it walks, crouches, picks up a part, and sorts it into a bin; when someone slides a closed box across the floor into its path, it adjusts and keeps working.

Also called

LBM, TRI LBM

Related

Toyota Research Institute · Diffusion Policy · Foundation Model · Behavior Foundation Model · Vision-Language-Action Model · Boston Dynamics Atlas (Electric)

Sources

TRI LBM 项目页：A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation (Chinese)
TRI News (2023-09-19): Toyota Research Institute Unveils Breakthrough in Teaching Robots New Behaviors
TRI News (2025-08-20): AI-Powered Robot by Boston Dynamics and TRI Takes a Key Step Towards General-Purpose Humanoids

As of

2025-08