TRI: pretrained large behavior models accelerate robot learning

The Toyota Research Institute (TRI) has advanced the development of Large Behavior Models (LBMs) to accelerate robot learning, demonstrating that a single pretrained LBM can learn hundreds of tasks and acquire new skills using 80% less training data. LBMs are trained on large, diverse datasets of robot manipulation, enabling general-purpose robots to perform complex, long-horizon behaviors such as installing a bike rotor. TRI’s study involved training diffusion-based LBMs on nearly 1,700 hours of robot data and conducting thousands of real-world and simulation rollouts, revealing that LBMs consistently outperform policies trained from scratch, require 3-5 times less data for new tasks, and improve steadily as more pretraining data is added. TRI’s LBMs use a diffusion transformer architecture with multimodal vision-language encoders and a transformer denoising head, processing inputs from wrist and scene cameras, proprioception, and language prompts to predict short action sequences. The training data combines real-world teleoperation data,