Boston Dynamics and Toyota Research Institute (TRI) have demonstrated Atlas, the humanoid robot, running on a Large Behaviour Model (LBM) that controls its entire body at once.
Shown in a new joint video, Atlas completes a long, continuous sequence of everyday tasks — walking, crouching, lifting, packing and sorting — while adapting on the fly when researchers add surprises, such as moving boxes mid-job.
Unlike earlier approaches that split leg control (for balance and walking) from arm control (for handling objects), the LBM directs the whole robot as a single system, treating hands and feet in much the same way.
The companies say this allows new skills to be added quickly from human demonstrations, without writing fresh code, and should scale better as tasks grow in length and complexity.
The work stems from a research partnership formed in October 2024 to speed up progress in smart, general-purpose robots. The teams argue that LBMs could help humanoids operate directly in homes and workplaces that were designed for people—without the need for bespoke programming for each new task.
Boston Dynamics vice-president of robotics research Scott Kuindersma said training one neural network to handle many “long-horizon” manipulation tasks should improve generalisation, with Atlas offering a strong platform for collecting the whole-body data those skills require.
Toyota Research Institute senior vice-president of large behaviour models Russ Tedrake said previous programming methods did not scale, while LBMs let engineers add skills rapidly from human demos; as the models improve, they need fewer demos to achieve more robust behaviour.
Co-led by Kuindersma and Tedrake, the project is investigating how large models can deliver whole-body control, advanced manipulation and dynamic movements in a single system — seen by both teams as a key step towards practical humanoid assistants.
