Researchers led by Lechen Zhang have built Body Schema Policy, or BSP, an imitation learning policy for robots that walk and handle objects at the same time, according to a paper on arXiv. It is modelled on the body schema with which people organise what they see and how they move.

BSP places tokens for the joints and links of the robot on its kinematic tree, and each body part fetches its own visual information to decide its actions. It returned action chunks faster than every policy it was compared with. Commands for the hands came to depend mostly on the objects of the task, while those for the legs responded to the whole scene.

On nine simulated whole-body tasks with a Unitree G1 humanoid, BSP did as well as the Ψ0 humanoid foundation model while using 75 times fewer parameters, and it almost doubled the success rate of Diffusion Policy. The results held on a real G1 and on a Fetch mobile manipulator.

Source: arXiv Robotics (cs.RO)

Background

Lechen Zhang

Robotics researcher

Robotics researcher.

Unitree G1

Humanoid robot

The Unitree G1 is a small humanoid robot from China's Unitree Robotics, and research teams, including one at Apple, use it to test new ways of training and controlling humanoids.