Researchers led by Sol Choi have found a way to reuse one reinforcement-learning recipe for humanoid walking on robots whose leg joints are laid out differently, according to a paper on arXiv. Rewards, action scales and other settings are usually tied to each robot's own joint coordinates, which makes recipes hard to move from one robot to another.

The team adds a body-aligned virtual joint space: the policy's actions and observations follow one anatomical convention, and a fixed kinematic layer turns the virtual targets into each robot's own joint angles, leaving the rest of the control stack as it is. With one unchanged recipe on three humanoids, it was the only interface that tracked commands accurately, kept the feet pointing in the direction of travel and trained stably on all three.

On a real Unitree G1, it reduced heading drift by nearly a factor of six.

Source: arXiv Robotics (cs.RO)

Background

Sol Choi

Robotics researcher

Researcher at the Korea Institute of Science and Technology and Yonsei University.

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.

Unitree Robotics

Robot maker

Unitree Robotics is a Hangzhou company that makes four-legged robot dogs and humanoid robots, including the G1, which research teams at Apple and other labs use.