Human motion data helps a Unitree G1 humanoid estimate its speed without fine-tuning

Researchers led by Yutian Chen have built a unified inertial encoder that uses human motion data to help humanoid robots estimate their own movement and relate it to text descriptions, in a paper posted on arXiv.

The encoder synthesises IMU signals at matching body sites for humans and humanoids, so it can learn from human motion and from humanoid motion retargeted from it. It worked on a physical Unitree G1 with no physics simulation, no training data from the real robot and no fine-tuning.

It had the lowest velocity error of the methods tested on five datasets, including 0.099 m/s on the real G1, and could find descriptions of G1 motion types that appeared only in the human training data.