Researchers led by Xingyi Wang have built ECHO, a behaviour foundation model for humanoid robots that turns recent interaction experience into control suited to the current conditions, according to a paper on arXiv. It keeps the robot's general range of movements apart from adjustments for specific dynamics.
ECHO first learns a behaviour prior from targeted practice and broad motion coverage. It then reads the dynamics from recent interaction history and corrects the frozen prior without updating its parameters while running. The authors say it outperformed all baselines across motions and physical conditions. On a Unitree G1 it was used for whole-body teleoperation, including acrobatics, and adapted to payloads in manipulation.



