DKWBC splits humanoid control into model-based tracking and a learned balance policy

Researchers led by Dingkun Wei and Xiaowei Zhou have described DKWBC in an arXiv paper, a whole-body control framework for humanoid robots. Reinforcement learning trackers usually learn motion tracking and balance together, which can make learning harder and cause jittery control; DKWBC instead supplies tracking torques from inverse dynamics and joint feedback, and a residual policy learns only the extra torques needed to stay balanced.

In simulation and on real hardware, a single learned balancing policy worked across motion datasets while tracking motions accurately and keeping balance on different ground, with attached payloads and during physical interaction, the authors report.