This is HOVER: a neural whole-body controller for humanoid robots developed by NVIDIA, Carnegie Mellon University, UC Berkeley, UT Austin, and UC San Diego. It is based on goal-conditioned reinforcement learning for humanoid control. The main idea is to unify different control modes into a single policy, which existing methods handle separately. HOVER unifies kinematic position tracking, joint angle tracking, and root tracking via a multi-mode command space with masking mechanisms.
This approach outperforms task-specific policies in simulation across 12 metrics. This was tested on a Unitree H1 robot. You can read about it at the link below.
[HT]