Learning-Based Whole-Body Control

A neural network, trained with reinforcement learning in simulation, that coordinates a humanoid robot's entire body from one policy.

Traditional whole-body control (WBC) solves a prioritized quadratic program from a dynamics model every control cycle to get full-body joint torques, which requires an accurate model and heavy tuning. Learning-based whole-body control instead trains a neural-network policy with reinforcement learning in simulation: it takes in the robot's own state plus a high-level command (walking speed, joint angles, or target positions for the hands and head, sometimes a snippet of human motion) and outputs full-body joint target positions, which per-joint PD controllers then execute; training commonly adds domain randomization to help the policy transfer to the real robot. The motion data used for training is usually retargeted from human motion-capture datasets such as AMASS. Notable examples include ExBody, OmniH2O, HOVER, and SONIC. It often serves as the ‘cerebellum’ in a brain-cerebellum split, receiving high-level commands from a VLA model or a teleoperator.

ExampleHOVER (Tairan He et al., ICRA 2025) unified several command modes — root-velocity tracking, local joint-angle tracking, and keypoint-position tracking — into one policy on a 19-degree-of-freedom Unitree H1, using masking, and distilled it from a privileged teacher policy via DAgger; the unified policy beat specialized controllers on at least 7 of 12 metrics under each mode.

Also called

Neural WBC, Neural Whole-Body Controller, RL Whole-Body Control

Related

Whole-Body Control · HOVER · Motion Tracking · Teacher-Student Distillation · RL-based Locomotion Control · Brain–Cerebellum Architecture

Sources

HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots (arXiv 2410.21229)
HOVER 全文（HTML 版，动作空间与 DAgger 蒸馏细节） (Chinese)
A Survey of Behavior Foundation Model: Next-Generation Whole-Body Control System of Humanoid Robots (arXiv 2506.20487)

As of

2025-11