Perpetual Humanoid Control

A physics-based humanoid controller that tracks huge libraries of human motion in simulation and gets back up on its own after falling.

PHC (Perpetual Humanoid Control) is a physics-based motion-imitation controller from Zhengyi Luo and colleagues at Carnegie Mellon University and Meta Reality Labs, presented at ICCV 2023 and trained with reinforcement learning in Isaac Gym. The goal is for a simulated humanoid to track noisy reference motions in real time — motions coming from video-based pose estimation or text-to-motion generation — without any external stabilizing forces. The key idea is Progressive Multiplicative Control Policy (PMCP): whenever the controller hits a motion it cannot learn, a new sub-network is added to handle it, which lets it learn nearly 10,000 motion clips from the AMASS motion-capture library without catastrophic forgetting (losing previously learned skills while acquiring new ones). The paper reports a 98.9% success rate on the training set and 96.4% on the held-out test set. PHC also learned to recover naturally after falling and resume tracking. Its open-source code was later extended to Unitree's H1 and G1 humanoid models, and follow-up work such as PHC+ and PULSE builds on it.

ExampleA simulated character is driven in real time by poses estimated from a monocular video; if it gets tripped partway through, PHC lets it stand back up on its own and keep tracking the reference motion from where it left off.

Also called

PHC, Perpetual Humanoid Control for Real-time Simulated Avatars

Related

Motion Tracking · DeepMimic · MaskedMimic · AMASS (Archive of Motion Capture as Surface Shapes) · H2O · Catastrophic Forgetting

Sources

arXiv 2305.06456: Perpetual Humanoid Control for Real-time Simulated Avatars
GitHub: ZhengyiLuo/PHC

As of

2026-09