HybridMimic: Hybrid RL-Centroidal Control for Humanoid Motion Mimicking

A research team has introduced HybridMimic, a novel control framework that merges reinforcement learning (RL) with model-based centroidal dynamics to create more stable and physically plausible humanoid robots. The core innovation is a learned policy that dynamically modulates a centroidal-model-based controller by predicting continuous contact states and desired centroidal velocities. This hybrid approach generates feedforward torques grounded in real physics, ensuring commands remain feasible even when the robot encounters unexpected environments, a common failure point for pure RL methods that often bypass explicit dynamics reasoning.

In hardware experiments conducted on the Booster T1 humanoid robot, HybridMimic demonstrated significant practical improvements. It reduced the average base position tracking error by 13% compared to a leading RL-only baseline. The system is trained using physics-informed rewards to efficiently utilize the underlying controller's optimization, outputting precise control targets and reference torques. This represents a key advance over previous hybrid methods, which were often limited by predefined contact timing, by enabling dynamic and versatile full-body motion mimicking from human demonstrations.

The work, published on arXiv, addresses a critical challenge in robotics: bridging the gap between agile learned policies and the hard constraints of the physical world. By ensuring control outputs are informed by centroidal dynamics—which model the robot's overall momentum—HybridMimic enhances robustness against domain shifts. This makes humanoid robots more reliable for complex, real-world tasks where pure simulation-trained policies can fail, marking a step toward more deployable and capable bipedal machines.