Learning Reusable Hybrid Motion Priors for Humanoid Locomotion from Motion Imitation

How can humanoid robots acquire diverse, reusable motor skills that are not tied to a specific task, enabling them to perform various locomotion objectives without needing to relearn basic balance and coordination for every new command?

The authors propose a three-stage pipeline to create a Hybrid Motion Prior (HMP). First, an expert policy is trained via reinforcement learning to imitate human motion-capture data. Second, this expert is distilled into a frozen architecture consisting of a proprioceptive encoder, a residual vector-quantized (RVQ) codebook, and an action decoder. This distillation uses the 'rotation trick' to improve gradient flow through the discrete bottleneck. Third, task-level policies are trained to solve specific objectives—such as velocity tracking, point-goal navigation, and fall recovery—by selecting discrete indices from the frozen codebook, effectively using the HMP as a modular motor interface.

The HMP successfully preserves the motor competence of the original imitation expert while providing a flexible interface for downstream tasks. The learned codebook exhibits an interpretable structure where the number of active RVQ stages modulates gait patterns. The system was successfully deployed on a real Unitree G1 robot for velocity tracking, demonstrating that the same frozen prior can be reused across different tasks without retraining the low-level motor controller.

This work addresses the inefficiency of training separate policies for every new robotic task. By decoupling motor competence (the HMP) from task-level decision-making, the approach enables more efficient learning of complex behaviors and facilitates the transfer of human-like motion skills to real-world humanoid platforms.

Key Terms

Hybrid Motion Prior (HMP)

A frozen, reusable architecture that maps proprioceptive observations and discrete codebook selections to joint-level motor actions.

Residual Vector Quantization (RVQ)

A technique that decomposes a latent vector into a sequence of discrete codebook entries, allowing for hierarchical refinement of motion intent.

Rotation Trick

A gradient estimation technique used during the training of discrete codebooks to improve the alignment between the latent space and the discrete code vectors.

Motion Imitation

A reinforcement learning paradigm where a robot learns to perform tasks by tracking reference trajectories derived from human motion-capture data.

Proprioceptive Encoder

A component of the HMP that processes the robot's internal state (e.g., joint positions and velocities) to provide a baseline for motion generation.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.