Multi-Gait Learning for Humanoid Robots Using Reinforcement Learning with Selective Adversarial Motion Prior

A research team has published a paper detailing a novel reinforcement learning (RL) framework that enables a humanoid robot to learn five distinct locomotion gaits within a single, unified policy. The approach, centered on a 'selective Adversarial Motion Prior (AMP)' strategy, intelligently applies adversarial training only to periodic, stability-critical motions like walking and stair climbing. This selective application accelerates learning and suppresses erratic behavior for these gaits, while deliberately omitting AMP for highly dynamic skills like running and jumping, where its regularization would over-constrain and limit agility.

The policies were trained in simulation using Proximal Policy Optimization (PPO) with domain randomization and then deployed via zero-shot sim-to-real transfer onto a physical 12-degree-of-freedom (DOF) humanoid robot. Quantitative results show the selective AMP method outperforms a policy using a uniform AMP approach across all five gaits. It achieved faster convergence, lower tracking error, and higher success rates on stability-focused tasks without sacrificing the performance of dynamic ones, demonstrating a more efficient and capable framework for versatile robot locomotion.