Generate, Track, Improve: Perceptive Multi-Skill Humanoid Locomotion with RL-Fine-Tuned Motion Generators

TL;DR

This paper presents a two-layer locomotion architecture for humanoid robots, combining a perceptive motion generator with a tracking policy trained using control-guided RL. The architecture enables multi-skill locomotion and improves terrain consistency on unseen geometries. It achieves successful terrain traversals and skill selection improvements.

Why It Matters

The proposed architecture enables humanoid robots to navigate complex outdoor environments without requiring odometry or height maps. This contributes to the development of more robust and autonomous humanoid robots. The approach can be applied to various robotic platforms, enhancing their locomotion capabilities.

Key Findings

• up to 25% increase in successful terrain traversals
• up to 80% improvement in skill selection
• sample-efficient off-policy RL fine-tuning