Agile Legged Locomotion Control Based on Bootstrap Your Own Latent

This study presents a novel locomotion control framework that enhances the agility and robustness of legged robots by learning to encode their dynamic responses to environmental disturbances into latent features for control. Our method leverages bootstrap your own latent-based contrastive learning to derive hybrid latent embeddings from the proprioceptive history, integrating explicit velocity signals with implicit stability-related features. This approach offers the following three key advantages: first, it relies solely on proprioceptive input, eliminating the need for privileged environmental information; second, it maintains consistent observations between simulation and reality, enabling zero-shot sim-to-real transfer; and third, it demonstrates strong robustness to noise and improved sample efficiency. Experimental results on a Unitree Go2 robot demonstrated that the policy quickly adapts to various disturbances and traverses complex terrains with agility. Extensive real-world tests further confirmed the strong generalization of the proposed method in open environments, where it performed reliably even in challenging and previously unseen scenarios.