Cross-Modal Distillation With Latent Consistency for Proprioceptive Quadruped Locomotion

Abstract

Proprioception-only locomotion over complex terrain is essential for quadruped robots when external perception is unavailable or unreliable. Existing methods either rely on privileged physical quantities for supervision or infer environmental information from short-term state evolution, which may lead to overemphasis on physical-parameter regression or insufficient environmental semantic guidance. To address these limitations, this paper proposes a proprioceptive reinforcement learning framework that combines cross-modal distillation with latent consistency. During training, privileged information is used as a semantic anchor to guide the encoded proprioceptive history, enabling the policy to learn task-relevant environment-aware latent representations without requiring privileged inputs during deployment. Meanwhile, latent consistency improves the temporal stability and compactness of the learned representations, helping capture low-frequency environmental semantics over complex terrain. Experiments show that the proposed method improves training efficiency and deployment performance, achieving lower velocity tracking error, yaw error, lateral displacement, and peak motor torque during stair traversal. Foot-trajectory visualization and unseen-terrain tests further demonstrate more stable gait patterns and stronger zero-shot terrain adaptability.