MGDP: Mastering a Generalized Depth Perception Model for Quadruped Locomotion

ABSTRACT Perception‐based Deep Reinforcement Learning (DRL) controllers demonstrate impressive performance on challenging terrains. However, existing controllers still face core limitations, struggling to achieve both terrain generality and platform transferability, and are constrained by high computational overhead and sensitivity to sensor noise. To address these challenges fundamentally, we propose a generalized control framework: Mastering a Generalized Contrastive Depth Model (MGDP). We leverage NVIDIA Warp to enable efficient parallel computation of depth images, thereby mitigating the inherent high computational cost. MGDP extracts low‐dimensional terrain feature representations from multi‐modal inputs (depth images and height maps) and integrates an explicit depth map denoising mechanism. This process not only facilitates effective decoupling of perception from dynamics but also significantly reduces the memory. Furthermore, we design terrain‐adaptive reward functions that modulate penalty strengths according to terrain characteristics, enabling the policy to acquire complex locomotion skills (e.g., climbing, jumping, crawling, squeezing) in a single training stage without relying on distillation. Experimental results demonstrate that MGDP not only endows the policy with superior cross‐terrain generalization capability but also enables fast and efficient fine‐tuning across diverse quadruped robot morphologies via its pre‐trained, dynamics‐decoupled perception model. This vigorously advances the development of unified, efficient, and generalized frameworks for quadrupedal locomotion control.