Vision-based terrain-aware quadruped locomotion via dual monocular depth fusion and information-decoupled adaptation

Reinforcement learning-based quadruped locomotion control typically relies on ground-truth terrain heightmaps that are readily available in simulation but inaccessible during real-world deployment. Existing solutions either forgo exteroceptive perception or depend on costly depth cameras and LiDAR, which suffer from structural blind spots beneath the robot. This paper proposes a complete, low-cost solution using only monocular RGB cameras. First, we introduce a stereo-matching-supervised domain-adaptive fine-tuning scheme that adapts a lightweight FastDepth network to the downward-facing terrain perception scenario unique to quadruped robots, requiring no manual depth annotation. Second, a dual-camera fusion pipeline with pitch-angle geometric compensation converts monocular depth estimates into policy-compatible terrain heightmaps in real time. Third, we propose an information-decoupled adaptation module that strictly separates terrain observation from dynamics estimation at the history-buffer level, preventing visual noise from degrading online environment parameter inference. Simulation experiments in IsaacGym demonstrate that the full model achieves the best velocity tracking accuracy across five terrain types, and the decoupled adaptation module retains over 93% tracking performance under realistic noise levels. Real-world experiments on a Unitree Go1 robot validate successful stair climbing and proactive obstacle negotiation using only RGB cameras, confirming effective sim-to-real transfer of terrain-aware locomotion.