CATR: Confidence-Aware Terrain Reconstruction for Vision-Based Quadruped Locomotion

Abstract

In vision-based legged locomotion, terrain maps provide essential environmental cues for robots to perceive and traverse upcoming terrain. However, exteroceptive perception is inherently uncertain due to sensor noise, occlusions, and missing depth measurements, which often leads to inaccurate terrain representations and consequently unstable locomotion. To address these challenges, we propose CATR, a learning-based terrain reconstruction framework that uses an observability-based confidence mask to improve robustness under occlusions, missing depth measurements, and sensing noise. The framework combines confidence-aware perception with a FiLM-gating mechanism. The former encodes observation support to distinguish directly observed, missing, and VAE-completed terrain regions, while the latter adaptively modulates terrain features using proprioceptive context. Extensive simulation and real-world experiments on the Unitree Go1 show that CATR enables zero-shot real-world deployment without real-world policy fine-tuning, achieving robust traversal across diverse challenging terrains.