ETD

• bipedal‑wheeled robot
• deep reinforcement learning
• depth camera
• end‑to‑end RL
• Perceptive locomotion
• sim‑to‑real

Wheeled-legged robots offer unprecedented agility by combining the energetic efficiency of wheels with the terrain versatility of legs. However, executing highly dynamic maneuvers across complex 3D obstacles remains an unsolved challenge. Traditional perception pipelines are often too slow to support the highly dynamic and underactuated nature of this locomotion. In this paper, we present the first Deep Reinforcement Learning (DRL) framework that enables perceptive 3D parkour for wheeled-legged robots directly from egocentric depth-camera inputs. To solve the active perception problem and bridge the information gap between privileged training and vision-based deployment, we propose a novel multi-teacher distillation architecture combined with a student-informed teacher mechanism. Extensive evaluations demonstrate that our approach significantly outperforms proprioceptive-only baselines, successfully unlocking complex perceptive 3D parkour behaviors for wheeled-legged systems.