一段视频，让机器人学会开门并穿越：Video2DoorTraversal 如何打通 Real-to-Sim-to-Real

TL;DR - Video2DoorTraversal is a preprint describing a single-video Real-to-Sim-to-Real framework that reconstructs a door as a simulated digital twin, generates valid trajectories, and trains a wheeled-legged robot to open and pass through it. It reduces real-world data collection and achieved a 96.57% average success rate on five reconstructed doors.

• A standard RGB video provides geometry, scale, camera motion, and articulated components for a simulation-ready DoorTwin with visual and physical properties.
• A simulation agent searches parameterized approach, grasp, handle-turning, pushing, and traversal skills, retaining trajectories that pass task, collision, and kinematic checks.
• The ArticuACT policy uses two depth-camera views and robot state to jointly control the base, arm, and gripper; domain randomization improves transfer robustness.
• The robot succeeded in 169 of 175 trials across five doors and averaged 80.95% zero-shot success on three unseen, structurally similar doors, with each traversal taking about 13 seconds.