How LiDAR helps robots navigate new scenes with only a camera

Robots navigating to GPS coordinates fail when moved to new buildings or outdoor spaces because they learn appearance shortcuts specific to training scenes. This work uses LiDAR depth as a teacher during training to guide visual representations toward geometry rather than surface details, then removes the LiDAR at deployment. The frozen encoder feeds into a reinforcement learning policy, keeping representation learning separate from task learning. Testing on diverse simulated environments shows substantial improvements over large pretrained vision models, plus a new multimodal dataset and open-source code.