Hell, you could even use slower offline 3d reconstruction of vision data for training, and still ultimately rely on runtime LiDAR.