Differentiable Raycasting for Self-Supervised Occupancy Forecasting

Motion planning for safe autonomous driving requires learning how the environment around an ego-vehicle evolves with time. Ego-centric perception of driveable regions in a scene not only changes motion actors environment, but also movement itself. Self-supervised representations proposed large-scale planning, such as ego-centric freespace, confound these two motions, making representation difficult to use downstream planners. In this paper, we geometric occupancy natural alternative view-dependent freespace. Occupancy maps naturally disentagle from ego-vehicle. However, one cannot directly observe full 3D (due occlusion), it signal learning. Our key insight is differentiable raycasting “render” future predictions into LiDAR sweep predictions, which can be compared ground-truth sweeps self-supervised The allows emerge internal within forecasting network. absence groundtruth occupancy, quantitatively evaluate raycasted and show improvements upto 15 F1 points. For planners, where emergent used guide non-driveable regions, relatively reduces number collisions objects by up 17% freespace-centric