Abstract

What does bumping into things in a scene tell you about scene geometry? Inthis paper, we investigate the idea of learning from collisions. At the heartof our approach is the idea of collision replay, where we use examples of acollision to provide supervision for observations at a past frame. We usecollision replay to train convolutional neural networks to predict adistribution over collision time from new images. This distribution conveysinformation about the navigational affordances (e.g., corridors vs open spaces)and, as we show, can be converted into the distance function for the scenegeometry. We analyze this approach with an agent that has noisy actuation in aphotorealistic simulator.

 

Quick Read (beta)