Direct Perception (DP) is a known approach for autonomous driving in which a monolithic convolutional neural network is used to extract the most useful driving-parameters from raw images. We use a similar new approach that extracts abstract model of the road called Model Perception (MP), and then use continuous control with deep reinforcement learning (DDPG) to train a driving agent on top of this model. The driving agent controls both steering and throttle, it develops a safe and smooth driving behavior, and we show that it can drive on never-before seen tracks, all while using only images as input. Imagry is developing a cameras-only level 4/5 self-driving platform that amounts to a fraction of the cost of traditional LiDar, Radar & HD GPS-based solutions. The AI technology is based on Deep Inverse Reinforcement Learning algorithms which accelerate the training and performance of its unified software solution (end-to-end perception, planning, and control), especially in complex unseen scenarios.
Hall: Hall D