Perceive, Predict, and Plan: Safe Motion Planning Through Interpretable Semantic Representations

In this paper we propose a novel end-to-end learnable network that performs\njoint perception, prediction and motion planning for self-driving vehicles and\nproduces interpretable intermediate representations. Unlike existing neural\nmotion planners, our motion planning costs are consistent with our perception\nand prediction estimates. This is achieved by a novel differentiable semantic\noccupancy representation that is explicitly used as cost by the motion planning\nprocess. Our network is learned end-to-end from human demonstrations. The\nexperiments in a large-scale manual-driving dataset and closed-loop simulation\nshow that the proposed model significantly outperforms state-of-the-art\nplanners in imitating the human behaviors while producing much safer\ntrajectories.\n