Deep Learning Model Predictive Control for Autonomous Driving in Unknown Environments

Abstract A dynamic obstacle avoidance Model Predictive Control (MPC) method is introduced for autonomous driving that uses deep learning technique for velocity-dependent collision avoidance in unknown environments. The objective of the method is to control an autonomous vehicle in order to perform different traffic maneuvers in a safe way with maximum comfort of passengers, and in minimum possible time, accounting for maneuvering capabilities, vehicle dynamics, and in the presence of traffic rules, road boundaries and static and dynamic unknown obstacles. Here, by defining local coordinates and collision regions, the dynamic collision avoidance problem is translated into a static collision avoidance problem which makes the method easier and faster to be solved in dynamical environments. In order to provide safety, an ensemble of deep neural networks is used to estimate the probability of collision and to form an uncertainty-dependent collision cost which prioritizes between mission and safety. The collision cost is a product of the probability of collision and vehicle’s velocity in the directions with high collision-risk. The dynamic obstacle avoidance optimization method minimizes the velocity in the obstacle cones where the probability of collision is high or in unfamiliar environments, and increases the velocity when probability and variation in predicted values of the ensemble are low. The predicted trajectory from MPC is used in learning procedure in order to assign labels that makes it possible to predict the collision in advance. Simulation results show that the proposed method has good adaptability to unknown environments.