An Integrated Reward Function of End-to-End Deep Reinforcement Learning for the Longitudinal and Lateral Control of Autonomous Vehicles.

Based on the end-to-end deep reinforcement learning framework, we present an integrated reward function (IRF) that improves the driving safety and efficiency of autonomous vehicles. Its longitudinal component employs an incremental slope incentive as well as the incentive/penalty of safety distance, while the lateral components exploits a symmetric piecewise-linear reward function to model the lateral and heading deviations from the lane centerline. The CARLA simulation results show that the IRF agent vehicle can achieve an excellent driving performance in both longitudinal and lateral directions.