End-to-End Autonomous Driving Decision Based on Deep Reinforcement Learning

End-to-end autonomous driving decision-making is a popular research field in autonomous driving. In this paper, we propose an end-to-end decision-making model based on DDPG deep reinforcement learning. Firstly, we establish an end-to-end decision-making model to map driving state (such as tangential angle of vehicle, velocity of vehicle, distance of road) to driving action (steer, accelerate, brake) continuously. Next we train and valid our agent in different scenarios on TORCS platform. The results show that DDPG algorithms can achieve end-to-end autonomous driving decisions. Finally, we visualize the agent by analyzing which state contributes to decision.