Decision-making in Autonomous Driving by Reinforcement Learning Combined with Planning & control

In recent years, there are more and more studies on reinforcement learning (RL) and RL algorithms have achieved good performance in autonomous driving simulation experiments. But the simulation experiments are too idealistic, their safety cannot be guaranteed in practical application, with the action implementation error not being taken into account. To this end, a method combining the RL-based decision-making model with the planning & control module is proposed in this paper. In this paper, DDQN is selected as the algorithm for model-free deep reinforcement learning (DRL) decision-making model, A* is selected as the algorithm of path planner and PID is selected as the algorithm of controller, which is demonstrated in CARLA, to make accurate decisions while considering the real vehicle dynamics model and action implementation error, and improve the robustness and realistic transferability of the RL decision-making algorithm. Several improvement measures are proposed and experiment shows that the proposed GRSD-DDQN has better test performance with the free collision rate improved from 81.5% to 98.6% and the average speed improved from 45.5 (km/h) to 53.7 (km/h) compared with the Baseline-DDQN.