Autonomous Vehicles Roundup Strategy by Reinforcement Learning with Prediction Trajectory

Autonomous vehicles are increasingly applied on many situations, but their autonomous decision-making ability needs to be improved. Multi-Agent Deep Deterministic Policy Gradient(MADDPG) adopts the method of centralized evaluation and decentralized execution, so that the autonomous vehicle can obtain the whole-field status information and make decisions through the companion information. In the process of autonomous vehicle training, we introduce artificial potential field, action guidance and other methods to alleviate the problem of sparse rewards. At the same time, we add a repulsion function to consider the relationship between team vehicles. Extended Kalman Filter(EKF) is also applied to predict the autonomous vehicle trajectory, changing the training network state input information. At the same time, secondary correction of the predicted autonomous vehicle trajectory is made to change the prediction range with the training time, and improve the training convergence speed while the speed of opposite agents increases. Simulation experiments show that the convergence speed and win rate of MADDPG algorithm based on trajectory prediction and artificial potential field is significantly improved, and it also has strong adaptability to various task scenarios.