Learning-based MPC for Autonomous Motion Planning at Freeway Off-ramp Diverging

Off-ramp diverging road segment is the preparation area for vehicles driving away from the freeway, while it causes more traffic conflicts making it a typical safety bottleneck. Focusing on improving driving safety, a Learning-based Model Predictive Control (LMPC) strategy is proposed to deal with the motion planning of autonomous vehicles (AVs) considering the dynamic driving behavior of surrounding traffic occupants in this area. The traditional MPC is combined with the reinforcement learning (RL) method to impose constraints for limitations to ensure safety while reducing conservatism. The Q-network of RL is used to build a terminal cost function to dynamically adjust the scale of the prediction horizon, which improves the speed of convergence while maintaining exploration of LMPC. Experimental testing scenarios are setup with considering different driving behaviors of surrounding vehicles to validate the performance of the proposed approach. The simulation results show that the LMPC can stably and flexibly plan the motion behavior of AVs, achieving a planning accuracy of 99.31%, which has the potential for practical application.