A Novel Hybrid Framework for Motion Planning in Autonomous Vehicles Using Reinforcement and Imitation Learning

Abstract

Autonomous cars are considered to revolutionize the transportation sector significantly, increasing safety and accessibility, particularly within the smart city context. The current paper proposes a novel motion planning approach for self-driving cars and integrates both Deep Reinforcement Learning (DRL) and Deep Imitation Learning (DIL) approaches This proposed Hybrid DRL-DIL algorithm aspires to improve the decision-making capabilities of self-driving vehicles through the integration of Reinforcement Learning together with Imitation Learning strategies. The model has been trained in an urban traffic environment using the Robotics System Toolbox in MATLAB along with high-end simulation environments of Unity to create a realistic simulated urban driving scenario. Furthermore, the Deep Learning Toolbox in MATLAB is used for the implementation and training of both DRL and DIL models. The implemented model reduced the collision frequency by 35% and the lane-keeping precision by 20% compared to the existent rule-based models.