A Novel Route Planning Approach Based on Energy-Based Action Sample Reinforcement Learning

Route planning for driverless vehicles in given environment has been widely studied. In this paper, a novel route path planning approach is proposed for self-driving based on energy-based action sample reinforcement learning. Firstly, based on the established graph, Q-learning is adopted in agent training with an novel considering steering penalty into reward function. Secondly, an innovative energy-based action sampling method is used to choose the most valuable action. Thirdly, the agent interacts with environment to plan a global path from the start to the goal point. As the results shown in experiments, this approach can complete the route planning task, and perform better in path length and turns (i.e. evaluation indicators).