Deep, Consistent Behavioral Decision Making with Planning Features for Autonomous Vehicles

摘要: Autonomous driving promises to be the main trend in future intelligent transportation systems due its potentiality for energy saving, and traffic safety improvements. However, traditional autonomous vehicles’ behavioral decisions face consistency issues between decision trajectory planning shows a strong dependence on human experience. In this paper, we present planning-feature-based deep behavior method (PFBD) complex, dynamic traffic. We used reinforcement learning (DRL) framework with twin delayed deterministic policy gradient algorithm (TD3) exploit optimal policy. took into account features of topological routes making vehicles, through which path layers can guaranteed. Specifically, route extracted from space are shared as input states decision. The actor-network learns near-optimal feasible safe candidate emulated routes. Simulation tests three typical scenarios have been performed demonstrate performance policy, including comparison rule-based expert considering partial information contour. results show that proposed approach achieve better decisions. Real-time test an HQ3 (HongQi third ) vehicle also validated effectiveness PFBD.