Decision-Making in Fallback Scenarios for Autonomous Vehicles: Deep Reinforcement Learning Approach

This paper proposes a decision-making algorithm based on deep reinforcement learning to support fallback techniques in autonomous vehicles. The fallback technique attempts to mitigate or escape risky driving conditions by responding to appropriate avoidance maneuvers essential for achieving a Level 4+ autonomous driving system. However, developing a fallback technique is difficult because of the innumerable fallback situations to address and eligible optimal decision-making among multiple maneuvers. We employed a decision-making algorithm utilizing a scenario-based learning approach to address these issues. First, we crafted a specific fallback scenario encompassing the challenges to be addressed and matched the anticipated optimal maneuvers as determined by heuristic methods. In this scenario, the ego vehicle learns through trial and error to determine the most effective maneuver. We conducted 100 independent training sessions to evaluate the proposed algorithm and compared the results with those of heuristic-derived maneuvers. The results were promising; 38% of the training sessions resulted in the vehicle learning lane-change maneuvers, whereas 9% mastered slow following. Thus, the proposed algorithm successfully learned human-equivalent fallback capabilities from scratch within the provided scenario.