Risk-Aware High-level Decisions for Automated Driving at Occluded Intersections with Reinforcement Learning

Reinforcement learning is nowadays a popular framework for solving different\ndecision making problems in automated driving. However, there are still some\nremaining crucial challenges that need to be addressed for providing more\nreliable policies. In this paper, we propose a generic risk-aware DQN approach\nin order to learn high level actions for driving through unsignalized occluded\nintersections. The proposed state representation provides lane based\ninformation which allows to be used for multi-lane scenarios. Moreover, we\npropose a risk based reward function which punishes risky situations instead of\nonly collision failures. Such rewarding approach helps to incorporate risk\nprediction into our deep Q network and learn more reliable policies which are\nsafer in challenging situations. The efficiency of the proposed approach is\ncompared with a DQN learned with conventional collision based rewarding scheme\nand also with a rule-based intersection navigation policy. Evaluation results\nshow that the proposed approach outperforms both of these methods. It provides\nsafer actions than collision-aware DQN approach and is less overcautious than\nthe rule-based policy.\n