Risk assessment and interactive motion planning with visual occlusion using graph attention networks and reinforcement learning

This study proposes an innovative framework that integrates risk assessment and interactive planning for autonomous vehicles (AVs) navigating unprotected left turns at occluded intersections. The upper risk assessment module of this framework synergizes Expert-Informed Graph Attention Networks (EIGAT) with Mixture Density Network (MDN) to predict the probabilistic distributions of the potential risk of the occluded zone. And the lower interactive planning module, utilizing Adaptive Loss Enhanced Reinforcement Learning (ALERL), further develops an interactive policy that integrates additional considerations for prediction accuracy of blind zones, potential risk measure of conditional value at risk (CVaR), and encourage of exploratory interaction. Simulation tests are conducted in occluded intersection scenarios with various traffic density level. Both qualitative and quantitative performance validate the effectiveness and adaptability of our proposed controller in risk assessment and interactive planning for AVs compared with other baseline methods.