Predictive model for hazardous scenarios in autonomous driving based on graph neural networks and variational autoencoders

The evaluation and improvement of autonomous vehicles necessitate the scalable generation of realistic and challenging hazardous scenarios that can be safely resolved. Current research often employs traditional kinematic models to control a single planner, without filtering the generated scenarios for effectiveness. This study introduces a more expressive, datadriven motion prior to control multiple adversarial agents and proposes a novel method for automatically generating challenging scenarios that induce adverse behaviors in a given planner. This paper utilizes an approach that integrates Convolutional Neural Networks (CNN), Graph Neural Networks (GNN), and Conditional Variational Autoencoders (CVAE). Initially, CNNs extract environmental features for each agent from data sources. Subsequently, GNNs are introduced to understand and leverage the interactions among agents. Finally, CVAEs are used to optimize the latent space of the traffic model to generate trajectories that collide with the given planner. Experimental results demonstrate that this method can generate challenging yet solvable scenarios. It offers a new approach for constructing hazardous scenarios for autonomous vehicles, as well as for behavior planning and trajectory optimization, thereby opening new possibilities for simulation testing of autonomous vehicles.