GINK: Graph-based Interaction-aware Kinodynamic Planning via Reinforcement Learning for Autonomous Driving

Applying reinforcement learning to autonomous driving entails certain challenges, primarily due to massive traffic flows, which change dynamically. To address such challenges, it is necessary to quickly determine response strategies to the changing intentions of surrounding vehicles. Accordingly, we propose a new policy optimization method for safe driving using graph-based interaction-aware constraints. In this framework, the motion prediction and control modules are trained simultaneously, while sharing a latent representation that contains a social context. Further, to reflect social interactions, we express the movements of agents in the graph form and filter the features. This helps preserve the spatiotemporal locality of adjacent nodes. Furthermore, we create feedback loops to combine these two modules effectively. As a result, this approach encourages the learned controller to be safe from dynamic risks and also renders the motion prediction robust under various situations. In the experiment, we set up a navigation scenario comprising various situations, with CARLA, an urban driving simulator. The experiments show state-of-the-art performance on the sides of both navigation strategy and motion prediction compared to the baselines.