With the application of auto-piloting systems on household automobiles, learning-based path prediction systems have drawn much attention. In production environments, safety is always one of the biggest concerns, which requires robust and consistent performance
of systems in safety-critical situations. However, there has been no available dataset that contains driving scenarios with potential safety hazards, which are rarely observed and hard to gather in the real world. Therefore, here we propose to use simulated driving scenarios
as training examples. Specifically, we developed a driving scenario collecting toolkit based on the popular video game Grand Theft Auto V (GTA V), which has realistic physical and graphical modelings for vehicles, pedestrians, and driveways. With the toolkit, we collected real-time navigation trajectories of game agents under safety-critical settings, such as vehicle driving through crossroads with no signals or encountering pedestrians in a close distance. We classified the collected scenarios by the vehicle maneuvers and named the
resulting dataset as GTA P2, reflecting the fact that the dataset was collected on GTA V and that the purpose of the dataset is to facilitate the research in path prediction (PP, or VP2) tasks. To motivate the study of multi-agent path prediction with our dataset, we tested multiple multi-agent-path-prediction models on the dataset, including our newly proposed learning model Message Aggregation Network (MAN). MAN is inspired by Gated Graph Neural Network (GG-NN) and performs message aggregation for agents involved in a scene. It shows state-of-art performance on prediction tasks for the GTA P2 dataset.