Endpoint Conditioned Multimodality Trajectory Prediction Using Voronoi Tessellation

Movement and navigation are key aspects of many games. Endpoint and trajectory prediction in games is thus becoming an emerging matter, also because they can serve as important components for downstream tasks such as real-time assistants, AI behaviour selection, or bot detection. However, such predictions can become costly due to the large volumes of data. In this paper, we present first steps towards a lightweight modular pipeline for endpoint and trajectory prediction based on Voronoi tessellation for compact and efficient data storage. The model outputs probability distributions, allowing for multimodality and easy processing by downstream tasks. We illustrate and evaluate the proposed approach using data from the team-based computer game World of Tanks. First results suggest that the proposed pipeline performs well in predicting trajectories, while keeping memory and computation requirements small.