PlayGen-MoG: Framework for Diverse Multi-Agent Play Generation via Mixture-of-Gaussians Trajectory Prediction

Kevin Song's PlayGen-MoG framework addresses a critical gap in sports AI: generating diverse, coordinated multi-agent plays from static formations alone. Traditional approaches like Conditional Variational Autoencoders (CVAE) and diffusion models struggle with this task, often suffering from posterior collapse or converging to average plays. PlayGen-MoG's three key innovations—a shared Mixture-of-Gaussians output head, relative spatial attention encoding player relationships, and non-autoregressive displacement prediction—enable it to generate eight distinct play scenarios while maintaining realistic spatial coordination between all players.

On American football tracking data, PlayGen-MoG demonstrates impressive performance with 1.68 yard Average Displacement Error (ADE) and 3.98 yard Final Displacement Error (FDE). Crucially, it achieves full utilization of all eight mixture components with an entropy score of 2.06 out of 2.08, proving it avoids the mode collapse that plagues other generative models. By eliminating dependence on observed trajectory history and predicting absolute displacements directly from initial formations, the framework enables coaches and analysts to design and visualize plays from scratch rather than just forecasting from existing game footage.

The framework's architecture represents a significant technical advancement in multi-agent trajectory generation. The shared mixture weights across all agents ensure that a single play scenario selection couples every player's movements coherently, while the relative spatial attention mechanism learns pairwise player positions and distances as attention biases. This combination allows PlayGen-MoG to generate plays that are both diverse and strategically plausible, making it valuable for play design, opponent simulation, and training scenario generation in professional sports analytics.