Multi-agent trajectory generation in team sports presents a complex challenge, demanding models capable of capturing both the diversity of possible plays and realistic spatial coordination among players. However, standard generative approaches like Conditional Variational Autoencoders (CVAE) and diffusion models often struggle with this task, exhibiting issues such as posterior collapse or convergence to the dataset mean.
Furthermore, most existing trajectory prediction methods operate within a forecasting regime that necessitates multiple frames of observed history. This significantly limits their utility for practical play design scenarios where only the initial formation is available.
To address these challenges, researchers have introduced PlayGen-MoG, an extensible framework for formation-conditioned multi-agent play generation. PlayGen-MoG tackles these limitations through three key design choices:
- **A Mixture-of-Gaussians (MoG) output head:** This component utilizes shared mixture weights across all agents, where a single set of weights effectively selects a play scenario that couples the trajectories of all players, promoting diverse generation.
- **Relative spatial attention:** This mechanism encodes pairwise player positions and distances as learned attention biases, crucial for understanding inter-player relationships.
- **Non-autoregressive prediction:** PlayGen-MoG directly predicts absolute displacements from the initial formation. This design choice eliminates cumulative error drift and removes the dependency on observed trajectory history, thereby enabling realistic play generation from a single static formation alone.
Evaluated on American football tracking data, PlayGen-MoG demonstrated robust performance. It achieved a 1.68-yard Average Displacement Error (ADE) and a 3.98-yard Final Displacement Error (FDE). Crucially, the framework maintained full utilization of all 8 mixture components with an entropy of 2.06 out of 2.08, qualitatively confirming its capability for diverse generation without succumbing to mode collapse.