RallyDiffuser: A Representation-Guided Diffusion Model Framework for Strategic Planning in Badminton

The rising interest in sports analysis has led to many studies from various perspectives, such as strategic insights and behavior prediction. In the rapid tactic nature of turn-based sports, badminton stands out as a compelling example of a game requiring players to make strategy-oriented decisions. Exiting planning works fail to capture its complex decision-making dynamics, particularly in balancing long-term strategy execution with immediate scoring opportunities in a turn-based setting. In this work, we propose RallyDiffuser, an innovative representation-guided diffusion model for strategic planning in badminton. We build a strategy latent space through representation learning that captures the variations in player strategies executed during rallies, and it identifies strategic anchors that guide agents in balancing long-term strategic objectives with short-term scoring opportunities. Our experiments demonstrate that RallyDiffuser outperforms existing planning methods, emerging as the only approach that achieves improved win rates across all strategies.