Interpretable Fuzzy Dynamics Encoding for Multi-Agent Trajectory Prediction in Autonomous Driving

Reliable multi-agent trajectory prediction is essential for autonomous vehicles to anticipate the future motion of surrounding agents in complex road environments. However, many existing methods encode motion dynamics implicitly from coordinate sequences, limiting their ability to represent behavior-level patterns with explicit physical meaning. This study proposes FZNet, an interpretable fuzzy dynamics encoding framework that extracts velocity, angular velocity, and acceleration from historical trajectories and transforms these continuous kinematic variables into semantic motion representations using learnable fuzzy membership functions and differentiable rule inference. The resulting motion embeddings are integrated with lane-level map features to construct a behavior-aware representation that jointly captures dynamic motion tendencies and road constraints. On the Argoverse benchmark, FZNet achieves a minimum average displacement error of 0.868, a minimum final displacement error of 1.477, and a miss rate of 0.184. Removing the fuzzy module increases the corresponding displacement errors to 0.872 and 1.488, while replacing fuzzy rule inference with multilayer perceptron aggregation produces errors of 0.875 and 1.492. Across five matched random seeds, FZNet shows statistically significant but small improvements over the no-fuzzy and MLP aggregation variants after Holm correction. These results indicate a modest and repeatable benefit of fuzzy rule inference in the evaluated setting. The proposed framework offers an explicit connection between physical kinematic variables and behavior-aware representations while maintaining comparable trajectory prediction accuracy for autonomous-driving applications.