An ASP Framework for Predictive Reasoning About Interacting Moving Objects

Summary

The paper presents a general declarative framework for anticipatory reasoning about interacting motion in complex dynamic environments. It aims to form explainable semantic space-time mental models that can heuristically guide lower-level motion computation. The hybrid, modular, and elaboration-tolerant method combines three elements: an ASP model for characterizing the dynamics of interactive motion trajectories, semantically guided prediction for generating branching trajectory futures, and an energy-based optimization model for evaluating, ranking, and filtering those anticipations. The framework is evaluated on the Argoverse 2 benchmark. Unlike end-to-end stochastic prediction, each anticipated trajectory includes an explicit, inspectable event-level hypothesis, making the reasoning behind a predicted future more accessible for analysis.