We introduce a general declarative framework for anticipatory reasoning about interactive motion in complex dynamic environments. The aim is to build an explainable space‑time mental model that heuristically guides low‑level motion computation at the semantic level. The framework is hybrid, modular, and elaboration‑tolerant, consisting of three key components:
- An Answer Set Programming (ASP) model that captures the dynamics of interacting motion trajectories, providing a formal representation of multi‑agent interactions;
- Semantically guided prediction that derives "branching trajectory futures", i.e., multiple possible future paths conditioned on different event contexts;
- An energy‑based optimisation model that evaluates, ranks, and filters the anticipated trajectories, retaining those that best satisfy physical constraints and task objectives.
We evaluate the approach on the Argoverse 2 benchmark. Each anticipated trajectory is accompanied by an explicit, inspectable event‑level hypothesis, contrasting with end‑to‑end stochastic predictors. Results demonstrate that the combination of ASP‑driven semantic constraints and energy optimisation improves both interpretability and prediction accuracy.
Review