Reasonable Motion: A General ASP Foundation for Environment Constrained Movement Trajectory Computation

This paper introduces a hybrid architecture for motion prediction that addresses the lack of interpretability in purely data-driven, end-to-end machine learning models. The approach operates in four distinct stages: first, it grounds the scene into a typed directed graph representing the environment; second, it uses Answer Set Programming (ASP) to non-deterministically generate 'reasonable' motion modes that satisfy both physical and normative constraints; third, it groups these modes based on high-level event sequences to reduce complexity; and finally, it converts these symbolic modes into continuous, smooth geometric trajectories.

In safety-critical domains like autonomous driving, knowing why an agent chooses a specific path is as important as the path's geometric accuracy. Current end-to-end stochastic models often function as black boxes, making it difficult to verify if a predicted trajectory adheres to traffic norms or environmental constraints. By using ASP, this framework provides a formal, verifiable link between the agent's high-level decision-making (e.g., 'turn left at the intersection') and its low-level movement, offering a transparent alternative to purely learned approaches.

The authors demonstrate the method's effectiveness using the Argoverse 2 benchmark, a large-scale dataset for autonomous driving. The evaluation highlights how the system can systematically enumerate distinct, geometrically admissible behaviors while maintaining traceability, allowing developers to inspect the underlying stable models that justify each predicted trajectory.