Multi-Goal Motion Planning with Physics-based Game Engines

Toward enhancing automation in video games, this paper proposes an efficient approach for multi-goal motion planning, where a mobile agent needs to visit several regions in a complex environment containing numerous obstacles. The approach works in conjunction with differential equations and physics-based simulations of vehicle dynamics, efficiently planning collision-free, dynamically-feasible, and low-cost solution trajectories. We combine discrete search with sampling-based motion planning to map this challenging task to graph search. The approach imposes a discrete abstraction obtained by a workspace decomposition and then precomputes shortest paths to each goal. As the sampling-based motion planner expands a tree of collision-free and dynamically-feasible trajectories, it relies on a fast TSP solver to compute low-cost tours which can effectively guide the motion-tree expansion. The tours are adjusted based on progress made and a partition of the motion tree into equivalent groups, giving the approach the flexibility to discover new tours that are compatible with the vehicle dynamics and collision constraints. Comparisons to related work show significant computational speedups and reduced solution costs.