Unifying Artificial Intelligence and Trajectory Optimization for UAV Guidance

Abstract

The envisioned scope of autonomy of Unmanned Aerial Vehicles (UAVs) has broadened from autonomous point-to-point motion to intelligent motion, which involves the completion of a high-level task. Algorithmic frameworks for defining and completing such tasks have been developed in the theory of artificial intelligence (AI). One such framework is of classical planning problems (CPP)s. We discuss the applicability of CPPs to formulate new types of complex UAV tasks. Next, we present a new technique to unify trajectory optimization algorithms with solutions of CPPs, thereby introducing a new class of UAV guidance techniques. The proposed approach is particularly suited to leverage existing algorithms for solving CPPs and, independently, algorithms for trajectory optimization. We introduce a family of graphs called lifted planning graphs parametrized by an integer H, and we map paths in these graphs to solutions of the CPP. We show that the overall cost of a high-level plan is a nonincreasing function of H, and that there exists a finite H for which an optimal path in the lifted planning graph is associated with the optimal solution of the CPP. We illustrate the proposed ideas with numerical simulation examples involving task-planning for a point-mass UAV model.