A Gradient Descent Method for The Dubins Traveling Salesman Problem

TL;DR

This paper introduces a combined bounding and gradient descent approach to efficiently solve the Dubins TSP, optimizing vehicle paths with curvature constraints for applications in robotics and UAVs.

Contribution

It presents a novel method integrating bounding procedures with gradient descent for curvature-constrained TSP, with linear scaling and high accuracy in practical scenarios.

Findings

Bounding procedure quickly finds the optimal visiting sequence.

Gradient descent converges within 1% of optimal in one iteration.

Method scales linearly with the number of points.

Abstract

We propose a combination of a bounding procedure and gradient descent method for solving the Dubins traveling salesman problem, that is, the problem of finding a shortest curvature-constrained tour through a finite number of points in the euclidean plane. The problem finds applications in path planning for robotic vehicles and unmanned aerial vehicles, where a minimum turning radius prevents the vehicle from taking sharp turns. In this paper, we focus on the case where any two points are separated by at least four times the minimum turning radius, which is most interesting from a practical standpoint. The bounding procedure efficiently determines the optimal order in which to visit the points. The gradient descent method, which is inspired by a mechanical model, determines the optimal trajectories of the tour through the points in a given order, and its computation time scales linearly with the number of points. In experiments on nine points, the bounding procedure typically explores no more than a few sequences before finding the optimal sequence, and the gradient descent method typically converges to within 1\% of optimal in a single iteration.