Neural Networks Based Path Planning and Navigation of Mobile Robots

Abstract

The path planning for mobile robots is a fundamental issue in the field of unmanned vehicles control. The purpose of the path planner is to compute a path from the start position of the vehicle to the goal to be reached. The primary concern of path planning is to compute collision-free paths. Another, equally important issue is to compute a realizable and, if possible, optimal path, bringing the vehicle to the final position. Although humans have the superb capability to plan motions of their body and limbs effortlessly, the motion planning turns out to be a very complex problem. The best known algorithm has a complexity that is exponential to the number of degrees of freedom and polynomial in the geometric complexities of the robot and the obstacles in the environment (Chen & Hwang (1998)). Even for motion planning problems in the 2-dimensional space, existing complete algorithms that guarantee a solution often need large amount of memory and in some cases may take long computational time. On the other hand, fast heuristic algorithms may fail to find a solution even if it exists (Hwang & Ahuja (1992)). In this paper we present a fast algorithm for solving the path planning problem for differential drive (holonomic)1 robots. The algorithm can be applied to free-flying and snake type robots, too. Generally, we treat the two-dimensional known environment, where the obstacles are stationary polygons or ovals, but the algorithm can easily be extended for the three-dimensional case (Kroumov et al. (2010)). The proposed algorithm is, in general, based on the potential field methods. The algorithm solves the local minimum problem and generates optimal path in a relatively small number of calculations. The paper is organized as follows. Previous work is presented in Section 2. In Section 3 we give a definition of the map representation and how it is used to describe various obstacles situated in the working environment. The path and the obstacle collisions are detected using artificial annealing algorithm. Also, the solution of the local minima problem is described there. In Section 4 we describe the theoretical background and the development of a motion