Path planning using a Multiclass Support Vector Machine

Highlights

• Generation of a smooth path based on the hyperplane generated by the SVM training stage.

• The margin maximization constraint of the SVM ensures that the generated path will be as far as possible from the obstacles in the environment.

• Robustness against uncertainty and the noise generated by sensors.

• Generation of a decision graph that allows using higher level methods to select the final path.

• The method is compared with a previous approach and in real world conditions.

Abstract

In this paper, a new path planning algorithm for unstructured environments based on a Multiclass Support Vector Machine (MSVM) is presented. Our method uses as its input an aerial image or an unfiltered auto-generated map of the area in which the robot will be moving. Given this, the algorithm is able to generate a graph showing all of the safe paths that a robot can follow. To do so, our algorithm takes advantage of the training stage of a MSVM, making it possible to obtain the set of paths that maximize the distance to the obstacles while minimizing the effect of measurement errors, yielding paths even when the input data are not sufficiently clear. The method also ensures that it is able to find a path, if it exists, and it is fully adaptable to map changes over time. The functionality of these features was assessed using tests, divided into simulated results and real-world tests. For the latter, four different scenarios were evaluated involving 500 tests each. From these tests, we concluded that the method presented is able to perform the tasks for which it was designed.

Introduction

From its inception, research on mobile robots has generated great expectations. To this day the field remains wide open and is constantly evolving. Mobile robots are used in applications such as industrial processes, oceanic and planetary exploration, military projects, rescue operations, and many more. One of the most interesting areas related to mobile robotics, one that has seen considerable research since the 1960s, is path planning. This area remains largely untested, however, and many methods are still under development, with some improving on previous methods or complementing existing techniques.

The problem we wish to address with the method presented in this paper is the generation of a trajectory which allows reaching any point on a map starting from the current position of a vehicle. The challenge is to be able to do this even if the map provided is noisy and unstructured. To this end, an innovative algorithm that combines techniques typically related to Artificial Intelligence and Machine Learning is used together with classical methods from graph theory. The main idea is to employ an SVM classifier in order to generate all the possible safe paths that a robot can follow between obstacles, creating a simple, weighted graph that allows the robot to easily compute the best paths towards a given destination.

The method described in this paper uses a map as its input. Obstacles in the map are represented and transformed into point features, which are tagged using a different label for each obstacle (or class). By doing so, every feature belonging to the same obstacle will have the same label, which will be unique to each obstacle. These generated features are used to train an SVM, yielding a hyperplane that divides all these classes, ensuring that the distance from each hyperplane to the nearest feature of each class is maximized. The intersections of this hyperplane with the plane of the map are extracted and joined together to form a graph (which is then cleaned, for optimization purposes). Edges in the graph will include the value that is desired to be minimized (in our tests, length of the sub-path). Once this graph is complete, it is quite easy and fast to travel along the map.

This document is divided as follows: in Section 2 a review of previous path planning related algorithms is presented. In Section 3 a Multiclass Support Vector Machine (MSVM) is briefly explained. The different steps of the algorithm are described in Section 4 and in Section 5 we show the results of several real application experiments. At the end of the document, in Section 6, we present some conclusions and discuss the advantages of our algorithm compared with similar methods.