A New Sample-based Strategy for Narrow Passage Detection Zahra Sadeghi Robolab, ECE department, College of Engineering, University of Tehran, Tehran, Iran zahra.sadeghi@ut.ac.ir Hadi Moradi Robolab, ECE department, Control and Intelligent Processing Center of Excellence College of Engineering, University of Tehran, Tehran, Iran moradih@ut.ac.ir Abstract— In this paper three different milestone generation methods are proposed to better overcome the narrow passage problem in Probabilistic Roadmaps (PRM). Unlike previous approaches the points generated on obstacles are preserved and their information is used to find the location of narrow passages. We have tested the proposed algorithms on several standard environments and compared its results with the results of two other well-known sample generation methods. The results show better performance in terms of runtime, uniform coverage of the configuration space, and success rate in narrow passage detection and final path generation. Index Terms - probabilistic motion planning; Narrow passage; I. INTRODUCTION Motion planning is the problem of finding collision free motions in order to move the robot from an initial position to a goal position. This problem, for high DOF robots, is often solved in the configuration space where the location of a robot is represented with robot’s joint variables. The set of all robots’ configuration which have collision with obstacles is called configuration space obstacle, and the complementary of configuration space which doesn’t have any intersection with obstacles is known as free space [1]. In recent years, one of the most popular methods for solving motion planning was known as Probabilistic Roadmaps (PRM) which is a sample based technique. PRM consists of a learning phase and a query phase. In the learning phase a large number of collision free configurations, which are called milestones, are created. Then a graph is constructed by connecting the neighboring samples with a collision-free path, using the local planner. In the query phase a path is tried to be found between the starting point and the goal point through the attained roadmap [2]. A typical and important issue in PRM is handling the narrow passages, i.e. a narrow area of the configuration space between two obstacles that are hard to generate random samples in between and it is harder to connect them to the rest of the PRM. Consequently different strategies are used for milestone generation in the sampling stage to overcome the narrow passage problem. In [3] Gaussian strategy is introduced with the intention to add samples near the obstacles. Thus a few samples would be created in narrow passage areas. In this approach, if one of the two consecutive random samples created with Gaussian distribution in a short distance, lies on obstacles and the other one lies on free space, the test is successful and the node in the free space is added. The 2 nd approach is the Bridge test in which two consecutive samples are generated at random with Gaussian distribution in a short distance from each other. If the two samples lie on obstacle space and the middle point between them is in free space, it is added as a new milestone [4]. The 3 rd approach, i.e. the hybrid approach, is the combination of Gaussian and bridge test and is explained in [5]. Other approaches are based on shrinking robots or dilation of free space [6], [7]. In the previous sample based approaches only the samples lied on the free space were preserved and the other samples lied in the obstacle space were abandoned Our strategy is based on the observation that in many environments, the obstacle space is much more expanded than free space, especially in cluttered environments, hence, most of the generated samples lie on obstacles. Thereupon, not only a considerable amount of time is needed to create more samples, but also the cost of the generation of all samples is not compensated. With regards to the mentioned problems it seems logical to try to take advantage of the samples generated on the obstacles. In strategies like the bridge test and Gaussian strategies the main attempt is to use the samples created on the obstacles to some extent. In the proposed strategy it is tried to highly employ the information associated to the samples, i.e. the detected parts of the obstacles. It is important to mention that the milestone generation and collision detection is, typically, 10 to 100 times faster than local planning between two milestones. Consequently, the majority of the sampling methods try to generate better milestones to reduce the cost of local planning, even if the number of generated milestones is bigger. We also rely on this approach to generate better milestones. It is noteworthy to mention that we recently found another obstacle based strategy proposed in [8]. While their strategy is similar to ours in using the obstacle space rather than the free space, we have taken different methods to solve the problem of narrow passages. II. THE PROPOSED METHODS We have proposed three different methods for finding narrow passages based on using the milestones lied on Proceedings of the 8th World Congress on Intelligent Control and Automation June 21-25 2011, Taipei, Taiwan 978-1-61284-700-9/11/$26.00 ©2011 IEEE 1059 Downloaded from http://www.elearnica.ir