An efficient on-line path planner for outdoor mobile robots

Abstract

Mobile robots operating in outdoor unstructured environments often have only incomplete maps and must deal with new objects found during traversal. Path planning in these environments must be incremental to accommodate new information and must use efficient representations. This article reports recent results in path planning using an efficient data structure (framed-quadtrees) and an optimal algorithm (D*) to incrementally replan optimal paths. We show how the use of framed-quadtrees leads to paths that are shorter and more direct than when other representations are used. We also show the difference in performance when the robot starts with no information about the world versus when it starts with partial information about the world. Our results indicate that, as would be expected, starting with partial information is better than starting with no information. However, in many cases, partial information results in performance that is almost as good as starting out with complete information about the world, while the computational cost incurred is significantly lower. Our system has been tested in simulation as well on an autonomous jeep equipped with local obstacle avoidance capabilities. Results from both simulation and real experimentation are discussed.

Introduction

Path planning for a mobile robot is typically stated as getting from one place to another. The robot must successfully navigate around obstacles, reach its goal, and do so efficiently. Outdoor environments pose special challenges over the structured world that is often found indoors. Not only must a robot avoid colliding with an obstacle such as a rock, it must also avoid falling into a pit or ravine and avoid travel on terrain that would cause it to tip over.

Vast areas have their own associated issues. Such areas typically have large open spaces, where a robot might travel freely, and are sparsely populated with obstacles. However, the range of obstacles that can interfere with the robot’s passage is large — the robot must still avoid a rock as well as go around a mountain. Large areas are unlikely to be mapped at high resolution a priori and hence the robot must explore as it goes, incorporating newly discovered information into its database. Hence, the solution must be incremental by necessity.

Another challenge is dealing with a large amount of information and a complex model of the vehicle. Taken as a single problem, so much information must be processed to determine the next action that it is not possible for the robot to perform at any reasonable rate. We deal with this issue by using a layered approach to navigation. That is, we decompose navigation into two levels — local and global. The job of local planning is to safeguard the robot, avoiding obstacles and reacting to sensory data as quickly as possible [4], [6]. A more deliberative process, operating at a coarser resolution of information, is used to decide how best to direct the robot toward the goal. This approach has been used successfully in the past in several systems at Carnegie Mellon [2], [4], [16].

Approaches to path planning for mobile robots can be broadly classified into two categories: those that use exact representations of the world (e.g. [7], [8]), and those that use a discretized representation (e.g. [1], [7], [9]). The main advantage of discretization is that the computational complexity of path planning can be controlled by adjusting cell size. In contrast, the computational complexity of exact methods is a function of the number of obstacles and/or the number of obstacle facets, which we cannot normally control. Even with discretized worlds, path planning for outdoor environments can be computationally expensive, and on-line performance is typically achieved by use of specialized computing hardware [9]. By comparison the proposed method requires general-purpose computing only. This is made possible by precomputing an optimal path off-line given whatever a priori map is available, and then efficiently modifying the path as new map information becomes available, on-line.

Methods that use uniform grid representations must allocate large amounts of memory for regions that may never be traversed or that may not contain any obstacles. Efficiency in map representation can be obtained by the use of quadtrees, but at a cost of optimality. Kambhampati and Davis [5] introduced hierarchical path-searching using quadtrees, but the method produces suboptimal paths and suffers from the fundamental problem of misrepresentation of terrain data in quadtree-nodes. Zelinsky [19] proposed strategies to smooth paths found using quadtrees, which include the use of a quadtree-based visibility graph, but the resulting paths are not optimal to the resolution of the smallest cell. Both methods require complete replanning whenever new information becomes available.

Recently, a new data structure called a framed-quadtree has been suggested as a means to overcome some of the issues related to the use of quadtrees [3]. We have used this data structure to extend an existing path planner that has in the past used uniform (regular) grid cells to represent terrain. This path planner, D* [14], [15], has been shown to produce optimal paths in changing environments by incorporating knowledge of the environment as it is incrementally discovered. Coupling the two provides a method that is correct, resolution-complete and resolution-optimal (D* search on framed-quadtrees is performed on the highest-resolution cells of framed-quadtrees). It also does this efficiently. Its paths are always shorter, and in all but the most cluttered environments it executes faster and uses less memory than when regular-grids are used [18]. In general, the sparser or the more unknown the world, the greater the advantage of using framed-quadtrees.