RESAMPL: A Region-Sensitive Adaptive Motion Planner [chapter]

Samuel Rodriguez, Shawna Thomas, Roger Pearce, Nancy M. Amato
Springer Tracts in Advanced Robotics  
Automatic motion planning has applications ranging from traditional robotics to computer-aided design to computational biology and chemistry. While randomized planners, such as probabilistic roadmap methods (prms) or rapidly-exploring random trees (rrt), have been highly successful in solving many high degree of freedom problems, there are still many scenarios in which we need better methods, e.g., problems involving narrow passages or which contain multiple regions that are best suited to
more » ... rent planners. In this work, we present resampl, a motion planning strategy that uses local region information to make intelligent decisions about how and where to sample, which samples to connect together, and to find paths through the environment. Briefly, resampl classifies regions based on the entropy of the samples in it, and then uses these classifications to further refine the sampling. Regions are placed in a region graph that encodes relationships between regions, e.g., edges correspond to overlapping regions. The strategy for connecting samples is guided by the region graph, and can be exploited in both multi-query and single-query scenarios. Our experimental results comparing resampl to previous multi-query and single-query methods show that resampl is generally significantly faster and also usually requires fewer samples to solve the problem. The general motion planning problem consists of finding a valid path for an object from a start configuration to a goal configuration. Traditionally, a valid path is any path that is collision-free, e.g., avoiding collision with obstacles in the environment and avoiding self-collision. Motion planning has applications in robotics, games/virtual reality, computer-aided design (CAD), virtual prototyping, and bioinformatics. While an exact motion planning algorithm exists, its complexity grows exponentially in the complexity of the robot [17] . Instead, research has turned towards randomized algorithms. One widely used and quite successful randomized algorithm is the Probabilistic Roadmap Method (prm) [11] . prms operate in configuration space (C-space), where each point in C-space corresponds to a specific robot configuration/placement. While not guaranteed to find a solution, prms are probabilistically complete, i.e., the probability of finding a solution given one exists approaches 1 as the number of samples in the roadmap approaches ∞. Issues: The motion planning problem is significantly more challenging when there are difficult or narrow areas in C-space that must be explored. While there have been many attempts to generate samples in difficult or interesting areas of C-space [1, 4, 5, 7, 20] , they are typically applied over the entire C-space and do not allow for the identification and refinement of particular areas of C-space. Motion planning problems typically come in one of two types: multi-query path planning and single-query path planning. The goal of a multi-query planner is to efficiently model the entire free C-space so as to answer any query in that space. A single-query planner, however, is only concerned about the portion of free C-space needed for the query, so it is generally faster than a multi-query planner. Most randomized motion planners are well-suited to one of these problem types, but not to both. Our Contribution: In this work, we propose resampl, a motion planning strategy that uses local region information to make intelligent decisions about how and where to sample, which samples to connect together, and to find paths through the environment. Based on an initial set of samples, we classify regions of C-space according to the entropy of their samples. We then use these classifications to further refine the sampling. For example, we increase sampling in "narrow" regions and decrease sampling in "free" regions. Regions are placed in a region graph that encodes relationships between regions, e.g., edges correspond to overlapping regions. We use the region graph to determine appropriate connection strategies for multi-query planning and to extract a sequence of regions on which to focus sampling and connection for single-query planning. Our experimental results comparing resampl to previous multi-query and single-query methods show that it is generally significantly faster and also usually requires fewer samples to solve the problem. Hence, resampl's region-based approach to motion planning addresses both issues mentioned above. • Regions: Considering local information when deciding where and how to refine sampling and connection enables us to focus on difficult areas instead of continuously searching in the entire space as is done by most previous methods. • Region Graph: The relationships between regions can be exploited during connection in both multi-query and single-query situations. Related Work There has been extensive work on randomized motion planners for both multi-query and single-query problems. In this section we give an overview of some of the methods that have been proposed. Multi-Query Planning. One widely used and quite successful multi-query randomized planner is the Probabilistic Roadmap Method (prm) [11] . prms consist of two phases, a preprocessing/roadmap construction phase and a query phase. During roadmap construction, robot configurations are first randomly sampled from C-space. Samples are kept if they are in the feasible region of C-space (C-free). Connections are then attempted using a simple local planner between neighboring configurations. Valid connections are stored as edges in the roadmap. Although prms have been successful in solving previously unsolvable problems, they have difficulty when the solution path must pass through a narrow passage in the C-space. Attempts have been made to generate configurations in interesting areas of C-space that are difficult to discover using uniform random sampling. For example, [1, 4, 9, 16] attempt to generate samples near the surface of C-space obstacles. In [20] , samples are
doi:10.1007/978-3-540-68405-3_18 fatcat:ltty3ywjmzaitkm7pwqfrhgcju