Region Identification Methods for Efficient and Automated Motion Planning

Realistic environments for motion planning problems typically consist of multiple areas, e.g., in a house there are rooms, hallways, and doorways. Decomposition of the environment into smaller areas, or regions, has been shown to impact the quality of planning for adaptive methods. Also other complex problems, such as planning for groups of agents with environmental preferences, can be aided with automated region identification. However, despite the advantages of regions, automatic region identification is not always easy, inexpensive, or consistent. In this paper we explore and compare three methods of automatic region identification and quantify the effects of the resulting regions on planning. The three strategies used for region identification are the statistical methods: k-means clustering, PG-means clustering, and Hierarchical clustering. First, we explore the differences in the regions produced by these three methods in a variety of rigid body and articulated linkage environments. Then, we quantitatively compare the usefulness of the region methods in automated motion planning through the application of the regions within the Unsupervised Adaptive Strategy for Motion Planning (UAS). Our results indicate both regions that seem intuitive and non-intuitive to the problems solve motion planning problems with increased efficiency and automation. That is, we found that planning generally benefitted from using regions, and the benefit was not tied to the naturalness of the region.