Authors: Chinwe Ekenna, Shawna Thomas, Nancy M. Amato

Venue: N/A
DOI:
Link to Publication

Find PDF here: PDF

Abstract:
Abstract. Sampling-based motion planning has been successful in planning the motion for a wide variety of robot types. An important primitive of these methods involves connecting nodes by selecting candidate neighbors and checking the path between them. Recently, an approach called Adaptive Neighbor Connection (ANC) was proposed to automate neighbor selection using machine learning techniques. It adaptively selects a neighbor connection strategy from a candidate set of options and learns which strategy to use by examining their success rates and costs. In this work, we extend ANC’s reward function by characterizing the types of nodes added (including information about their connectivity) after each connection attempt. In doing so, we gain insight into the nature of these nodes and their ability to improve roadmap quality. We also refine ANC’s cost function by considering the computation time spent during each connection attempt so as to potentially learn faster and more efficient connectors. We show improved performance over selecting a single connection strategy and over ignoring node characterization during learning. We present results in a variety of 2D and 3D environments and are able to solve queries in half the time on average compared to the best single connection strategy and to the original ANC work. 

@misc{Ekenna-ancunc-2014, 
 author = {Chinwe Ekenna and Shawna Thomas and Nancy M. Amato}, 
 title = {Adaptive Neighbor Connection using Node Characterization}, 
 year = {} 
}