Skip to Main Content
The representation of probabilistic graphical model often encodes a network whose size is unboundedly large. Such networks pose particular challenges to inference algorithms, specifically making the task of robot path queries highly inefficient due to poor locality of memory references. Whereas a more predictable, resolution complete method yields a highly compact graph structure that captures much of the signal in distributing the configuration free space. In this paper we demonstrate an efficient data parallel algorithm for mapping the computationally intensive, Reachability Roadmap method on the GPU. For our implementation on the recently introduced NVIDIA's Fermi architecture, we show roadmap construction time under twenty seconds for a closure resolution of 55×55×55 cells. Moving forward, our system is well positioned to address smooth navigation of robots in a dynamically changing 3D virtual environment.
Date of Conference: 18-22 Oct. 2010