Duke University researchers say they have developed a custom processor to perform robotic motion planning that accelerates the process by three orders of magnitude while using 1/20th the power.
Motion planning for a robotic arm involves checking for all potential collisions across the robot's entire range of motion by generating a probabilistic road map (PRM). Precomputation generates a single, massive PRM of 150,000 edges representing all possible robot motions that avoid collisions with stationary objects and other parts of the robot. However, the size of the PRM is limited to the number of circuits that can fit on a field-programmable gate array (FPGA), as each collision-detection circuit corresponds to an edge in the PRM.
To trim the PRM, the researchers simulated 10,000 scenarios and eliminated infrequently used edges from the PRM, leaving about 1,000 edges. No matter how many edges the PRM produces, the FPGA takes 50 nanoseconds per pixel to determine all potential collisions.
In one example, the processor took a little more than 0.6 milliseconds to plan its course, while a software-based planner on a quad-core Intel Xeon processor took nearly three seconds.
From IEEE Spectrum
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA
No entries found