Automatically Tuning Control Systems for Simulated Legged Robots

Robert Ringrose

Rather than create a control system from scratch each time we build a new robot creature, we would like to generate control systems automatically. I have implemented an algorithm which, given a control system that works well for one creature, automatically tunes it to work for a new, similar creature. Using this approach, the control system for a horse might be adjusted for use with elephants, giraffes, and dogs. The adjustment is accomplished by gradually altering the original creature to make it like the new one and repeatedly tuning the control system as these changes are made. Because the creature’s alteration is gradual, the control system can be tuned using a local search such as gradient descent. In simulation tests, the tuning algorithm has successfully tuned the control system of a planar quadruped simulation to accommodate a reduction in leg length by a factor of two, an increase in body mass by a factor of three, and changes in the commanded speed while trotting.

This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.