Open Challenge Abstract

lation league competition, due in large part to a fast and robust humanoid walk on an approximate model of the Nao robot in the ODE-based RoboCup 3D simulator [3]. Using the CMA-ES evolutionary learning algorithm [2] for param-eter optimization and using the Condor system to perform large-scale distributed computing, the team was able to run thousands of simulated trials to find parameter values that transformed a relatively slow walk into the most effective simulated walk for the Nao in the simulation environment. The walk engine and the intial (seed) walk were ported from the real Nao robots based on the work of Graf et al. [1], mak-ing this the first time, to our knowledge, that control code on a real robot was used to improved control in a simulator, rather than vice versa. In simulation there is little cost to running and testin