Learning in simulation for real robots
Simulation is often used in research and industry as a low cost, high efficiency alternative to real model testing. Simulation has also been used to develop and test powerful learning algorithms. However, optimized values in simulation do not translate directly to optimized values in application. In fact, heavy optimization in simulation is likely to exploit the simplifications made in simulation. This observation brings to question the utility of learning in simulation. The UT Austin Villa 3D Simulation Team developed an optimization framework on which a robot agent was trained to maximize the speed of an omni-directional walk. The resulting agent won first place in the 2011 RoboCup 3D Simulation League. This thesis presents the adaptation of this optimization framework to learn parameters in simulation that improved the forward walk speed of the real Aldebaran Nao. By constraining the simulation with tree models learned from the real robot, and manually guiding the optimization based on testing on the real robot, the Nao's walk speed was improved by about 30%.