Learning biped locomotion from first principles on a simulated humanoid robot using linear genetic programming

We describe the first instance of an approach for control programming of humanoid robots, based on evolution as the main adaptation mechanism. In an attempt to overcome some of the difficulties with evolution on real hardware, we use a physically realistic simulation of the robot. The essential idea in this concept is to evolve control programs from first principles on a simulated robot, transfer the resulting programs to the real robot and continue to evolve on the robot. The Genetic Programming system is implemented as a Virtual Register Machine, with 12 internal work registers and 12 external registers for I/O operations. The individual representation scheme is a linear genome, and the selection method is a steady state tournament algorithm. Evolution created controller programs that made the simulated robot produce forward locomotion behavior. An application of this system with two phases of evolution could be for robots working in hazardous environments, or in applications with remote presence robots.