Application of evolved locomotion controllers to a hexapod robot

This paper shows that genetic algorithms used to evolve dynamical neural networks for controlling the locomotion of a simulated hexapod agent are capable of directing the walking of a real six-legged robot, and that many of the desirable properties observed in simulation carry over directly to the real world.

Abstract

In previous work, we demonstrated that genetic algorithms could be used to evolve dynamical neural networks for controlling the locomotion of a simulated hexapod agent. We also demonstrated that these evolved controllers were robust to loss of sensory feedback and other peripheral variations. In this paper, we show that these locomotion controllers, evolved in simulation, are capable of directing the walking of a real six-legged robot, and that many of the desirable properties observed in simulation carry over directly to the real world. In addition, we demonstrate that these controllers are amenable to hardware implementation and can thus be easily embodied within the robot.