Evolutionary optimization of a bipedal gait in a physical robot

Evolutionary optimization of a gait for a bipedal robot has been studied, combining structural and parametric modifications of the system responsible for generating the gait. The experiment was conducted using a small 17 DOF humanoid robot, whose actuators consist of 17 servo motors. In the approach presented here, individuals representing a gait consisted of a sequence of set angles (referred to as states) for the servo motors, as well as ramping times for the transition between states. A hand-coded gait was used as starting point for the optimization procedure: A population of 30 individuals was formed, using the hand-coded gait as a seed. An evolutionary procedure was executed for 30 generations, evaluating individuals on the physical robot. New individuals were generated using mutation only. There were two different mutation operators, namely (1) parametric mutations modifying the actual values of a given state, and (2) structural mutations inserting a new state between two consecutive states in an individual. The best evolved individual showed an improvement in walking speed of approximately 65%.