Embodied Evolution of Locomotion in Modular Robots

Control parameters tuned in a simulated environment does not always transfer well on to the real robots, as there exists a reality gap between the simulated environment and the real world. To bridge this gap, candidate controllers are evaluated on a real robot, wherein the controller is a simple sinusoidal oscillator per module. Amplitude, offset, phase-shift of each module's controller, and a common frequency parameter is optimized using GA. Optimization is performed on the real robot, wherein each candidate solution (a set of sinusoidal oscillator parameters) are evaluated on the physical robot. Fitness function is the absolute distance traveled by the robot in a fixed amount of time, which is calculated with the help of an overhead webcam and a color marker placed on the robot.

Highlights Video