Learning Robot Locomotion

With the growth of computational power in recent years, complex autonomous robot systems have become a possibility. To increase their flexibility and adaptability, learning mechanisms are necessary for robot control. In this paper, I present two recent legged robot-systems that use machine-learning in their foothold-selection process. It was found that currently both systems have a number of issues, including high failure rates in autonomous navigation and the necessity of exact maps and external positioning information. The combination of both systems may lead to an overall improvement in performance.