Go to content Go to navigation Go to search

Reinforcement Learning In Control Systems For Walking Hexapod Robots

Ivo Hrdlička, Patrik Kutílek

Abstract


We have developed a controller for a six-legged robot that allows it to walk on rough terrain or where mines have been detected. Intention of this article is to describe application of reinforcement learning methods for implementation of an acyclic walking. Cyclic methods of a walking are not feasible in difficult terrain because doesn’t consider subsequent undercarriage states of the robot in future. For basic (not so rough) terrain we used new designed heuristic algorithm. In complex environment, computational more intensive approach (i.e. Q-learning) is applied. Overall control algorithm of the walking undercarriage combines these two presented techniques. Results of the new algorithm of control to the problem of learning to walk with a six legged robot are presented by demining robot.
Full paper

2005-09-10

commenting closed for this article

The Locomotion Control of the Concyclical Walking Robot GFORCE – Tool for Flight Simulations