A Modular Approach to Construction and Control of Walking Robots

In our view walking machines is not a goal in itself. Of course, on one hand there are interesting applications, especially for exploration tasks, but on the other hand, a walking robot serves as demonstrator for nonlinear and adaptive control tasks with a high number of degrees-of-freedom and fast-changing sensor inputs. Here we describe a modular approach, not only for the construction, but also for control aspects: The control technique is called Pose Fitting Networks (PFN) and is able to adapt a basic set of small, recurrent neural networks to a user-defined sequence of robot poses, such that the output nodes of the net can drive the robot’s legs periodically through the sequence of poses.