A Genetic Cerebellar-Based Controller for Dynamic State Estimation

Abstract : We aimed to produce a general method for controller development that requires minimal domain expertise and development time, yet outperforms traditional control methods. Our testbed system was a high degree-of-freedom 3-D biped robot. In simulation, we successfully demonstrated that surprisingly simple controllers can be evolved to exploit the passive dynamics of the robot and thus permit it to stand and to walk over rugged terrain (-6% leg length perturbations). A 'gray box' modeling approach is then used to integrate data gathered during walking trials on the actual (hardware) biped to adaptively improve the hardware model.