| This work aims to explore the power of learning and speed of one type of the artificial neural networks, namely the Cerebellar Model Articulation Controller (CMAC). The capabilities and limitations of CMAC are investigated and compared with other types of neural networks. Then, a CMAC-based learning algorithm is applied to the kinematic control of a walking machine. The stability of the CMAC-based control scheme is also studied and a mathematical proof of the asymptotical stability for the regulation control is derived. The developed algorithm is then extended to control both the position and force of a two degrees-of-freedom leg walking on soft terrain. It is demonstrated that the CMAC-based learning system performs better than the mere feedback control in terms of speed and accuracy. Furthermore, the CMAC-based hybrid force/position control is applied to control a quadruped walking machine walking on a flat and soft terrain. Finally, the proposed walking control are simulated using a self-developed animation software package. The entire learning control process was accomplished on a PC/AT personal computer with a CMAC board. |
