Stable on-line neural control of systems with closed kinematic chains

Abstract

Artificial neural networks have been used extensively in control research. In industrial systems, however, it is crucial to adopt neural control structures which have a guaranteed proof of stability, especially if control system failure were to endanger life (e.g. in fast moving manipulators or transportation). In the paper, the neural control of robotic systems with closed kinematic chains is discussed and theorems guaranteeing the control stability of such systems are developed. The first class of systems have a single serial chain with a prescribed contact force when moving across a surface, i.e. the problem of hybrid position/force neural control. The second class of systems considered includes hexapod walking machines, which have a varying topology of closed kinematic chains during walking. The equations of motion can be solved by optimising contact forces according to a predefined cost function, and so the hybrid/position neural controller is extended to this class. A novel control structure which makes no initial assumptions about the system is also presented, using the concept of `virtual neural networks': a projection of the neural controllers into the underconstrained space of the generalised co-ordinates of the equations of motion. This approach can be applied to a large number of different systems, including parallel manipulators and Stewart platforms, and it is also extended to include neural networks implemented on digital microprocessors.