I. Introduction
With the development of robots and control technology, robots are applied extensively in industrial and agricultural operation. Robots are a complicated nonlinear multivariable with and strong coupling. Because of the complexity of operation condition, and random distribution of operated object, it is also time-varying system. By the general model-based control method, robots are not controlled accurately, so their tracks are not kept better. Sliding mode control (SMC) has complete adaptability for system disturbance and stirring, which is extensively applied in robots[1]. Fuzzy control does not need precise mathematical model and can decouple joints, but fuzzy control system is easily influenced by non-linear, time-varying and random disturbance [2]. Neural network control has many advantages, such as self-learning, selforganizing, self-adaptive capacity, fault-tolerance, nonlinear and parallel distributed processing, noise treatment, inadequate data treatment, and so on. However, it also has the congenital defects, such as it falls into local minimum easily, and it is weakly normalized for few samples[1]. These defects make it difficult to meet the requirements of precise control for robot.