Abstract: Pneumatic flexible-bending joint can be used in the design of multi-fingered hand. To discuss its characteristics and control method, the structure and principle of bending joint were introduced and the static and dynamic models were analyzed. A neural Proportional, Integral and Differential(PID) controller was designed for this kind of joint. The algorithm mainly includes two neural networks: Neural Network Identifier (NNI) and Neural Network Controller (NNC). NNI can perform on-line identification of the bending joint and NNC can accomplish real-time adaptation of the coefficient. Therefore, the system has property of self-adaptability. Experimental results show that the neural PID controller can efficiently track the movement of the joint and it is an appropriate controller for this kind of joint.