Position control of a robotic manipulator using neural network and a simple vision system

• Authors: Bach H. Dinh;Matthew W. Dunnigan;Donald S. Reay
• Venue: AIKED'08 Proceedings of the 7th WSEAS International Conference on Artificial intelligence, knowledge engineering and data bases
• Year: 2008

This paper describes a new approach for approximating the inverse kinematics of a manipulator using a RBFN (Radial Basis Function Network). In fact, there are several traditional methods based on the known geometry of the manipulator to calculate the relationship between the joint variable space and the world coordinate space. However, these traditional methods are impractical if the manipulator geometry cannot be determined, in a robot-vision system for example. Therefore, a neural network with its inherent learning ability can be an effective alternative solution for the inverse kinematics problem. In this paper, a training solution using the strict interpolation method and the LMS (Least Mean Square) algorithm is presented in a two phase training procedure. The strict interpolation method with regularly-spaced-position training patterns in the workspace can produce an appropriate approximation of the inverse kinematic function. However, this solution has the main difficulty in how to collect accurate training patterns whose inputs are selected at pre-defined positions in the workspace. Additionally, the LMS algorithm can improve the approximate function iteratively through on-line training with arbitrary position patterns. A training approach combining two mentioned techniques can be applied for the inverse kinematics problem even with an inaccurate training set. It consists of two steps, firstly producing an inaccurate inverse kinematic approximation by strict interpolation (reflecting the situation that the initial setup and application environments are different) and then re-correcting the network through on-line training by the LMS. The advantages of both training methods can deal with the difficulty of collecting training patterns for practical applications. To verify the performance of the proposed approach, a practical experiment has been performed using a Mitsubishi PA10-6CE manipulator observed by a webcam. All application programmes, such as robot servo control, neural network, and image processing were written in C/C++ and run in a real robotic system. The experimental results prove that the proposed approach is effective.