Approximation of the inverse kinematics of a robotic manipulator using a neural network
Abstract
A fundamental property of a robotic manipulator system is that it is capable of accurately
following complex position trajectories in three-dimensional space. An essential component
of the robotic control system is the solution of the inverse kinematics problem which allows
determination of the joint angle trajectories from the desired trajectory in the Cartesian
space. There are several traditional methods based on the known geometry of robotic
manipulators to solve the inverse kinematics problem. These methods can become
impractical in a robot-vision control system where the environmental parameters can alter.
Artificial neural networks with their inherent learning ability can approximate the inverse
kinematics function and do not require any knowledge of the manipulator geometry.
This thesis concentrates on developing a practical solution using a radial basis function
network to approximate the inverse kinematics of a robot manipulator. This approach is
distinct from existing approaches as the centres of the hidden-layer units are regularly
distributed in the workspace, constrained training data is used and the training phase is
performed using either the strict interpolation or the least mean square algorithms. An
online retraining approach is also proposed to modify the network function approximation
to cope with the situation where the initial training and application environments are
different. Simulation results for two and three-link manipulators verify the approach.
A novel real-time visual measurement system, based on a video camera and image
processing software, has been developed to measure the position of the robotic manipulator
in the three-dimensional workspace. Practical experiments have been performed with a
Mitsubishi PA10-6CE manipulator and this visual measurement system. The performance
of the radial basis function network is analysed for the manipulator operating in two and
three-dimensional space and the practical results are compared to the simulation results.
Advantages and disadvantages of the proposed approach are discussed.