Hybrid memory-based control of robotic manipulators
Choon-Young Lee, Ju-Jang Lee · 2002
Robotics research aims to realize some aspects of human control skills in a mechanical system. We considered another approach for the control of robotic manipulators using a multi-valued function regularization network (MVRN) approximating a multiple inverse kinematics solution. We assume that we have only the input-output data pairs of joint and Cartesian space for the unknown forward kinematics relation. Using these data, we approximate global inverse kinematics mapping using the MVRN. After approximating the inverse kinematics mapping, we find collision-free joint trajectories like a human being for the given task and the environment. We also adopt an adaptive neural network control scheme for motion control with unknown dynamics. From the global viewpoint, the MVRN may be considered as the long-term memory for relatively unchanging information for the robot manipulator and an adaptive neural network controller can be thought of as the short-term memory for the time-varying information for the change of load and dynamic parameters of the plant. Using these two components, we construct a more general and global control scheme for robotic manipulators. Simulation results are presented to illustrate the overall scheme of the proposed method.