Real time learning algorithm for redundant manipulator movement control

We propose a new learning control strategy to solve the ill-posed inverse kinematics of a redundant robot manipulator. Four distinct characteristics are observed: 1) the inverse solution is context-sensitive, which is a requisite when the manipulator starts from an arbitrary joint configuration or moves in a complex environment; 2) learning and execution are both memory-based and can be implemented in real time; 3) the property of conventional pseudoinverse control, i.e. keeping the incremental changes of joint angles minimum, is intrinsic in our scheme; and 4) control is goal-directed in that only the current end-effector position relative to the goal position is needed.