Learning control for redundant manipulators

An iterative scheme is proposed for learning the input torques that produce a specified repetitive end-effector trajectory for a redundant manipulator, without explicit knowledge of the robot dynamic model. The approach does not rely on a specific inverse kinematic solution. In the learning problem, the number of driving error signals (in the task space) is strictly less than the number of inputs to be determined (in the joint space). During the iterative process, control effort is transferred from a linear feedback law designed for the end-effector error to the learned feedforward term. Joint velocity damping stabilizes the closed loop system. The basic learning algorithm is designed in the frequency domain, while a digital implementation of the necessary signal filtering improves the speed and uniformity of convergence of the learning performance. Simulations are reported for a three-link planar manipulator. The inclusion of a kinematic singularity avoidance scheme is also illustrated