Combining reinforcement learning and differential inverse kinematics for collision-free motion of multilink manipulators

This paper presents a class of neural controllers that learn goal-oriented obstacle-avoiding strategies for multilink manipulators. They acquire these strategies on-line through reinforcement learning from local sensory data. These controllers are mainly mode of two neural modules: a reinforcement-based action generator and a module for differential inverse kinematics (DIV). The action generator generates actions with regard to a goal vector in the manipulator joint space. Suitable goal vectors are provided by the DIV module. This module is based on the inversion of a neural network that has been previously trained to approximate the manipulator forward kinematics in polar coordinates. Results for two-and three-link planar manipulators are shown. These controllers achieve a good performance quite rapidly and exhibit good generalization capabilities in the face of new environments.