Learning reaching strategies through reinforcement for a sensor-based manipulator

This paper presents a neural controller that learns goal-oriented obstacle-avoiding reaction strategies for a multilink robot arm. It acquires these strategies on-line from local sensory data. The controller consists of two neural modules: an actor–critic module and a module for differential inverse kinematics (DIV). The input codification for the controller exploits the inherent symmetry of the robot arm kinematics. The actor–critic module generates actions with regard to the Shortest Path Vector (SPV) to the closest goal in the configuration space. However, the computation of the SPV is cumbersome for manipulators with more than two links. The DIV module aims to overcome the SPV calculation. This module provides a goal vector by means of the inversion of a neural network that has been trained previously to approximate the manipulator forward kinematics. Results for a two-link robot arm show that the combination of both modules speeds up the learning process.