Adaptive torque control using a connectionist reinforcement learning agent.

The correction of angular misalignment between mating components is a fundamental requirement for their successful assembly. In this paper we present how a learning agent based on the Stochastic Gradient Ascent (SGA) algorithm can be used to command appropriate rotations of a robotic manipulator's hand in order to minimise such misalignments and allow the completion of a peg in hole task. The learning capabilities of the proposed methodology are proven both with simulation and real time implementation using a PUMA 6 DOF industrial manipulator.