A Deep-Reinforcement-Learning Approach to the Peg-in-Hole Task with Goal Uncertainties

Abstract

The thesis proposed a framework to train deep-reinforcement-learning agents for fine manipulation tasks with goal uncertainties. It consisted of three aspects: state-space formulation, artificial training-goals uncertainties, and progressive training. The framework was used in a simulation for two fine manipulation tasks, square Peg-in-Hole and round Peg-in-Hole. The resulting behaviours were then transferred a physical robotic manipulator and compared to traditional training methods. The deep-reinforcement-learning agents trained using the framework in this work outperformed those trained with definite goals.