High Precision Peg-in-Hole Assembly Approach Based on Sensitive Robotics and Deep Recurrent Q-Learning

Abstract

Sensitive robot systems are used in various assembly and manufacturing technologies. Assembly is a vital activity that requires high-precision robotic manipulation. One of the challenges faced in high precision assembly tasks is when the task precision exceeds the robot’s precision. In this research, Deep Q-Learning (DQN) is used to perform a very tight clearance Peg-in-Hole assembly task. Moreover, recurrence is introduced into the system via a Long-Short Term Memory (LSTM) layer to tackle DQN drawbacks. The LSTM layer has the ability to encode prior decisions, allowing the agent to make more informed decisions. The robot’s sensors are used to represent the state. Despite the tight hole clearance, this method was able to successfully achieve the task at hand, which has been validated by a 7-DOF Kuka LBR iiwa sensitive robot. This paper will focus on the search phase. Furthermore, our approach has the advantage of working in environments that vary from the learned environment.