Reinforcement Learning Approaches for Collaborative Robot Control in Manipulation Tasks

Abstract

With the exponential growth in technological advancement and the increasing reliance on electrical and electronic equipment, the efficient treatment of end-of-life products has become essential for mitigating environmental impact. Remanufacturing presents an environmentally and economically advantageous approach to address these impacts. However, while automation has seen success in manufacturing, manual labour remains preferred in remanufacturing, particularly in disassembly, due to operational uncertainties. In this regard, reinforcement learning offers an alternative for decision-making and control in dynamic systems, yet the efficiency and generalisability of learning disassembly tasks remain unclear. This industrial doctoral thesis investigates the application of reinforcement learning techniques in the specific context of disassembling magnetic gaskets from refrigerator doors in a human-robot working environment, focusing on three core pillars of reinforcement learning at present: performance, sample efficiency, and generalisation. Building on these research areas, the thesis initially proposes a proof-of-concept balancing safety and workflow efficiency in a randomised human-robot disassembly environment. The study is then expanded, with the control policy being learned through an interactive reinforcement learning framework where the human role is replaced by an automated supervisor featuring constraint-based modelling techniques to enhance sample efficiency. The results for both studies are presented in simulation and real-world settings.