Robotic disassembly of screws for end-of-life product remanufacturing enabled by deep reinforcement learning

Robot-assisted screw removal can greatly facilitate the disassembly and remanufacturing automation of end-oflife products to realise circular economies. However, it is challenging to determine the exact positions of disassembling screws in practical remanufacturing environments. To tackle the issue, in this research, a novel approach designed based on reinforcement deep learning (DRL)-based optimisation processes is herein presented. First, to identify the search directions for the exact positions of disassembling screws, an analytical model was established to quantify geometrical relationships between a robotic screwdriver and the screws for removal. Secondly, a Markov model was built to represent key parameters related to robot configuration and the evolving relationships in the analytical model. Furthermore, a proximal policy optimisation (PPO) algorithm, which is a high-performing DRL algorithm, was developed to determine the optimal values of key parameters in the analytical model and the Markov model. Finally, experiments were conducted to identify optimal parameters by applying this approach and to benchmark its effectiveness. Experimental outcomes demonstrated that high success rates in screw positioning were achieved using this approach, and it outperformed the comparative approaches/optimisation algorithms in terms of screw positioning accuracy and efficiency.