Contact-Based Digital Twins Modeling for Reinforcement Learning of Robotic Disassembly Operations

Reinforcement learning (RL) holds great potential for robotic skill acquisition, but its practical deployment in industrial disassembly tasks is challenged by low sample efficiency and safety concerns in contact-intensive environments. This article presents a cyber-physical approach that enhances RL through simulation-to-reality (sim-to-real) skill transfer using a Digital Twin (DT). The DT models the physical environment and is calibrated via the Bees Algorithm, a metaheuristic optimisation method, to reduce the reality gap by minimising discrepancies between simulated and real-world responses. That enables more accurate simulation of contact dynamics without requiring manual parameter tuning or expert modelling. The method is validated on a representative task: removing a bolt from a door-chain groove, simulating the challenges of force-sensitive disassembly operations. Results demonstrate that the DT-assisted sim-to-real transfer improves learning efficiency, offering a scalable framework for deploying RL in cyber-physical systems for intelligent disassembly and circular manufacturing.