Reinforcement learning for collaborative robotic contact-rich disassembly tasks

Abstract

With the exponential growth of the world’s population and the resulting increase in consumption rates, the efficient treatment of end-of-life (EOL) products has become critical to mitigating environmental impacts. Remanufacturing offers an environmentally and economically beneficial approach to counteracting these impacts. While automation has been successful in assembly and manufacturing, manual labor is preferred in remanufacturing, especially disassembly, to cope with operational uncertainties. Reinforcement learning (RL) is presented as an alternative for decisión making and control in changing systems, but the extent to which disassembly tasks can be automatically learned and generalized is unknown. This doctoral dissertation, in Applied Engineering, explores the application of RL techniques for collaborative robot control to generalize disassembly tasks with uncertainties due to the variability of the geometric and physical properties of the manipulated objects. With this, a modular RL library that enables simultaneous training of agents in massively parallel environments is presented to reduce training time while consuming the same amount of resources and increasing the perceived reward. In addition, a control framework for KUKA LBR iiwa cobots that outperforms existing solutions and allows the use of different types of force overlays to reduce contact forces caused by friction and the probability of jamming states when performing disassembly tasks is presented. Furthermore, a collection of ready-touse packages for rapid prototyping and reducing the development and deployment time of collaborative robotic systems for assembly and disassembly is proposed.