Reinforcement Learning-Based Adaptive Control of Screwing Process Under Material Uncertainties

The quality of the final product in many manufacturing companies is determined by the accurate and precise control of assembly operations. Such operations are often automated to achieve higher throughput, consistent quality, and lower labor costs. The screwing process is one of the most commonly used assembly operations which in spite of its simple appearance is challenging to automate. The process of tightening a screw into a mating part depends on several parameters including material properties as well as geometric characteristics. In-advance knowledge of all involved parameters is hard to achieve, and potential uncertainties such as variations in material properties as well as tolerances in geometric characteristics make the model-based control of the process impractical. This has limited the applied control strategy in most of the industrial screwing machines to follow a straightforward passive scheme. This paper proposes an adaptive framework for controlling the screwing process at the presence of material uncertainties using reinforcement learning (RL). To address uncertainties, we suggest a data-driven approach augmenting the simulation environment of the RL framework with real-world process data. Screwing in wood is studied as a testbed since higher orders of material uncertainties are expected while screwing into an organic material. The proposed framework shows that the trained RL agent can adapt its behavior and accomplish screwing operations even at the presence of material uncertainties.