Collaborative Research: Unified Sensing and Control for Generalizable Real-Time Arc Welding

Advanced manufacturing increasingly relies on robotic systems to improve productivity, quality, and workplace safety in critical industries such as shipbuilding, energy infrastructure, transportation, and heavy equipment manufacturing. Arc welding is among the most widely used manufacturing processes, yet producing consistent, high-quality welds remains challenging because skilled operators must continuously adapt to variations in material properties, joint geometry, and thermal conditions. Conventional robotic welding systems have limited adaptability and often require extensive process-specific tuning, restricting their deployment in complex manufacturing environments. This award supports fundamental research to develop intelligent robotic welding systems that can autonomously perceive and adapt to changing welding conditions while maintaining consistent weld quality. By integrating advances in artificial intelligence, control theory, and welding science, the project will establish a unified sensing and control framework that enables robots to respond to molten weld-pool behavior in a manner similar to experienced human welders. The resulting technologies are expected to improve manufacturing productivity, reduce worker exposure to hazardous environments, and strengthen U.S. leadership in advanced manufacturing. The project will also provide interdisciplinary training for graduate and undergraduate students, integrate artificial intelligence and robotics into engineering curricula, and expand outreach activities that engage K–12 students in manufacturing and STEM education. Current robotic welding systems are typically designed for specific processes and operating conditions, making them difficult to adapt to new materials, joint configurations, and manufacturing environments without extensive retuning. To overcome these limitations, this research will develop a unified learning and control architecture that combines physics-informed state estimation, data-driven dynamics modeling, and adaptive control for a broad class of arc-welding processes. Dynamic variational autoencoders and deep Bayesian filtering will be employed to infer robust low-dimensional representations of weld-pool behavior from high-speed sensing data, while neural ordinary differential equations will model the continuous-time evolution of these latent states under varying process conditions. A hybrid control strategy will combine generative adversarial imitation learning to capture expert human welding behavior with iterative linear quadratic regulation to improve real-time control accuracy, stability, and disturbance rejection. Experimental validation on robotic welding platforms, together with virtual-reality-based human-in-the-loop studies, will evaluate the framework across multiple welding processes, materials, and joint geometries. The resulting research will advance fundamental knowledge in intelligent sensing, machine learning, and adaptive control for manufacturing and establish a generalizable foundation for autonomous robotic systems capable of operating reliably in complex industrial environments. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.