Robotizing welding for complex welding processes could be challenging as the design of an effective control algorithm that adjusts the welding parameters in real-time per process dynamic change traditionally requires a dynamic model for the complex process. In addition, traditional control theory requires the process feedback to be parameterized into variables while such parameterization for complex welding process is not straightforward. In solving these challenges, we propose to learn from human welders who naturally take un-parameterized process feedback as the input of their “control algorithms”, also without the need for an explicit process model, to adjust the welding parameters. They are quick learners who may quickly tune their “control algorithms” from a few trials. As such, learning from human welder provides an effective way to robotize a complex welding process. In this work, we designed an experimental system to allow a human welder to adjust their gas tungsten arc welding (GTAW) torch to follow a gas metal arc welding (GMAW) to maintain a path from the current to flow from the wire to the tungsten. This is the double-electrode GMAW that can reduce the heat input while maintaining the depositing rate . The key to operating this process is to dynamically place the tungsten in the right position of the gas metal arc. It is hypothesized that the image of the double-electrode arc, consisting of the main arc (arc between the wire and work-piece) and the bypass arc (arc from the wire to the tungsten), and its dynamic change (of the image) can provide the full process feedback to effectively adjust the welding parameters to maintain the stability of the arcing process for the DE-GMAW to operate. Experiments have been done for a human welder to operate the GTAW torch per the observation of the arcing image. The image is recorded together with the GTAW torch position in relation to the main arc as the input and output of the human welder. A deep learning model has been trained to predict human operation from the updated image sequence. The trained model will be used by a robot carrying the GTAW torch to adjust the GMAW torch's position and orientation in relation to the GMAW torch.
 Li, K.H., Chen, J.S. and Zhang, Y., 2007. Double-electrode GMAW process and control. WELDING JOURNAL-NEW YORK-, 86(8), p.231s.