| Abstract Scope |
Robotization of novel manufacturing is key to unlocking its full advantages for wide industrial applications. However, the inherently complex and not yet fully understood nature of such processes impedes the design of effective automation rules. Double-electrode gas metal arc welding (DE-GMAW) improves conventional gas metal arc welding (GMAW) by introducing a bypass electrode for flexible heat input control, yet its robotization requires adaptive and precise control of electrode positions, which remains a challenge. Humans possess unparalleled adaptability and learning capabilities for tackling unseen processes, but physical limitations and safety concerns restrict the full utilization of their cognitive skills. In this work, we address these challenges by demonstrating the robotization of DE-GMAW. We collect high-quality human demonstrations through human-robot collaboration and derive human intelligence from these demonstrations for deployment. First, a human-robot collaboration (HRC) system was established, featuring a virtual reality human interface, deep learning-enhanced visual feedback, and image-based electrode positioning. These features address the challenges in direct manual operation, enabling high-quality human demonstrations. Next, we proposed a combined method, using a variational autoencoder (VAE) and generative adversarial network (GAN), for representation learning and image generation to model the metal transfer phenomenon observed by humans. This approach reduces the dimensionality of complex human observations, facilitating the efficient learning of human intelligence from limited data. Subsequently, we learned the expert policy from human observation-action trajectories, employing a generative adversarial imitation learning (GAIL) algorithm. Finally, the expert policy was deployed on the human-robot collaboration system, achieving comparable performance to the human operator. We believe these methods are generalizable to other manufacturing processes, effectively addressing process complexity, human physical limitations, and data constraints in learning from humans. |