Collision-Free Welding Torch Pose Determination Through Imitation Learning with Expert Demonstrations

In collaboration with our industrial partners at Autocam [1], we developed this Deep Learning (DL) based assistance system for robotic welding. The system is tailored to address the challenges encountered in real-world industrial settings, where conventional programming methods face limitations, particularly in constrained spaces. Autocam provided us with authentic industrial datasets and 3D models. Our approach focuses on automating the identification of collision-free torch poses, crucial for navigating intricate environments commonly encountered in industrial robotic operations. To capture the intricacies of these environments, we utilize four virtual cameras, each strategically positioned to provide a unique perspective. This multi-perspective observation facilitates the development of a robust relationship between 3D space and collision dynamics. The DL model is trained using a combination of Imitation Learning and Deep Reinforcement Learning (DRL). This training process not only empowers the agent to autonomously learn collision-free poses but also allows for the assimilation of existing solutions derived from real-world industrial scenarios. By leveraging computer vision and combining the strengths of both imitation learning and DRL, our model aspires to offer a more adaptive and efficient solution for automation in challenging spatial configurations within industrial robotic operations.