Sim2Rob—A Sim2Real Framework for Smart Robotics in Assembly

Abstract Robot-based assembly automation struggles with flexibility and adaptability in today's dynamic production landscape. Intelligent, AI-based robotic systems bring great potential for addressing those challenges as well as raising the level of automation. However, the implementation of powerful deep learning (DL) algorithms poses the so-called data problem. Acquiring real, annotated training data sets of sufficient quality and quantity usually requires a significant manual effort. Realistic simulation environments open up the possibility of generating synthetic training data in order to train models for real-world application. Combined with techniques for stable Sim2Real transfer to address the domain gap, this enables rapid adaptation to new products and processes. This work proposes a novel holistic framework for autonomous assembly based on synthetic data generation that leverages the synergy of lightweight robots, AI and advanced simulation capabilities while incorporating domain knowledge.