Development of a toolchain for robust Sim2Real transfer of learning-based robotics applications

The increasing complexity of industrial production demands innovative automation solutions. This paper presents a toolchain developed for the robust Sim2Real transfer of reinforcement learning-based control policies in robotic systems, addressing challenges in the inefficient manual programming of task-specific control. Our toolchain facilitates the training within a simulated environment in NVIDIA Isaac or MuJoCo and the subsequent implementation on real-world robotic hardware using ROS 2 for seamless interaction with low-level robot controllers. Several methods to ensure robust performance on real hardware are an integral part of our approach. The first real-world experiments of a dual-arm robotic assembly task demonstrate a 100% success rate, with process times closely matching those in simulations, verifying the model's robustness to uncertainties in system dynamics and sensor noise.