Abstract
This paper presents a novel synthetic learning-based approach for solving the component-to-slot assignment problem in robotics using a SCARA robot. The method uses a fully simulated environment that generates and annotates scenes based on rules and visual features. Within this environment, we train a permutation-invariant neural model to predict correct assignments between detected components and predefined target slots. Set Transformer-based encoders are combined with a self-attention MLP scoring head. Assignment prediction is optimized using an improved soft Hungarian loss function. To increase data realism and generalizability, we implement a synthetic dataset generation module on the NVIDIA Omniverse platform. This setup enables precise control over scene composition and component placement. The resulting model achieves high matching accuracy on complex layouts with variable numbers of components and demonstrates strong generalization across multiple configurations. Our results validate the feasibility of learning bijective mappings in simulated assembly scenarios, providing a foundation for scalable real-world robotic pick-and-place tasks. Tests were also conducted on actual robot units.