The automated handling of rigid and especially flexible items poses significant challenges for industrial vision systems. Variations in shape, material properties, and deformation need to be covered, while the detection quality must be maintained over time when retraining. AI-based perception solutions, trained with synthetic data generated in photorealistic simulation environments, enable the robust detection and handling of such items. In industrial automation and logistics, the introduction of AI-models specialized by object category assure quality control of the training data. Practical examples illustrate this approach.