Performance benchmarking of multimodal data-driven approaches in industrial settings

Abstract

Data-driven solutions are increasingly transforming the industrial sector, yet collecting large-scale, multimodal datasets remains costly and challenging. This paper presents three synthetic multimodal datasets that replicate real-world industrial conditions across varying levels of complexity, designed to benchmark multimodal machine learning models. We validate their utility through a series of experiments. Cross-modal prediction and domain adaptation demonstrate that the datasets effectively capture strong multimodal correlations. Multimodal reconstruction experiments confirm the internal consistency and richness of the fused representations, indicating that the modalities complement each other in capturing underlying structure. Additionally, multimodal regression significantly outperforms unimodal baselines, underscoring the predictive strength gained through multimodal integration. Together, these results demonstrate the utility of our datasets, establishing a solid baseline for future research and encouraging further advancements in industrial data-driven solutions. • Developed 3 simulation environments for multimodal industrial robotics data. • Proposed ML tasks to benchmark performance using our multimodal datasets. • Evaluated cross-modal tasks to show dataset versatility and usefulness. • Evaluated datasets on multimodal learning tasks to benchmark models.