UniOMA: Unified Optimal-Transport Multi-Modal Structural Alignment for Robot Perception

This paper proposes UniOMA, a multimodal alignment method that augments contrastive losses with geometry-aware Gromov-Wasserstein distance regularization to ensure structural consistency across modalities. The approach is applied to robot perception tasks involving force and tactile signals with varying sample densities. Experiments on five robotic benchmarks demonstrate improved structural alignment and downstream performance compared to state-of-the-art multimodal contrastive baselines, with seamless integration into InfoNCE-like objectives.