TacFG-9: A Tactile-Vision-Language Dataset for Factorized Fine-Grained Tactile Semantics

Tactile perception provides essential physical cues for understanding object properties during robotic interaction. However, existing tactile-language datasets primarily rely on coarse labels and lack systematic annotations for fine-grained tactile semantics, which makes it difficult to generate semantically faithful descriptions under partial tactile observations. To address this issue, we introduce TacFG-9, a fine-grained tactile-vision-language dataset constructed through a three-level, nine-dimensional annotation framework. TacFG-9 integrates tactile observations, visual context, verified captions, and structured semantic labels into a unified multimodal semantic representation, enabling tactile semantics to be organized from coarse material identity to fine-grained local perceptual cues. To scale TacFG-9 while maintaining annotation reliability, we couple model-assisted expansion with uncertainty-aware verification to identify uncertain or weakly supported predictions for targeted human review. Building on TacFG-9, we further propose TacRAGate, a similarity-gated tactile captioning framework for reliable fine-grained description generation. TacRAGate fuses retrieved captions and structured attribute labels as semantic evidence, and adaptively adjusts its inference behavior according to the strength of the retrieved evidence. Experiments demonstrate that the structured tactile-vision-language supervision in TacFG-9 improves tactile representation learning, while TacRAGate produces more semantically faithful tactile descriptions by fusing retrieved semantic evidence under partial observations.