A new arXiv preprint introduces TacClaw-Infinity, a large-scale dataset designed for dexterous manipulation research. The dataset combines vision, tactile, language, and action data, aiming to support robot learning from diverse human demonstrations.

The authors note that while large-scale robot learning increasingly depends on such demonstrations, tactile manipulation data remain difficult to scale. They also point out that existing multimodal datasets often lack consistent quality control and structured organization, which TacClaw-Infinity seeks to address.

As the abstract is truncated, details on dataset size, collection methods, and evaluation are not yet available. The work is posted on arXiv under identifier 2610.20320.