Tactile sensing is critical for dexterous robot manipulation, but gathering real tactile data on robot hardware is expensive and hard to scale across different hand designs and sensors. In this preprint, Yeonguk Yu, Jaehong Kim, and Byungok Han introduce HuTact, a framework that learns tactile-aware hand representations from large-scale human hand-object interaction data. HuTact does not need real tactile signals during pretraining; instead, it captures interaction structure from human hand motions and later adapts to robot tactile observations using a small amount of unlabeled robot data.
HuTact was evaluated on grasp prediction, slip detection, object classification, and tactile-conditioned policy learning. The authors report that adding human HOI pretraining and a midtraining stage on unlabeled robot play data substantially improved data efficiency. For grasp prediction, HuTact reached 88.26% Macro-F1 while using only 25% of the labeled downstream data, matching the 88.23% score achieved by training from scratch on the full dataset.
These results suggest that human interaction data can serve as a scalable complementary source for robotic tactile representation learning, potentially reducing the amount of robot-collected data needed. The authors argue that this approach could help robots acquire useful tactile priors without building large tactile datasets for every new hand and sensor.