.png){kind=icon}
.png){kind=icon}
Return
A Sensor-Optimized Data Glove for Taxonomy-Level Human Grasp Recognition
.png){kind=icon}
PRE
AI
DOI：10.1109/JSEN.2026.3678973.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Embodied AI has advanced rapidly, yet the sim-to-real gap still limits the real-world transfer of skills learned in simulation, especially for fine-grained grasping. Human hand dexterity provides valuable priors for manipulation learning, but capturing and understanding detailed hand–object interactions remain challenging. Data gloves offer a promising solution; however, most prior work focuses on nongrasp gestures or object recognition, and few studies address fine-grained grasp types. We develop GloG, a custom data glove system that captures and recognizes grasp gestures. In the offline stage, we introduce a taxonomyguided sensor layout strategy that places thin-film pressure arrays and stretch sensors on critical hand regions to maximize discriminability while reducing redundancy. Using this glove, we compile a comprehensive dataset covering 33 typical grasp gestures following the Feix taxonomy. In the online stage, we propose a topology-adaptive graph Transformer (TAGT) for static grasp recognition by explicitly modeling intersensor coordination on the optimized sensor graph. Experiments show that the system can distinguish multiple grasps per object and generalize across the object shape and size variations. GloG provides a scalable interface for collecting fine-grained human grasp demonstrations and recognition labels, which may support future robotic manipulation learning.
Keywords:
Multimodal data glove
sensor layout optimization
static grasp recognition
topology-adaptive graph Transformer (TAGT)
Journal
IF:
4.5
Papers:
2.1W
Citations:
7.3W