Tactile Perception for Embodied Intelligent Dexterous Hands: Sensor Principles, Sensor Placement, and Multimodal Fusion

To perform these movements better, dexterous hands need to be able to feel contact with objects and the surrounding environment. This article mainly introduces tactile technology for dexterous hands, and it discusses these technologies from the view of embodied intelligence. The article mainly introduces five types of tactile sensors: piezoresistive, capacitive, piezoelectric, triboelectric, and optical sensors. These sensors are categorized by their different working principles. This article compares their sensitivity, sensing range, response speed, ability to detect small details, and the difficulty of integrating them into the system. Secondly, the deployment strategies of sensors at fingertips, palms, full hand surfaces, and soft-bodied hands, as well as the engineering constraints they face, are analyzed. Thirdly, multimodal fusion methods of touch, vision, and force, as well as their roles in enhancing grasping stability, are discussed, along with preliminary explorations of combining tactile information with visual-linguistic-action models. The study finds that piezoresistive sensors achieve a 92.8% classification accuracy in fabric texture recognition, outperforming human subjective classification. Soft-bodied humanoid dexterous hands can achieve all grasping poses in the GRASP classification by integrating distributed tactile sensors. Based on literature analysis, this paper points out that the core bottlenecks of current research include insufficient long-term stability of sensors, lack of interfaces between tactile representation and basic models, and limited computational power for real-time closed-loop control. It also proposes future research directions such as self-powered sensing, tactile basic models, and simulation-to-reality transfer.