Feel Robot Feels: Tactile Feedback Array Glove for Dexterous Manipulation
Authors: Feiyu Jia, Xiaojie Niu, Sizhe Yang, Qingwei Ben, Tao Huang, Feng zhao, Jingbo Wang, Jiangmiao Pang
Organizations: University of Science and Technology of China · Shanghai AI Laboratory · The Chinese University of Hong Kong · Shanghai Jiao Tong University
Teleoperation is a key approach for collecting high-quality, physically consistent demonstrations for robotic manipulation. However, teleoperation for dexterous manipulation remains constrained by: (i) inaccurate hand-robot motion mapping, which limits teleoperated dexterity, and (ii) limited tactile feedback that forces vision-dominated interaction and hinders perception of contact geometry and force variation. To address these challenges, we present TAG, a low-cost glove system that integrates precise hand motion capture with high-resolution tactile feedback, enabling effective tactile-in-the-loop dexterous teleoperation. For motion capture, TAG employs a non-contact magnetic sensing design that provides drift-free, electromagnetically robust 21-DoF joint tracking with joint angle estimation errors below 1 degree. Meanwhile, to restore tactile sensation, TAG equips each finger with a 32-actuator tactile array within a compact 2 cm^2 module, allowing operators to directly feel physical interactions at the robot end-effector through spatial activation patterns. Through real-world teleoperation experiments and user studies, we show that TAG enables reliable real-time perception of contact geometry and dynamic force, improves success rates in contact-rich teleoperation tasks, and increases the reliability of demonstration data collection for learning-based manipulation.
Figures & tables
Fig. 2: Hardware architecture of TAG. (a) Exploded view of the joint encoder, utilizing a magnetometer and ring magnet for angle sensing. (b) Exploded view of the tactile feedback module, a multi-layer EOP stack sandwiching a membrane between PCBs. (c) TAG is compatible with diverse arm tracking solutions, such as 6DoF trackers or exoskeletons. (d) Kinematic configuration distributing 21 encoders for full-hand motion capture.
Fig. 3: Physical dimensions and form factor of the TAG feedback module. The entire assembly measures only 29×18.4×5.5 mm.
Fig. 4: Dynamic demonstration of the TAG feedback module. (a) Active area expansion in Pressure Mode under varying pressure levels. (b) Dynamic force profiles simulating forward-to-backward motion.
Fig. 5: Exoskeleton glove tracking performance. (a) Experimental setup for joint-tracking. (b) Real-time tracking results for discrete and continuous trajectories. (c) Long-term stability test over 1000s of operation. (d) Statistical distribution of instantaneous tracking errors with Gaussian fit.
Fig. 6: EMI resilience evaluation. (a) Setup for near-field interference testing. (b) Angular stability comparison between the Manus glove and TAG.
Fig. 7: Tactile sensing and shape mapping performance across various end-effectors. Comparison of raw sensing inputs (i) and physical rendering outputs on TAG (ii) across various robot end-effectors (right). TAG effectively translates diverse sensor geometries into coherent tactile feedback: (a) LeapTac: linear ridge; (b) Inspire Hand: circular area; (c) XHand: dual contact points; (d) GelSight Mini and (e) TacSL: complex shapes.
TABLE I: Comparative Analysis of Tactile Sensing Performance. Res: Spatial Resolution. Thres: Minimum Force Threshold. Var. indicates sensitivity is tunable via gel stiffness. EM: Electromagnetic.
Fig. 8: User Study. (a): The operator, blindfolded and wearing headphones, identifies the shape of the object pressed against the index finger of G1 solely through tactile feedback. (b): Four objects with different contact geometries: single-point, two-point, line, and plane.
Fig. 9: Classification results of the user study. The Sankey diagrams visualize the mapping between actual stimuli and user predictions. The width of the flows represents the number of trials.
Fig. 10: Real-world teleoperation tasks using TAG. (a) Multimodal lamp control : Hierarchical interaction via short press, press-and-turn, and subtle rotation. (b) Sequential filament pinching : Distinguishing successful capture of a 1.75mm filament from empty pinches via localized tactile feedback. (c) Multi-level spring compression : Maintaining three distinct pressure levels (0.5N, 3.5N, 7.5N) guided by rendered tactile feedback.
Teleoperation System
Success Rate ( ↑ )
Average Time ( ↓ )
Stage I: Power Toggling (Short Press)
HOMIE [ 1 ]
12/15
6.1s
TAG w/o feedback
14/15
3.2s
TAG w/ feedback
15/15
3.6s
Stage II: CCT Adjustment (Press & Turn)
HOMIE [ 1 ]
9/15
23.5s
TABLE II: Results of multimodal desk lamp control.
Teleoperation System
Success Rate ( ↑ )
Average Time ( ↓ )
HOMIE [ 1 ]
5/15
42.8s
TAG w/o feedback
7/15
35.6s
TAG w/ feedback
13/15
22.4s
TABLE III: Results of sequential filament pinching (2 consecutive trials).
Teleoperation System
Blindfolded ( ↑ )
With Visual ( ↑ )
HOMIE [ 1 ]
2/15
10/15
TAG w/o feedback
3/15
12/15
TAG w/ feedback
13/15
15/15
TABLE IV: Success rate of multi-level spring compression.
Fig. 11: Imitation learning. (a) Experimental setup for real-world tasks. (b) Task illustration: CCT Adjustment and Fully Spring Compression.
Fig. 12: Tactile Feedback Module and Control Circuit. (a) Top view of the Tactile Feedback Module and its single-layer PCB; (b) Five modules correspond to the five fingers, each independently driven by an HV507 chip.
Fig. 13: Hand retargeting performance. The optimization-based method is applied to three distinct robotic hands: XHand, LeapHand and InspireHand. Despite differences in kinematics and dimensions, all hands successfully reproduce the target gesture captured from TAG.
Fig. 14: Illustration of teleoperation tasks in Press Mode. (a) Multimodal Control of a Smart Desk Lamp. (b) Multi-level Spring Compression.
Fig. 15: Illustration of Filament Pinching and Pull-out Task. The schematic shows the thumb and index finger pinching the filament.
Fig. 16: Physical prototype and schematic of the LeapTac sensor. (a) The two-layer Flexible Printed Circuit (FPC) electrodes and the fully assembled LeapTac module. (b) The circuit schematic illustrating the readout mechanism for individual sensing points (taxels).