Chapter 7: TacGlove — Stretchable Tactile Data Glove

Summary

TacGlove [#26] extends Park et al.'s [2024] [#6] stretchable eGaIn glove from 3 fingers to 5 fingers and adds 8 OSMO [#18]-style 3-axis magnetic tactile sensors placed on 5 fingertips + 3 palm regions, making it the first stretchable data glove that simultaneously measures joint angles and tactile forces. Developed in Prof. Park Yong-Rae's lab at SNU, TacGlove closes the gap where existing data gloves were split into kinematics-only or tactile-only devices.

7.1 Introduction: The Need for an Integrated Kinematics+Tactile Glove

The core hardware gap repeatedly identified throughout Parts I and II is the absence of an integrated pipeline that simultaneously measures joint angles and tactile forces (Chapter 2, Chapter 3).

• Kinematics-only gloves lack contact force information, limiting their utility for contact-rich tasks.
• Tactile-only gloves require separate exoskeletons or camera-based systems for kinematics, constraining wearability and scalability.
• No system integrates both modalities in a single stretchable device.

This gap is TacGlove's starting point. TacGlove provides the hardware for simultaneous kinematics + tactile data collection, serving as the physical foundation for TacTeleOp's (Chapter 8) large-scale data collection and TacPlay's [#27] (Chapter 9) autonomous learning.

7.2 Survey of Existing Data Gloves

Existing data gloves fall into three categories: kinematics-only, tactile-only, and kinematics+tactile integrated.

Kinematics-Only

• Park et al.^[1] [#6]: 9 eGaIn strain sensors measuring joint angles and bone lengths of 3 fingers. Published in Nature Communications. Joint angle error 4.16 degrees, bone length error 2.1mm. The only research glove using stretchable materials, but no tactile sensors, 3 fingers only.
• DOGlove ^[4]: 21-DoF kinematics + force feedback. Under $600. Provides force feedback, not sensing.
• TAG ^[5]: 21-DoF kinematics + haptic feedback. Under $500. Feedback only, same as DOGlove.
• Commercial (Manus, SenseGlove, HaptX): Joint measurement + force feedback. No tactile sensing capability.

Tactile-Only

• OSMO ^[2] [#18]: 12 three-axis magnetic tactile sensors. No kinematics measurement — requires a separate Manus glove. Embodies the Embodiment Bridge concept for human/robot shared use.
• ViTaM [Nature Comms, 2024]: 1,152-channel pressure sensors. Kinematics estimated via external camera. Not self-contained.
• STAG [MIT, Nature 2019]: 548 tactile sensors. 76% object identification accuracy. No kinematics measurement.

Kinematics+Tactile Integrated

• Zhu et al. [PKU, Engineering 2023]: 15 IMUs + 14 Velostat force taxels. 5 fingers. The only existing system that simultaneously measures kinematics and tactile. However, Velostat is binary/sparse (no shear force), low resolution (14 taxels), and has never been used for end-to-end robot policy training.

Data Collection Frameworks

• DexCap ^[8]: A data collection framework combining magnet-based 3D hand tracking with SLAM-based localization. Demonstrated retargeting quality to LEAP Hand, but lacks tactile sensors and suffers from 3D-printed exoskeleton deformation issues.
• Skill Capture Glove ^[13] [#29]: A wearable data glove co-designed with "exactly the same geometry and sensor layout" as the robot hand. Over 2,000 units deployed, collecting 10M+ episodes, with Skill Transform (90% conversion success rate) converting human data to robot data. Key limitations: restricted to gripper form factor (2-DoF), no tactile sensors, proprietary (not reproducible).

Comparison Table

The message from this table is clear: no existing system integrates kinematics and multi-axis tactile sensing in a single stretchable glove.

7.3 Limitations of Existing Gloves

The limitations of existing gloves are systematized into three categories:

Limitation 1 — No kinematics+tactile integration. Only Zhu et al. ^[3] integrates both modalities, but their 14 binary taxels cannot measure 3-axis forces and have never been used for robot policy learning. OSMO has excellent tactile but no kinematics, requiring a separate Manus glove. This split architecture causes time synchronization errors, wearing discomfort, and limited scalability.

Limitation 2 — No stretchable materials. OSMO, DOGlove, and commercial gloves all use rigid or semi-rigid structures. Rigid gloves impede natural hand motion, and as DexUMI's 3D-printed exoskeleton demonstrated (Chapter 2), deformation and slippage issues arise. Only Park et al.^[1] uses eGaIn-based stretchable materials.

Limitation 3 — No industrial durability validation. Among 40+ related papers, zero gloves have been validated in actual factory environments (Chapter 1). No durability data exists for 8-hour continuous wear, washing, or shift changeover.

7.4 TacGlove Design

TacGlove extends Park et al.'s [2024] eGaIn glove in two directions.

Design Principles

Derived from the analyses in Chapters 2 and 3:

From Park et al.^[1] to TacGlove

Extension 1 — 5 fingers: Adding ring and pinky fingers for whole-hand coverage. The eGaIn sensor placement pattern is replicated on the additional fingers. Sensor count increases linearly from 9 to approximately 15, a manufacturing process extension rather than a structural redesign.

Extension 2 — Tactile sensors: Adding 8 OSMO-style 3-axis magnetic tactile sensors. Placement: 5 fingertips + 3 palm sections (thenar, hypothenar, central). This prioritizes fingertip contact information relative to OSMO's 12-sensor layout. This 5+3 placement aligns with ^[14]'s quantitative analysis — on a DRL-based in-hand reorientation benchmark, the contribution of upper-palm tactile locations is essentially on par with fingertip distal phalanges (27.87 vs 28.21 consecutive successes), and for small objects the configuration using all five palm locations achieves the highest score of any configuration. For TacGlove's target industrial tasks (small cosmetic containers, small assembly parts), the palm thenar/hypothenar/central triad is not an "auxiliary surface" but a dominant sensing site.

The palm-sensor evolution can be understood along a hardware-variant spectrum — from passively compliant palms (ROMEO ^[15]) to high-density visuotactile palms (TacPalm SoftHand ^[16]) to actuated palms (with a comprehensive taxonomy in ^[17]). TacGlove occupies the coordinate "all-day human-wearable while sharing a human–robot sensing space," which the sister directions TacPlay (Chapter 9) and TacTeleOp (Chapter 8) exploit.

Smart Glasses Time Synchronization

TacGlove synchronizes with a headband webcam or smart glasses to simultaneously record joint angles + tactile + egocentric RGB. Data schema:

{
  timestamp: float64,         // millisecond precision
  hand_joints: float32[15],   // MANO-compatible joint angles
  tactile: float32[8][3],     // 8 sensors × 3-axis forces
  egocentric_rgb: uint8[H][W][3],  // 30 Hz
  head_pose: float32[6]       // 6-DoF
}

MANO-compatible joint angles are directly compatible with EgoVLA's ^[10] unified action space and EgoDex's ^[11] per-finger tracking, enabling use of existing pretrained models.

7.5 Data Usefulness Demonstration Strategy

TacGlove is a hardware paper. How should the hardware's usefulness be demonstrated? We analyze existing gloves' demonstration strategies ranked by difficulty, then propose a 3-stage strategy tailored to TacGlove.

Existing Glove Demonstration Precedents

That Park et al.^[1] was published in Nature Communications with sensor accuracy metrics alone demonstrates that, for hardware papers, sensor precision and reliability can constitute sufficient contribution.

Recommended 3-Stage Demonstration Strategy

Stage 1: Sensor accuracy benchmarking (Months 1-2). Quantitative evaluation at Park et al.^[1] level. Target: joint angle error <5 degrees, tactile classification accuracy >95%. Demonstrate that the 5-finger extension maintains the original 3-finger accuracy and that the 3-axis tactile sensors are significantly superior to Zhu et al.'s ^[3] binary Velostat.

Stage 2: Cosmetics task dataset collection + grasp/task classification (Months 2-4). Collect cosmetics manual process data using headband webcam + TacGlove. 3-5 workers perform 10+ tasks (capping, label application, assembly packaging, etc.). Report grasp type classification and task classification accuracy from collected data. Using STAG's 76% object identification as a reference point, show that TacGlove achieves higher classification accuracy through the combined kinematics+tactile signal.

Stage 3: Vision vs. tactile comparison — information invisible to cameras (Months 3-5). This is TacGlove's most compelling demonstration. Visually show that torque/force changes during capping tasks are invisible in camera footage but clearly captured in tactile signals. Example: under-tightened vs properly tightened caps are indistinguishable in egocentric RGB, but the torque profiles in tactile time series are clearly different.

Core message: "TacGlove is the first stretchable glove that simultaneously measures joint angles and contact forces, and we demonstrate that the collected tactile signals contain task-discriminative information invisible to vision."

7.6 Hardware Differentiation from OSMO

The hardware differences between TacGlove and OSMO are threefold:

1. Stretchable vs Rigid. OSMO uses a rigid structure. TacGlove uses Park et al.'s [2024] eGaIn silicone-based stretchable material that does not impede natural hand motion. This is essential for extended wear (8 hours/shift) and naturalistic demonstration data collection.

2. Integrated kinematics. OSMO is tactile-only and requires a separate Manus glove for kinematics measurement. TacGlove self-measures joint angles via eGaIn strain sensors, collecting both modalities from a single device. This eliminates time synchronization errors, improves wearing comfort, and reduces system complexity.

3. Factory durability design. OSMO was designed for laboratory environments. TacGlove targets factory conditions (washing, shift changeover, vibration, contamination) through silicone shell and modular sensor replacement as design goals. This is the prerequisite for TacTeleOp's (Chapter 8) 800-hour large-scale collection.

4. Differentiation from Skill Capture Glove. The Skill Capture Glove takes the most aggressive approach to eliminating the kinematics gap through hardware co-design, but is restricted to gripper form factors and includes no tactile sensing. TacGlove takes the opposite strategy: integrating distributed tactile (24 channels) into a dexterous 5-finger form factor, targeting the contact-rich task domain that the Skill Capture Glove cannot reach.

7.7 Limitations and Open Questions

1. 3-to-5 finger extension timeline. Hardware development is slower than software. Extending from Park et al.'s [2024] 3-finger prototype to 5 fingers is a linear sensor count increase, but wiring complexity and manufacturing yield may present unexpected challenges. Piloting with the 3-finger prototype first, then extending to 5 fingers, is the realistic path.

1. Magnetic sensor factory interference. OSMO's 3-axis magnetic sensors use MuMetal shielding, but the interference level in motor- and electromagnetic-equipment-dense factory environments requires pre-measurement. If interference is severe, calibration protocols or alternative sensor types must be considered.

1. Glove durability and lifespan. Park et al.^[1] reported eGaIn sensor stability over thousands of deformation cycles, but no data exists for continuous industrial use over 8 hours × 20 days. Long-term stability verification for sensor drift, silicone fatigue, and wiring disconnection is required.

References

1. Park, M., & Park, Y.-L. et al. (2024). Stretchable Glove for Hand Motion Estimation. Nature Communications. https://www.nature.com/articles/s41467-024-50101-w #6 scholar
2. Yin, J., et al. (2025). OSMO: A Large-Scale Tactile Glove. arXiv. https://arxiv.org/abs/2512.08920 #18 scholar
3. Zhu, Z., et al. (2023). A Soft Robotic Glove with Integrated Sensing and Haptic Feedback. Engineering. https://doi.org/10.1016/j.eng.2023.01.011 scholar
4. Ding, Z., et al. (2025). DOGlove: Dexterous Manipulation with a Low-Cost Open-Source Glove. RSS 2025. scholar
5. Shanghai AI Lab (2026). TAG: Teleoperation Accessible Glove. arXiv. scholar
6. Sundaram, S., et al. (2019). Learning the Signatures of the Human Grasp Using a Scalable Tactile Glove. Nature. (STAG) scholar
7. Han, S., et al. (2024). ViTaM: A High-Resolution Vision-Based Tactile Sensor. Nature Communications. scholar
8. Chi, C., et al. (2024). DexCap: Scalable and Portable Dexterous Manipulation Data Collection. RSS 2024. scholar
9. Liu, Q., et al. (2025). VTDexManip: Visual-Tactile Dataset for Dexterous Manipulation. ICLR 2025. scholar
10. Yang, R., et al. (2025). EgoVLA: Egocentric VLA with MANO. arXiv. scholar
11. Hoque, M., et al. (2025). EgoDex: Learning Dexterous Manipulation from Egocentric Video. arXiv. scholar
12. Xu, M., et al. (2025). DexUMI: Universal Manipulation Interface. arXiv. #8 scholar
13. Sunday Robotics (2025). ACT-1: Skill Capture Glove & Skill Transform. #29 scholar
14. Almeida, J. D., Falotico, E., Laschi, C., & Santos-Victor, J. (2025). The Role of Touch: Towards Optimal Tactile Sensing Distribution in Anthropomorphic Hands for Dexterous In-Hand Manipulation. IEEE ICNSC 2025. https://arxiv.org/abs/2509.14984 #41 scholar
15. Liu, S. Q., & Adelson, E. H. (2024). A Passively Bendable, Compliant Tactile Palm with RObotic Modular Endoskeleton Optical (ROMEO) Fingers. ICRA 2024. https://arxiv.org/abs/2404.08227 scholar
16. Zhang, N., Ren, J., Dong, Y., Gu, G., & Zhu, X. (2025). Soft Robotic Hand with Tactile Palm-Finger Coordination (TacPalm SoftHand). Nature Communications 16:2395. https://doi.org/10.1038/s41467-025-57741-6 #40 scholar
17. Pozzi, M., Malvezzi, M., Prattichizzo, D., & Salvietti, G. (2024). Actuated Palms for Soft Robotic Hands: Review and Perspectives. IEEE/ASME Transactions on Mechatronics, 29(2):902–921. scholar