Tactile Data Collection Devices for Dexterous Manipulation
Sources: Research synthesis, 2026-04-05; Session notes, 2026-04-09; Session notes, 2026-04-15; arXiv digests, 2026-W07–W32; Song et al. (MIT et al.), arXiv:2512.16842, 2025-12 Raw: Dexterous Hand Tactile Data Collection Devices for LfD; Wuji Glove Commercial Landscape; Lai Lab Soft Electronics Glove; W26; OpenTouch Updated: 2026-08-12
Overview
Wearable tactile devices for dexterous manipulation Learning from Demonstration (LfD) split along two axes: sensing direction (record from human for policy input, vs. feedback to human to improve demonstration quality) and sensor technology (magnetic, optical FBG, camera-based, resistive). Open-source and sub-$600 devices now match commercial performance. Operator haptic feedback matters as much as tactile recording — operators with haptic feedback produce qualitatively better demonstrations.
A critical market confusion: most commercial gloves are Type B (relay robot tactile → human operator) while researchers typically want Type A (record human tactile → robot policy). The commercial Type A market remains immature.
Two Device Goals (Complementary, Not Competing)
Type A — Sensing gloves (record human → policy input): capture tactile data from the operator for the policy’s input. The human’s contact signals during demonstration are recorded and used as conditioning signals in downstream policies.
- Examples: OSMO, TacCap, Wuji Glove, Pressure Profile TactileGlove II
Type B — Feedback gloves (relay robot → human operator): relay robot-side contact back to the operator during teleoperation to improve demonstration quality. The policy receives standard proprioceptive/visual inputs; the haptic feedback improves the human’s demonstration.
- Examples: TAG, DOGlove, HaptX, SenseGlove, Manus
These are complementary — you can use both simultaneously.
Device Profiles
OSMO (Meta FAIR, Dec 2025) — Open Source
Sensor: 12 three-axis magnetic sensors on fingertips and palm, 0.3–80 N range.
Key claim: demonstrated zero-real-robot-data training — policies trained on human hand demonstrations alone (no robot present) achieved 71.69% success on marker wiping tasks.
Pipeline: SAM2 + HaMeR hand pose estimation → IK retargeting → MuJoCo joint commands. Compatible with Meta Quest 3, Aria Gen 2, Manus Quantum, HaMeR.
Best for: in-the-wild scalable collection without a physical robot. Open hardware, open-source.
TacCap (Stanford, Mar 2025) — Open Source
Sensor: FBG (fiber Bragg grating) thimble-form-factor sensors on fingertips. 2 kHz sampling, 0.028 N minimum detectable force, 360° finger coverage (vs. GelSight’s front-only).
Key principle: mount identical sensors on both human and robot fingers — eliminating the cross-sensor domain gap entirely. 100% grasp stability in teleoperation mode.
Trade-offs: FBG is EM-immune but requires an expensive bench-bound optical interrogator. Fingertip-only coverage (no palm).
Best for: maximum tactile fidelity with zero domain gap. High cost from interrogator equipment.
TAG / Feel Robot Feels (Mar 2026, <$500)
Sensor: 21-DOF magnetic joint tracking (±0.35°, drift-free) + 32-actuator EOP (Electro-Osmotic Pump) tactile arrays per fingertip. Novel actuator: drives fluid through microchannels, achieving higher resolution than vibrotactile at lower cost than pneumatic.
Evidence: filament pinching: 87% success with TAG feedback vs. 33% without. Downstream IL policies from TAG demonstrations: 87–100% success.
Best for: teleoperation with bidirectional tactile feedback where feedback quality matters more than recording-side sensor coverage.
DOGlove (Tsinghua, Feb 2025, <$600) — Open Source
Sensor: 21-DOF motion capture + 5-DOF bidirectional cable-driven force feedback + 5 LRA vibrotactile actuators. FSR sensors on robot fingertips provide the contact signal retargeted to operator hand.
Evidence: in-hand carton rotation: 10/10 with DOGlove vs. 1/10 with AnyTeleop (vision-only). Downstream IL benchmark: 85–90% success with 3D Diffusion Policy. Tested on LEAP Hand (real), Shadow/Inspire/Allegro (simulated).
Best for: teleoperation with feedback; has the most validated downstream IL results of any device in this class; open-source.
DEXOP (MIT/Stanford, Sep 2025) — Research Prototype
Design: passive mechanical exoskeleton coupling human fingers to a co-designed passive robot hand via 4-bar linkages. Three variants (12, 9, 7 DOF). Camera-based GelSim(ple) sensors for whole-hand coverage including palm and proximal phalanges.
“Perioperation” paradigm: human directly manipulates via exoskeleton, feeling contact through proprioception. Data collection: 11s vs. 86s per task compared to standard teleoperation.
Best for: whole-hand contact capture including palm; maximally natural demonstration kinematics via mechanical coupling.
Wuji Glove (Commercial)
Sensor: 6-axis IMU + tactile matrix + joint tracking. 1000 Hz via ROS2. Tactile matrix density is not publicly disclosed.
Integration: Designed as a matched pair with Wuji Hand; supports Apple Vision Pro retargeting. Commercial plug-and-play is the primary advantage — no custom hardware assembly needed.
Type A only: Records human contact data for policy input. Does not provide feedback to the operator.
Pressure Profile TactileGlove II (Commercial Type A)
65 sensing points, Bluetooth 5. Currently the most complete commercial Type A glove. Designed for ergonomics/human factors research rather than robotics — no ROS integration. Requires custom software bridge for LfD use.
Manus (Commercial Type B — NVIDIA Partnership)
Manus was selected as NVIDIA Isaac Teleop’s official data glove partner at 2026 GTC, accelerating commercial glove adoption into embodied AI training pipelines. Primarily Type B (feedback to operator), but the NVIDIA integration adds data collection use cases.
DexUMI (Stanford/Columbia/CMU, May 2025) — Research Prototype
Design: wearable exoskeleton co-designed per target robot hand (Inspire Hand, XHand). 3.2× data collection efficiency vs. teleoperation. Video inpainting removes human hand from training observations to close visual domain gap.
Best for: multi-platform deployment with per-robot optimized exoskeletons.
OpenTouch (MIT et al., Dec 2025) — Research Dataset
Sensor: FPC (Flexible Printed Circuit) tactile glove with 169 taxels uniformly covering fingers and palmar surface; synchronized with hand-tracking and egocentric video cameras.
Key capability: first in-the-wild egocentric full-hand tactile dataset — 5.1 hours of synchronized video-touch-pose data; 2,900 curated clips with text annotations. Benchmarks for cross-sensory retrieval and tactile classification.
Multimodal advantage: combining video, pose, and tactile data substantially outperforms individual modalities. Tactile signals — sparse and highly structured unlike natural images — are best processed by lightweight CNN encoders rather than ResNet-18 backbones.
Cross-dataset generalization: models trained on OpenTouch successfully retrieve tactile patterns from Ego4D video queries, indicating cross-dataset generalization.
Affiliations: MIT, Duke University, Brown University, University of Washington, Harvard University.
Best for: in-the-wild egocentric tactile data collection; cross-sensory retrieval research; tactile classification benchmarking. Not designed for real-time robot teleoperation.
Sensor Technology Comparison
| Device | Type | Sensor | Coverage | EM Immunity | Cost | Open |
|---|---|---|---|---|---|---|
| OSMO | A | Magnetic 3-axis | 12 points (fingertips + palm) | Low | Open | Yes |
| TacCap | A | FBG optical | Fingertips only (360°) | High | ~$10k (interrogator) | Yes |
| Wuji Glove | A | IMU + tactile matrix | Undisclosed | — | Commercial | No |
| Pressure Profile TG II | A | Resistive (65 pts) | Whole hand | — | Commercial | No |
| TAG | B | EOP feedback | Per-fingertip feedback | High | <$500 | Partial |
| DOGlove | B | FSR (robot side) | 5 fingertips | High | <$600 | Yes |
| DEXOP | A | Camera-based | Whole hand | High | Research | No |
| DexUMI | A | FSR / EM (robot side) | Fingertips | Medium | Research | No |
| OpenTouch | A | FPC (169 taxels) | Full hand (fingers + palm) | Medium | Research | Yes |
Choosing a Device
In-the-wild collection without a robot: OSMO. No robot required, open-source, magnetic sensors on fingertips and palm.
Maximum tactile fidelity, zero domain gap: TacCap. Identical FBG sensors on human and robot; trade-off is expensive interrogator and fingertip-only coverage.
Teleoperation with haptic feedback, validated IL results: DOGlove. <$600, open-source, tested on multiple robot hands with quantified downstream policy performance.
Teleoperation with high-resolution tactile feedback: TAG. Novel EOP actuators provide better spatial resolution than vibrotactile; fewer published downstream IL results than DOGlove.
Whole-hand contact including palm, most natural kinematics: DEXOP. Mechanical coupling eliminates retargeting; requires per-robot exoskeleton design.
Emerging Research: Soft Electronics Approach
Soft electronics (self-healing skins, liquid metal fibers, transparent flexible substrates) offer a differentiated capability profile compared to commercial rigid IMU gloves like Wuji:
Distributed pressure arrays across the full palm surface: rigid IMU gloves place discrete sensors at fixed points (Wuji: undisclosed density; OSMO: 12 magnetic sensors at fingertips + palm). Soft sensing skin can cover the full palm surface with continuous distributed pressure measurement — the sensing gap most commercial gloves have not solved.
Self-healing electronic skin for long LfD campaigns: flexible sensing layers that recover from physical damage reduce maintenance cost for repeated demonstration tasks (hundreds of trials per day). This is relevant for building large-scale robot training datasets where glove wear is a practical bottleneck.
Liquid metal fiber integration: potential for partial self-powering — converting hand mechanical energy to electrical energy — reducing battery constraints for long-duration data collection. Not yet commercially deployed.
Low motion interference: transparent (88% in lab demonstrations), thin-layer substrates do not alter the demonstrator’s proprioceptive feel, producing more natural demonstration kinematics and higher data quality.
Technical gaps vs. commercial baseline:
- Sampling rate typically lower than rigid IMU (commercial rigid IMU like Wuji: 1000 Hz; soft sensor sampling rates are device-dependent and often lower)
- Calibration pipelines more complex — flexible geometry changes with hand pose
- ROS2 joint state output requires custom engineering (commercial gloves provide this out-of-the-box)
- Laboratory demonstrations do not yet show robotics-integration maturity
Physiological signal multiplexing: flexible electrodes that measure PPG, EDA (skin conductance), and skin temperature alongside motion — potentially useful for surgical training applications and operator-state monitoring in HRI research. Secondary capability; not the primary LfD use case.
Foundation Tactile Policy: Cross-Sensor Generalization
FTP-1 (arXiv:2606.13102, 2026): the first generalist foundation model for tactile manipulation — analogous to foundation vision models but for tactile sensing.
Scale: pretrained on ~3,000 hours of data from 26 sources spanning 21 different tactile sensors.
Architecture: heterogeneous encoders project diverse tactile inputs (image-based, array-based, state-based) → unified morphology-aware latent tokens → shared tactile Transformer expert.
Results:
- +17.2% on seen sensor setups
- +31% success rate gain on two previously unseen tactile sensor configurations (zero-shot cross-sensor generalization)
Significance: prior tactile policies were hardware-specific — trained on one sensor, unusable on another. FTP-1 demonstrates cross-sensor generalization for the first time at this scale. The +31% gain on unseen sensors is the key capability; the architecture’s morphology-aware tokenization is what enables it.
AT-VLA (arXiv:2605.07308): adaptive tactile injection for VLAs — adds tactile signal conditioning to standard VLA architectures without requiring tactile sensors at training time. Enables deploying existing vision-only VLA architectures with tactile input during inference.
TAMEn (arXiv:2604.07335): Tactile-Aware Manipulation Engine — closed-loop data collection system for contact-rich tasks, addressing the data collection bottleneck for tactile policies.
Open Questions
- What is the minimum tactile resolution needed for a downstream diffusion policy to benefit? Devices range from 5 sensors (TacCap per hand) to 12 (OSMO) to full palm coverage (DEXOP).
- Can OSMO-style magnetic gloves maintain calibration across multi-day collection campaigns (sensor drift, mechanical wear)?
- How much does tactile signal quality degrade when robot hand morphology significantly differs from human morphology (e.g., 3-finger gripper vs. 5-finger hand)?
- At what sampling rate does soft-sensor data become useful for manipulation-speed tasks? Is the rigid-IMU 1kHz rate necessary, or is lower-frequency distributed pressure sufficient?