Data gloves, haptic feedback and touch sensors: how robots get a sense of touch from people, which techniques work, who supplies them and who argues what, with a source for every fact.
Gloves have become the bridge between people and robot hands. A data glove records how the fingers move; a haptic glove lets a teleoperator feel what the robot touches; a tactile glove records where and how hard the hand presses; and an exoskeleton glove copies the robot hand's joints so the recording fits the robot directly. Behind them sits a fast-growing market of touch sensors for robot fingertips and skins. The research case got stronger in 2025 and 2026: a glove with force feedback lifted a carton task from 4 to 10 successes out of 10 (DOGlove, arXiv 2502.07730), a robot trained only on tactile-glove data reached 77% on force-sensitive tasks (Feel the Force, arXiv 2506.01944), and MANUS became the official data glove of NVIDIA's Isaac Teleop (MANUS, 22 May 2026). Money followed: China's touch-sensor makers raised more than a billion yuan each in single rounds in 2026 (Caixin, 7 Mar 2026; TechNode, 20 Jul 2026). The open questions are which glove fits which job, whether the operator needs to feel, whether touch recorded on a glove transfers to a robot's different sensor, which sensing principle wins, whether sensors survive a working shift, and who supplies at scale.
Which glove for which job?
Tracking gloves record finger motion. Bend and stretch sensors are cheap and the most used (68 papers in a 2024 review) but need calibration per user and drift; motion-sensor gloves drift and suffer from magnetic fields; electromagnetic trackers do not drift but metal distorts them (systematic review, arXiv 2405.15417).
Haptic gloves give the operator feedback: vibration motors (bHaptics, $249 a pair), force-feedback brakes or actuators (SenseGlove Nova 2 at $5,999 at launch; SenseGlove R1 for humanoid teleoperation) or air-driven skin displacement (HaptX G1, $5,495 a pair in a 2025 review) (bHaptics; Auganix; Knoxlabs; PR Newswire).
Tactile gloves record pressure on the hand: ART-Glove records joints and contact together with 2,048 touch points at 120 Hz (arXiv 2606.16370); PaXini's PXCap gloves use its own 6D touch sensors (PR Newswire).
Does the operator need to feel what the robot touches?
The only direct test of data quality, on eight real doors, found haptic feedback raised data throughput by 6% and the success of the trained policy by 11% (Cuan, Okamura, Khansari, arXiv 2211.03020).
DOGlove's carton task went from 4/10 without feedback to 9/10 with force and 10/10 with force and fingertip haptics; policies trained on its data reached 85% to 90% (arXiv 2502.07730).
HapCompass lifted key insertion from 63% to 100% and cut peak force from 98 N to 66 N, in a small preliminary study (arXiv 2603.30042); FACTR's force-feedback teleoperation raised completion by 65% (arXiv 2502.17432).
Does touch recorded on a glove transfer to a robot's sensor?
Same sensor on both: OSMO puts the same tactile glove on the person and the robot, so no conversion is needed; a wiping policy trained only on human data reached 72%, beating vision only (arXiv 2512.08920).
Aligning different sensors: TactAlign maps glove touch to robot-hand touch with less than five minutes of human data (arXiv 2602.13579); GIFT transferred glove-sensed force to a robot hand with no touch sensors and cut grip force by 53% (arXiv 2609.14173).
General touch models are not yet general: T3, trained across 13 sensors, transferred to an unseen sensor with only minor improvement over random guessing for classification (arXiv 2406.13640); Meta's Sparsh encoders help across tasks but remain tied to camera-based sensors (arXiv 2410.24090).
Which sensing principle wins on the robot?
Camera-based touch (GelSight, Digit 360, Daimon, Xense, ViTai, Sharpa) gives image-like data that suits modern models, but soft gels wear, each sensor needs calibration, and force models transfer poorly between units (survey, arXiv 2608.15490). GelSight Mini sells for $510 and replacement gels for $42 to $70 (GelSight store).
Magnetic skins (XELA uSkin, ReSkin, AnySkin) are thin, cheap and swappable in 12 seconds, but vary between units and react to outside magnetic fields (arXiv 2409.08276).
Pressure films (piezoresistive) are the cheapest: a dense pad with its reader costs about $20 (3D-ViTac, arXiv 2410.24091).
Do touch sensors survive a working shift?
Makers publish cycle ratings: Robotiq rates TSF-85 above 1.5 million cycles; Tacterion's plyon flex datasheet says above 5 million; Meta rated ReSkin for more than 50,000 interactions (Robotiq; Tacterion datasheet; Meta). These are company claims; no field failure rate has been published.
Gloves wear too: Pressure Profile Systems describes copper traces cracking at the thumb in its TactileGlove, fixed with serpentine routing (PPS).
Swap time is becoming the metric: AnySkin replaces a skin in 12 seconds against 58 seconds for DIGIT and 82 to 236 seconds for ReSkin (arXiv 2409.08276).
Who supplies touch at scale, and at what price?
China leads on money: PaXini raised more than RMB 1 billion in a Series B at a valuation above RMB 10 billion and started IPO preparation in September 2026 (Caixin; KuCoin); Yimu raised more than RMB 1 billion in a Series E (TechNode); Daimon raised an RMB 1 billion A round (Huxiu); Tashan claims more than 80% of China's humanoid touch-sensor shipments (company claim, same source).
Prices are falling: a PaXini module starts from RMB 199 according to Huxiu; CICC models a humanoid touch sensor falling from RMB 400 in 2024 to RMB 200 in 2030 (Sina Finance, 24 Jun 2025).
Volume is starting: Fulai reported more than 30,000 touch sensors delivered to the hand maker Linkerbot (search snippet of Gasgoo).