Dexterous Hand Tactile Data Collection Devices for LfD

Research Question

What wearable devices with integrated tactile sensing are suitable for collecting dexterous manipulation demonstrations for Learning from Demonstration (LfD) / Imitation Learning? Covering sensor technology, hardware design, data collection pipeline, robot hand compatibility, and commercial vs. research prototype availability.

Knowledge Map

Tactile sensing modalities — different transduction principles (magnetic, FBG optical, capacitive, camera-based, FSR resistive) have different tradeoffs in bandwidth, coverage, cost, and EM interference immunity. Choosing a device implicitly chooses a sensing paradigm.
Teleoperation vs. direct demonstration — some devices (gloves with haptic feedback) are teleoperation interfaces; others (wearable sensors) are worn during direct human manipulation. The data collection pipeline differs significantly.
Embodiment gap — the kinematic and morphological difference between human and robot hands means joint angles, contact locations, and forces don’t map 1:1. Understanding this gap motivates designs that either minimize it (same-sensor approach, co-designed exoskeleton) or bridge it via retargeting.
Imitation learning policy inputs — what modalities the downstream policy consumes determines what the device must capture. Vision-only policies need only pose; tactile-conditioned policies need synchronized contact signals at sufficient resolution and frequency.
Diffusion policy / action chunking — the dominant policy architecture for contact-rich manipulation; compatible with multi-modal conditioning including tactile, which motivates investment in tactile data capture infrastructure.

Sources Gathered

New sources clipped and analyzed during this research:

Clippings-osmo-open-source-tactile-glove-human-to-robot-skill-transfer — Meta FAIR open-source magnetic tactile glove; zero real robot data needed
Clippings-taccap-wearable-fbg-tactile-sensor-human-to-robot-skill-transfer — Stanford FBG fingertip sensors; same sensor on human and robot eliminates domain gap
Clippings-feel-robot-feels-tactile-feedback-array-glove-dexterous-manipulation — TAG glove with EOP tactile feedback arrays; bidirectional teleoperation device

Existing vault notes referenced:

Clippings-doglove-dexterous-manipulation-with-a-low-cost-open-source-haptic-force-feedback — DOGlove: 21-DoF, 5-DoF haptic force feedback, <$600, open-source
Clippings-glovity-learning-dexterous-contact-rich-manipulation-via-spatial-wrench-feedback — Glovity: spatial wrench feedback teleoperation system
Clippings-dexwild-dexterous-human-interactions-for-in-the-wild-robot-policies — DexWild: in-the-wild dexterous demonstration collection

Key Findings

Two orthogonal device goals require different hardware. Devices split into two camps: (a) sensing gloves that capture tactile data from the operator for the policy’s input (OSMO, TacCap), and (b) feedback gloves that relay robot contact to the operator to improve demonstration quality (TAG, DOGlove). These are complementary, not competing, and can be combined.
“Same sensor on human and robot” is the cleanest skill transfer design. TacCap’s approach of mounting identical FBG sensors on both human and robot fingertips eliminates the cross-sensor domain gap at the cost of constraining which robot hand can be used. OSMO takes the opposite approach — instrument the human hand richly, then retarget.
**Open-source and sub- $600 d e v i ces n o w ma t c h co mm erc ia lp er f or man ce . * * D OGl o v e (<$ 600), TAG (<$500), OSMO (open hardware), and TacCap (open hardware) all achieve competitive task performance while remaining accessible to labs without industrial budgets.
Tactile feedback to the operator matters as much as tactile recording. TAG and DOGlove both demonstrate that operators with haptic feedback produce qualitatively better demonstrations (e.g., filament pinching: 87% vs. 33% without feedback). The quality of the demonstration dataset is upstream of any policy improvement.
In-the-wild vs. structured teleoperation is a collection paradigm choice. OSMO targets in-the-wild collection (human wears glove, manipulates objects naturally, no robot present). DOGlove/TAG require a robot to be physically present for teleoperation. The former scales faster; the latter captures richer force signals.

Open Questions

Can OSMO-style magnetic gloves maintain calibration over multi-day collection campaigns (sensor drift, mechanical wear)?
How does tactile signal quality degrade when the robot hand morphology significantly differs from the human (e.g., 3-finger gripper vs. 5-finger hand)?
What is the minimum tactile resolution needed for a downstream diffusion policy to benefit — current devices range from 5 sensors (TacCap per hand) to 12 (OSMO) to effectively higher with camera-based sensors?
Are there commercial products (SenseGlove Nova, HaptX Gloves G1) that warrant comparison against these research devices at higher cost? (部分已在「A 類別商用方案補充」中回答)
Wuji Glove 的 tactile matrix 實際解析度是否足夠訓練觸覺條件化策略？（規格未公開）
Manus Metagloves Pro 成為 NVIDIA Isaac Teleop 官方指定手套（2026 GTC），其資料格式與 Isaac Lab 的整合深度如何？

Report

Device Taxonomy

Wearable devices for dexterous LfD data collection divide along two axes: sensing direction (human → record → policy vs. robot → feedback → human) and sensor technology (magnetic, optical FBG, camera-based, resistive).

Category 1: Tactile Sensing Gloves (record contact for policy input)

OSMO (Meta FAIR, Dec 2025, open-source) The most complete open-source solution for in-the-wild tactile data collection. 12 three-axis magnetic sensors on fingertips and palm, 0.3–80 N range. Unique advantage: demonstrated zero-real-robot-data training — policies trained on human demonstrations alone achieve 71.69% success on marker wiping. Pipeline: SAM2 + HaMeR hand pose estimation → IK retargeting → MuJoCo joint commands. Compatible with Meta Quest 3, Aria Gen 2, Manus Quantum, HaMeR.

TacCap (Stanford, Mar 2025, open-source) Thimble-form-factor FBG fingertip sensors. Key principle: mount identical sensors on human and robot fingers — eliminating the cross-sensor domain gap entirely. 2 kHz sampling, 0.028 N minimum detectable force, 360° finger coverage vs. GelSight’s front-only. Achieves 100% grasp stability in teleoperation mode. FBG technology is EM-immune but requires an optical interrogator (expensive, bench-bound). Open hardware at Stanford site.

Category 2: Haptic Feedback Gloves (relay contact from robot to operator)

TAG (Feel Robot Feels, Mar 2026, <$500) 21-DoF magnetic joint tracking (±0.35°, drift-free) + 32-actuator EOP tactile arrays per fingertip. Bidirectional teleoperation: captures hand pose AND feeds back contact geometry/pressure. On filament pinching: 87% with TAG vs. 33% without. IL policies from TAG demonstrations: 87–100% success. EOP (Electro-Osmotic Pump) is a novel actuator that drives fluid through microchannels — higher resolution than vibrotactile at lower cost than pneumatic.

DOGlove (Tsinghua, Feb 2025, <$600, open-source) 21-DoF motion capture + 5-DoF bidirectional cable-driven force feedback + 5 LRA vibrotactile actuators. Tested on LEAP Hand (real), Shadow/Inspire/Allegro (sim). On in-hand carton rotation: 10/10 with DOGlove vs. 1/10 with AnyTeleop (vision-only). IL benchmark: 85–90% success (3D Diffusion Policy). Force sensing resistors on robot fingertips provide the contact signal retargeted to the operator’s hand.

Category 3: Exoskeleton Systems (mechanical coupling + full-hand sensing)

DEXOP (MIT/Stanford, Sep 2025) Passive mechanical exoskeleton connecting human fingers to a separate passive robotic hand via 4-bar linkages. Three variants (12, 9, 7 DoF). Uses GelSim(ple) camera-based tactile sensors on fingertips and palm — whole-hand coverage. Key innovation: “perioperation” paradigm — the human directly manipulates via the exoskeleton, feeling the contact through proprioception without active actuation. Data collection speed: 11s vs. 86s per task compared to teleoperation. Co-designed DEXOP-7 + EyeSight Hand for zero kinematic retargeting.

DexUMI (Stanford/Columbia/CMU, May 2025) Wearable exoskeletons co-designed per robot hand, with FSR tactile sensors on Inspire Hand and electromagnetic sensors on XHand. 3.2× data collection efficiency vs. teleoperation. Video inpainting pipeline replaces human hand with robot hand in training observations — handles visual domain gap.

Glovity (arXiv:2510.09229, existing vault) Spatial wrench (force + torque) feedback teleoperation. Captures 6-DoF interaction forces at each fingertip rather than simple normal force, providing richer contact characterization for contact-rich manipulation tasks.

Sensor Technology Comparison

Device	Sensor Type	Coverage	EM Immunity	Cost	Open
OSMO	Magnetic 3-axis	12 pts (fingertips + palm)	Low	Open	Yes
TacCap	FBG optical	Fingertips only (360°)	High	~$10k (interrogator)	Yes
TAG	EOP (feedback)	Per-fingertip feedback	High	<$500	Partial
DOGlove	FSR (robot side)	5 fingertips	High	<$600	Yes
DEXOP	GelSim camera	Whole hand	High	Research	No
DexUMI	FSR / EM (robot side)	Fingertips	Medium	Research	No

Choosing a Device

For in-the-wild, scalable collection without a robot present: OSMO. No robot required; compatible with off-the-shelf hand tracking; open-source.

For maximum tactile fidelity with zero sensor domain gap: TacCap. Mount identical FBGs on both human and robot; get 100% grasp stability. Trade-off: bench-bound interrogator, fingertip-only coverage.

For teleoperation with operator haptic feedback: TAG or DOGlove. Both sub-$600. TAG offers higher tactile resolution (EOP) but is newer. DOGlove has more validated downstream IL results and tested on more robot hands.

For whole-hand contact capture including palm and proximal phalanges: DEXOP. Camera-based sensors provide the richest spatial coverage, but the system requires co-designing the exoskeleton with the target robot hand.

For maximum compatibility with multiple robot hands: DexUMI. Exoskeleton is optimized per-robot but the framework supports multiple platforms (Inspire Hand, XHand).

中文版

研究問題

哪些穿戴式觸覺感測裝置適合搭配靈巧機器手進行 LfD 示範資料收集？涵蓋商用與研究原型、感測技術比較與選型建議。

知識地圖

觸覺感測模態 — 磁性、FBG 光纖、相機、電阻式各有頻寬/覆蓋/成本/抗電磁干擾的取捨
遙操作 vs. 直接示範 — 兩種不同的資料收集正規性，影響裝置設計和管線
具身體差距（Embodiment Gap） — 人手與機器手形態不同，需要重定向（retargeting）
下游策略輸入 — 策略需要哪些模態決定了裝置必須捕捉什麼
擴散策略 — 接觸密集操作的主流策略架構，支援觸覺多模態輸入

關鍵發現

裝置分兩個方向：感測（記錄操作者觸覺→策略輸入）vs. 回饋（機器人觸覺→操作者感受）
「人手機器手用同一感測器」是最乾淨的遷移方案（TacCap 原則）
開源低成本裝置已達商用水準：DOGlove < $600 、 T A G <$ 500
操作者的觸覺回饋品質和記錄同樣重要：有回饋的操作者能收集更高品質的示範
原野採集（OSMO）vs. 結構化遙操作：前者規模更快，後者力訊號更豐富

A 類別商用方案補充

純 A 類（感測/記錄，無需機器人在場）的商用市場尚未成熟，多數商用產品以動作捕捉或觸覺回饋為主。以下是目前最接近的商用/準商用選項：

XELA Robotics uSkin

商用模組化 3D 磁性觸覺感測器，每個感測點 4×4 mm
可客製化為手套/皮膚陣列形式
主要定位為機器人手的感測皮膚（Tesollo DG-5F、LEAP 等），但提供 wearable 整合方案
優勢：商用支援、軟體生態（uAi Software）、高密度三軸感測
限制：不是現成手套產品，需整合工程；報價需聯絡廠商
網站：xelarobotics.com

Pressure Profile Systems TactileGlove II

現成商用觸覺手套：65 個感測元件，100 Hz，Bluetooth 5 無線
靈敏度：0.04 N/元件，量程 80 psi（55 N/cm²），S/N > 500:1
原本針對人因工程/製造安全設計，但技術規格符合 LfD 資料收集需求
限制：不針對機器人領域，無 ROS 整合、無 retargeting 管線；需自行開發軟體橋接
網站：pressureprofile.com

SenseGlove R1（偏向 B 類但有感測能力）

20 DoF 次毫米動作捕捉 + 內建力感測控制迴路
1 kHz 回饋頻率；設計用於遙操作機器人手
同時記錄操作者手指力道和動作，有 IL 資料集應用案例
研究顯示有回饋條件下資料品質提升 11%
限制：主要是 B 類產品，觸覺記錄非核心功能；價格較高
網站：senseglove.com

AnySkin（NYU/Meta，2024，開源準商用）

磁性頂針感測器，100 Hz，$10/個，12 秒可更換
關鍵特性：跨感測器實例泛化 — 在一個感測器上收集的資料可直接用於新的感測器實例，不需要重新校準
可掛在人手指尖，也可直接裝在 xArm/Franka/LEAP Hand
限制：每個感測器只覆蓋指尖正面，非 360°；無手腕/掌心覆蓋
網站：any-skin.github.io

SynTouch BioTac（較舊，仍在使用中）

流體填充指尖感測器，量測力、振動、溫度
廣泛用於學術研究（Shadow Hand 搭配使用），有大量公開資料集
限制：公司已縮減，新購困難；不是穿戴式而是機器手指尖形式

VTDexManip 架構（ICLR 2025，研究但使用商用零件）

客製化觸覺手套（20 個 FSR 感測器）+ HoloLens 2 ego-centric 視覺 + 商用 F/T 感測器校準
資料集：2032 筆操作序列，10 種日常任務，182 個物體，56.5 萬組視覺-觸覺配對
優勢：完整的多模態資料集，展示「低成本 FSR + 視覺」的可行性
限制：非現成商品，需自行搭建

市場現狀分析

產品	類型	商用?	觸覺感測點數	即插即用?
XELA uSkin	感測模組	是	高密度（客製）	需整合
TactileGlove II	完整手套	是	65 點	是（非機器人）
SenseGlove R1	遙操作手套	是	力感測（5 DoF）	是
AnySkin	指尖感測器	開源	1/指（需組裝）	接近
SynTouch BioTac	指尖感測器	是（有限）	1/指	需校準

核心結論：A 類別的「現成可買、直接用於 LfD」商用手套目前不存在。最接近的是 TactileGlove II（商用但非機器人定向）和 SenseGlove R1（機器人定向但以回饋為主）。研究實驗室通常選擇 AnySkin/uSkin 等感測器加自行整合，或直接採用 OSMO/TacCap 這類開源研究系統。

中國廠商補充（A 類別）

中國廠商的切入角度與歐美不同：歐美主流是「感測器戴在人手，直接示範（in-the-wild）」；中國主流是「觸覺皮膚裝在機器手端，配遙操作套件採集資料」。結果相同，但感測位置在機器人端。

戴盟機器人（Daimeng Robot）— 最接近 A 類別需求

DM-Tac 系列（W/W2/X/F）：多模態視觸覺感測終端
DM-EXton2：遙操作資料採集系統，配套 DM-Tac 使用
一條龍方案：感測器 + 遙操作採集 + VTLA 模型訓練
dmrobot.com

帕西尼（PaXini）— 高精度商用觸覺感測器

PX-6AX-GEN3：15 感測維度（含 6 軸力/紋理/彈性），0.01 N 解析度，$49 起
配套靈巧手（GMH18、DexH13）和人形機器人（TORA-ONE）
感測器定位，非手套產品；需自行整合至採集管線
paxini.com

汉威科技（Hanwei Technology）— 柔性觸覺感測器供應商

自研柔性微納感測器（壓力/壓電/應變/織物），四大系列七大產品
已與多家人形機器人廠商小批量供貨
定位：B2B 感測器供應，不含系統整合

灵巧智能（DexHand）— 遙操作資料採集套件

靈巧手（19 DoF）+ 遙操作套件 + 3D/VR 視覺 + 具身智能訓練平台
一站式方案，完整的感測 + 採集 + 訓練生態

悟機科技 Wuji Glove（2025，商用，配套 Wuji Hand）

硬體：觸覺矩陣（tactile matrix）+ 6 軸 IMU + 手指關節追蹤，網路連線（左手 192.168.1.100，右手 192.168.1.101）
軟體生態：Python SDK（串流/姿態/錄製）+ Wuji Studio（桌面 GUI）+ ROS2 Driver（1000Hz joint state publishing）
Vision Pro 整合：支援 Apple Vision Pro 手部追蹤做即時遙操作重定向
定位：配套 Wuji Hand 的一體化解決方案，A + B 類別兼顧（記錄 + 遙操作）
限制：觸覺矩陣密度/解析度規格未公開；與非 Wuji Hand 的機器手整合需額外工程
文件：docs.wuji.tech/docs/zh/wuji-glove/latest/

關鍵差異： 如果目標是無機器人在場的大規模原野採集（OSMO 路線），中國目前無對應現成產品，需自行整合帕西尼或汉威科技的感測器至可穿戴手套形式。Wuji Glove 是目前中國廠商中最接近「觸覺感測 + 遙操作整合」現成商用方案的產品。

未解問題

磁性手套長時間使用的感測器漂移問題
機器手形態差異大時觸覺訊號的跨具身遷移效果
下游策略需要的最低觸覺解析度閾值
TactileGlove II 是否可搭配 ROS 和 retargeting 管線使用（需實測）
中國路線（機器手端觸覺）vs. 歐美路線（人手端觸覺）哪種資料品質更高？

bot_vault

Explorer

靈巧手示範資料收集裝置：觸覺感測手套與穿戴式設備