本文由 AI 分析生成
建立時間： 2026-03-29 來源： https://arxiv.org/html/2510.09229v1
Summary
Glovity is an open-source, sub-$300 wearable teleoperation system that provides spatial wrench feedback (force/torque on the wrist) and fingertip haptic feedback for dexterous manipulation. The system integrates IMU-based hand tracking with Hall sensor calibration to close the embodiment gap. Demonstrations collected with Glovity are used to train diffusion-based imitation learning policies that incorporate 6D wrench signals as observations, achieving 80% success on contact-rich page-flipping tasks.
Glovity 是一個低於 $300 美元的可穿戴遙操作系統，提供手腕端的 6D 力/扭矩回饋與指尖觸覺回饋。用於收集 dexterous manipulation 示範資料，並將 wrench 信號整合進 diffusion policy，在頁面翻轉任務達到 80% 成功率。
Prerequisites
- Teleoperation data collection for imitation learning — 理解為何人類示範品質直接影響策略學習效果，是評估 Glovity 系統價值的前提
- Diffusion Policy / behavior cloning — Glovity 的策略訓練基於 DP-R3M（Diffusion Policy + R3M 視覺編碼），需了解行為克隆的基本框架
- 6-DoF force-torque sensing — wrench feedback 的設計基於 6D F/T sensor 輸出，理解力/扭矩的物理意義有助於評估 feedback mapping 設計
Core Idea
現有遙操作系統缺乏多模態回饋，操作員只能靠視覺判斷接觸力，導致接觸密集任務（contact-rich tasks）的示範品質差。Glovity 的核心設計是用4個小型伺服馬達與雙連桿機構將機器人末端的 6D wrench 即時對映到操作員手腕上的機械回饋，讓人直觀感受力的大小與方向，而不需要昂貴的外骨骼。手套側則用 IMU + Hall sensor 組合解決大拇指與食指捏取的精確校準問題，整套系統模組化可分離使用。
Results
| Task | With Wrench Feedback | Without Wrench Feedback |
|---|---|---|
| Page-flipping success rate | 39/50 (78%) | 24/50 (48%) |
| Mean task time (s) | 11.8 | 17.7 |
| Failures (excessive force) | 4 | 11 |
| Imitation learning (page-flip) | 16/20 (80%) | 9/20 (45%) |
| Imitation learning (hand-over) | 15/20 (75%) | 0/20 (0%) |
| Thin-object grasping (envelope) | 41/50 vs AnyTeleop 44/50 | Vrtrix Glove 17/50 |
Limitations
- Author-stated: wrench 回饋區域有限（只覆蓋手腕中段），無法準確區分手腕與指尖附近的接觸；手套目前只有6DoF，複雜手部操作受限；wrench 整合進 diffusion model 的方式較簡單，不一定泛化到長時域任務
- Unstated: 10位受試者樣本量小，外部有效性存疑；$300 成本假設所有元件可取得（XL330 servo 在某些地區難購得）；force-torque sensor (FT300s) 安裝在機器人末端，在真實部署中這是固定成本假設
Reproducibility
- Code: 承諾開源（論文撰寫時仍為預發布狀態）
- Datasets: 自行收集，30-50個示範/任務
- Compute: 單 GPU 即可（DP-R3M 訓練 300 epochs，規模小）；硬體成本 ~230 wrench + $70 手套）
Insights
「$300 開源遙操作系統」的意義在於降低 data collection 門檻。高品質接觸示範資料長期是 dexterous manipulation 的瓶頸，DOGlove、DexCap 等系統都在解同一問題。Glovity 選擇 wrench（端部力矩）而非手指觸覺作為主要回饋模態，這個設計決策反映了：操作員感受「我是否在用正確的力氣」比感受「哪根手指在接觸」更關鍵。
Connections
- Clippings-doglove-dexterous-manipulation-with-a-low-cost-open-source-haptic-force-feedback — 同類低成本觸覺手套，採用外骨骼+cable-driven力回饋方案
- Clippings-anyteleop-vision-based-dexterous-teleoperation — 視覺為主的遙操作系統，Glovity 在 thin-object grasping 中以此為基準線
- Clippings-dexterous-manipulation-imitation-learning-survey — 更廣泛的 dexterous imitation learning 文獻脈絡
- Clippings-touch-in-the-wild-learning-fine-grained-manipulation-with-a-portable-visuo-tacti — 類似的可攜式多模態抓取資料收集研究
- teleoperation
- imitation-learning
- haptics
Raw Excerpt
“Glovity combines a wearable wrench feedback mechanism, which provides intuitive force and torque cues without restricting arm mobility, with a haptic glove equipped with fingertip Hall calibration for precise grasping. Costing less than $300 and utilizing 3D-printed components, Glovity is accessible and replicable, requiring only a few hours for assembly.”