本文由 AI 分析生成
建立時間： 2026-03-25 來源： https://do-glove.github.io/
Summary
DOGlove (Tsinghua/Shanghai AI Lab) is a <$600 open-source wearable glove with 21-DoF motion capture and 5-DoF haptic force feedback, enabling precise dexterous hand teleoperation. It enables manipulation without visual feedback by providing tactile sensing, and demonstrations collected via DOGlove are used to train effective imitation learning policies.
DOGlove 是一款成本低於 600 美元的開源穿戴手套，具有 21 自由度動作捕捉和 5 自由度觸覺力回饋，能夠實現精確的靈巧手遠端操作，甚至可在無視覺回饋的情況下操作，並用所收集示範訓練模仿學習策略。
Prerequisites
- Dexterous hand teleoperation — DOGlove controls dexterous robot hands (Shadow Hand, LEAP Hand); understanding their DoF constraints is needed
- Force/haptic feedback — the glove’s key innovation is bidirectional sensing; haptic systems are different from visual feedback
- Imitation learning — collected demonstrations feed IL pipelines; behavior cloning from haptic-enriched demos is the downstream application
Core Idea
Current dexterous teleoperation relies on expensive MoCap systems or vision-only tracking that lacks force perception, making precise contact-rich tasks difficult. DOGlove solves this with a cable-driven torque transmission mechanism for force feedback and linear resonant actuators for fingertip haptic response. The retargeting framework maps human hand pose to diverse robot hand morphologies. Crucially, operators can feel contact forces during teleoperation — enabling tasks like regulating sauce flow or identifying objects by touch alone without visual information.
Results
| Task | Capability |
|---|---|
| Blind bottle grasping | Successful with haptic feedback only |
| Regulating condensed milk flow | Force-controlled via haptic feedback |
| In-hand rotation | Friction-based using force feedback |
| IL policy training | High success rates from collected demos |
Limitations
- Author-stated: 5-DoF force feedback (not full 21-DoF); covers finger-level but not palm/wrist forces fully
- Author-stated: System validated on specific robot hands; retargeting quality varies with morphological differences
- Unstated: Sub-$600 hardware might have durability issues for large-scale data collection campaigns
Reproducibility
- Code: Fully open-sourced at https://do-glove.github.io/
- Datasets: Task demonstrations collected with DOGlove
- Compute: Standard IL training; main investment is hardware assembly
Insights
The blind manipulation demonstrations (grasping without visual feedback) are the paper’s most compelling evidence — they show that haptic feedback provides genuinely irreplaceable information that vision cannot substitute. The assembly-in-hours / <$600 price point is a significant democratization compared to LEAP Hand + MoCap setups that previously cost 10-50x more. This may enable smaller labs and universities to run dexterous manipulation experiments.
Connections
Raw Excerpt
DOGlove can be assembled in hours at a cost under 600 USD. It features a customized joint structure for 21-DoF motion capture, a compact cable-driven torque transmission mechanism for 5-DoF multidirectional force feedback.