本文由 AI 分析生成
建立時間： 2026-04-05 來源： https://arxiv.org/html/2503.01789v1
Summary
TacCap is a thimble-form-factor wearable tactile sensor using Fiber Bragg Grating (FBG) optical fiber technology, designed to capture 360-degree contact information around human fingertips during manipulation demonstrations. The same sensor hardware is mounted on both human and robot fingers, eliminating the sensor domain gap during skill transfer. On grasp stability tasks, TacCap achieves up to 100% success vs. 10–80% without tactile data and outperforms GelSight DIGIT.
TacCap 是一款頂針形態的可穿戴觸覺感測器，採用光纖布拉格光柵（FBG）技術，提供 360° 手指觸覺覆蓋。關鍵設計是人手與機器手使用完全相同的感測器，消除感測器域偏差。抓握穩定性任務成功率高達 100%，優於 GelSight DIGIT。
Prerequisites
- Fiber Bragg Grating (FBG) sensing — FBG sensors measure strain via wavelength shifts in optical fiber; this is the core transduction mechanism and explains the sensor’s high sampling rate and immunity to electromagnetic interference.
- Tactile sensor calibration — contact position prediction from wavelength signals requires systematic ground-truth mapping; understanding the calibration pipeline is necessary to assess practical deployment cost.
- Imitation learning with multimodal input — the paper’s downstream task is training robot policies on tactile+proprioception data; understanding what contact events matter for grasp success motivates the sensor design.
Core Idea
The central insight is that the hardest problem in tactile-based skill transfer is not capturing contact data on the human side but ensuring that the signal produced is identical to what the robot would produce in the same contact configuration. By using a thimble-mounted identical sensor on both human and robot fingertips, TacCap sidesteps the cross-embodiment tactile retargeting problem entirely. FBG technology is chosen over capacitive or optical tactile sensors because it is inherently immune to electromagnetic interference (pure optical signal path), lightweight, and capable of 2 kHz sampling — high enough to capture transient slip events. The 360-degree coverage (vs. GelSight’s front-only view) is critical because natural grasping involves lateral and dorsal contacts that vision-based sensors miss.
Results
| Condition | Success rate |
|---|---|
| TacCap (teleoperation) | 100% |
| TacCap (wearable) | 90% |
| GelSight DIGIT | 40–75% |
| No tactile | 10–80% (task-dependent) |
Limitations
- Author-stated: Interrogator hardware (Micron Optics sm130) is expensive and bench-bound; not suitable for fully wearable mobile data collection.
- Unstated: Thimble form factor covers only fingertips — palm, proximal phalanges, and thumb base have no coverage. FBG interrogators add cost and complexity vs. purely electrical sensor approaches.
Reproducibility
- Code/Hardware: Open-source at sites.google.com/stanford.edu/taccap
- Datasets: Custom grasp experiments with Rokoko mocap glove; dataset not publicly released
- Compute: Contact prediction model is lightweight (encoder-decoder); training and inference both feasible on single GPU
Insights
TacCap represents an important design philosophy: instead of trying to retarget tactile signals across different sensor modalities, eliminate the retargeting problem by making the human and robot sensors identical. This is analogous to how DexUMI eliminates kinematic retargeting by co-designing the exoskeleton with the target robot hand. The FBG approach has a significant latent advantage for environments with strong EM fields (e.g., near robot motors) where capacitive or magnetic sensors degrade.
Connections
- Clippings-osmo-open-source-tactile-glove-human-to-robot-skill-transfer — complementary: OSMO uses magnetic sensors for broader glove coverage; TacCap uses FBG for fingertip precision
- Clippings-doglove-dexterous-manipulation-with-a-low-cost-open-source-haptic-force-feedback — DOGlove uses FSR sensors on robot fingertips for haptic feedback; TacCap solves the same interface but for data collection
- Clippings-glovity-learning-dexterous-contact-rich-manipulation-via-spatial-wrench-feedback — spatial wrench feedback for teleoperation, complementary sensing approach
Raw Excerpt
“Touch on the sides is very common in natural grasping — 360-degree sensing around the fingertip proved critical, as GelSight DIGIT’s front-only coverage missed the majority of contact events.”