arXiv: https://arxiv.org/abs/2503.01789 | cs.RO | 2025-03-03
Abstract
TacCap is a lightweight wearable fingertip sensor utilizing Fiber Bragg Grating (FBG) technology to capture tactile feedback during human manipulation demonstrations. The system bridges the sensorimotor gap between human and robotic execution by mounting identical sensors on both operators’ and robots’ fingertips, enabling effective skill transfer through imitation learning with tactile data.
Hardware Specifications
Sensor Architecture
- Single optical fiber (Corning Ultra SMF-28, 0.25mm diameter) embedded in 3D-printed structure
- Three-layer design: rigid PA6-CF inner layer, compliant Rigid 4K middle layer, rubber outer layer
- Thimble form factor (cylindrical body with dome-shaped top, minimum inner radius 8mm)
- Multiple FBGs with nominal wavelengths 1525–1565 nm
Performance Metrics
- Sampling rate: 2 kHz (via Micron Optics sm130 interrogator)
- Minimum detectable force: 0.028 N
- Rise time: 87 ms; Fall time: 92 ms
- Contact position accuracy: ~5.3–5.8 mm
Data Collection Pipeline
Calibration uses a three-axis linear stage with hemispherical probe and rotational stage for systematic contact mapping. Human wears TacCap thimbles on fingertips + Rokoko motion capture glove for hand pose; robot uses identical TacCap sensors. Contact prediction model: encoder-decoder with multi-task learning (contact position regression + binary presence detection) from FBG wavelength signals.
Key Results
Grasp Stability Performance
- TacCap (wearable): 90% success rate
- TacCap (teleoperation): 100% success rate
- No tactile data: 10–80% success (task-dependent)
- GelSight DIGIT comparison: 40% vs. TacCap’s 75–100%
Key Advantage
360-degree sensing around fingertip vs. DIGIT’s front-only coverage — critical since side contacts are common in natural grasping.
Open Source
Hardware and software: sites.google.com/stanford.edu/taccap