TacCap: A Wearable FBG-Based Tactile Sensor for Efficient Human-to-Robot Skill Transfer

TacCap is based on fiber Bragg grating (FBG) technology, which is inherently resistant to environmental factors such as light, water, and magnetic fields. FBGs function as highly sensitive optical strain gauges, capable of resolving strains as small as 10^-5. When strain is applied to the FBGs, it induces shifts in the reflected optical wavelengths, as shown in the figure below. Additionally, FBG technology allows multiple sensors to be integrated along a single optical fiber, each tuned to a different nominal wavelength. This configuration simplifies wiring while enabling the system to sample dozens of FBGs at rates of up to 2 kHz.

Finite element analysis (FEA) of a single‑point contact shows that strain is greatest near the applied force and decays with distance, confirming the locality of the sensor response. Guided by these results, we adopt a pseudo‑uniform spiral layout in which successive FBGs are evenly spaced along the axial direction.

We employ a three‑layer construction: an inner stiff layer that prevents finger squeezing of the TacCap, a middle layer that sensing strain with FBGs, and an outer compliant layer. This architecture yields a minimal human‑to‑robot transfer gap, as illustrated by the TacCap thumb signals shown below.

Since the raw signal consists of FBG wavelength data representing strain at specific locations on the sensor structure, TacCap is able to extract higher level information such as the contact position. We design a the calibration system that automaticly presses the sensor at different locations to collect training data. Then a contact prediction algorithm is trained to map the raw FBG wavelength data to the contact position.

In this section, we conduct a series of experiments to evaluate the performance of the TacCap sensor, with a particular focus on its effectiveness in human-to-robot demonstration transfer. In the following image, we demonstrate our setup: (a) we place the TacCap sensor on the Leap hand for robot execution and teleoperation data collection, and (b) we wear the TacCap sensor for data collection without the robot. We evaluate our model with grasp stability (without vision) and skill transfer (with vision). The first focuses only on the TacCap sensor, without being affected by the vision domain gap, while the second focuses on real-world performance.

Our experiments show that TacCap markedly improves grasp stability and skill transfer in both wearable and teleoperation settings. In the wearable setup, TacCap yields a higher success rate than configurations without tactile sensing by compensating for hand pose estimation errors and enabling a firmer grasp. In teleoperation, TacCap likewise produces more reliable, robust grasps.

When evaluating skill transfer with visual feedback, we observe a performance drop when training solely on wearable data due to visual domain differences. However, supplementing with a small amount of teleoperation data effectively bridges this gap and restores performance. Comparisons with the DIGIT sensor highlight TacCap's advantage: its ability to sense contact around the entire fingertip provides richer tactile information, while its lower-dimensional, force-correlated signals are easier for learning algorithms to utilize. In contrast, the DIGIT sensor's high-dimensional data and limited sensing area can hinder performance, and its hardware is not easily adapted for wearable use.

TacCap does not currently support force measurement or multi-contact detection. With additional FBGs and a more comprehensive calibration process, these capabilities may be achievable.

The three-layer design can feel encumbering for the user, and it likely reduces the ability to perceive fine object details while wearing the sensor.

Temperature changes affect the FBG signal. We observe signal drift for roughly the first five minutes after wearing TacCap; all reported data were collected after stabilization. In future work, we plan to add a temperature-compensation FBG to mitigate this effect.

The interrogator remains expensive: the fiber costs ~$150 and the interrogator ~$5k. Portable interrogators exist but currently require customization.