OSMO Tactile Glove

• OSMO tactile glove is an open-source wearable sensorized glove with 12 tactile sensors capturing high-bandwidth force signals during human demonstrations.
• It features a robust mechanical and electrical architecture with flexible sensor modules, integrated magnetometers, and low-noise calibration for precise force estimation.
• OSMO enables imitation learning by synchronizing multimodal data streams, allowing robotic policies to outperform vision-only methods in contact-maintenance tasks.

The OSMO tactile glove is an open-source, wearable sensorized platform developed for contact-rich human-to-robot skill transfer, with design and validation explicitly targeted toward in-the-wild imitation learning of dexterous manipulation tasks. Featuring 12 distributed, three-axis tactile sensors (“taxels”) across the fingertips and palm regions, OSMO natively captures high-bandwidth normal and shear force signals during human demonstrations, thereby minimizing the embodiment gap in visual and tactile signals when transferring skills to biomimetic hands mounted on robots. Training robotic policies solely on human demonstration data from the glove, researchers have shown that policies can execute complex contact-maintenance tasks (e.g., wiping) without access to robot-collected real-world data, outperforming vision-only methods through elimination of contact-regulation failure modes (Yin et al., 9 Dec 2025).

1. Mechanical and Electrical Architecture

The OSMO glove is constructed from a thin, stretchable fabric base designed to conform both to human hands and biomimetic robotic end-effectors such as the Psyonic Ability Hand. Twelve flexible tactile sensing modules are distributed across five fingertips and three distinct regions of the palm (distal, middle, proximal). Each taxel consists of a soft magnetic elastomer patch (EcoFlex 00-30 with MQFP-15-7 magnetic microparticles) directly adhered to a compact 12 mm×11 mm PCB. The PCB incorporates two 3-axis magnetometers (Bosch BMM350) and one 6-axis IMU (Bosch BHI360). To address cross-sensor noise, a 0.15 mm MuMetal shield is positioned between the fabric and PCB.

The sensors are cabled using tinsel wires in series along a shared I²C bus, terminating in a central STM32 microcontroller module located at the wrist. The full system is powered and communicates via a single USB-C cable, with 5 V at up to 1 A supplied and 3.3 V provided locally to sensors and controller via onboard regulators (Yin et al., 9 Dec 2025).

2. Sensing Principles and Performance

OSMO’s sensing principle leverages magnetic field perturbations induced by mechanical deformation of the elastomer, with magnetometers recording three-axis flux ($B_x,B_y,B_z$) in local PCB–fixed coordinate frames. Unlike optical and gel-based tactile systems, OSMO requires no imaging components. The operational measurement range is $0.3\,\mathrm{N} - 80\,\mathrm{N}$ (normal plus shear). Calibration reveals that $1\,\mathrm{N}$ input produces a $\sim 300\,\mu\mathrm{T}$ field response. Noise floor for detection is approximately $0.1\,\mu\mathrm{T}$, with crosstalk RMS noise reduced to $30–70\,\mu\mathrm{T}$ (finger motion) and $1–18\,\mu\mathrm{T}$ (adjacent contact) post-shielding, in contrast to $150\,\mu\mathrm{T}$ in unshielded configurations.

Differential sensing across paired magnetometers on each PCB ($\Delta B_i(t) = B_i^{(1)}(t) - B_i^{(2)}(t)$) further mitigates environmental interference and commutes the Earth's magnetic field. Each sensor is polled at up to 50 Hz, but system data for policy learning is streamed at 25 Hz via ROS2.

Force estimation is achieved through offline calibration applying known triaxial loads to extract a linear mapping:

$F_i(t) = C_i [\Delta B_i(t) - \Delta B_{0,i}]$

where $C_i \in \mathbb{R}^{3 \times 3}$ is a taxel-specific calibration matrix and $\Delta B_{0,i}$ indicates the rest (no-load) offset. Firmware employs a first-order low-pass filter ($\tau = 20\,\mathrm{ms}$) for initial denoising, complemented by Savitzky–Golay filtering in post-processing for training.

3. Firmware, Data Handling, and Synchronization

The STM32 firmware polls all magnetometers in round-robin at 50 Hz. Each packet includes: (i) $\Delta B_i \in \mathbb{R}^3$ per taxel, and (ii) raw $B_i^{(1)}$ and $B_i^{(2)}$ for calibration. Hardware time-stamping synchronizes the data streams, which are logged (or SD-card recorded) alongside video and stereo-IR frames within ROS2 at 25 Hz. Data acquisition also facilitates complete extrinsic calibration between RGB/IR cameras and the Franka robot world frame. Although the PCB-mounted IMU is connected, it is not used in current deployments but remains available for potential future fusion during pose estimation (Yin et al., 9 Dec 2025).

4. Human Demonstration Recording and Policy Learning Pipeline

The glove’s appearance and embedded profile are compatible with standard vision-based hand trackers (HaMeR, Dyn-HaMR, Aria Gen 2, Quest 3), and can be overlaid with an optically tracked Manus Quantum II glove without electromagnetic interference, due to the distinct operating modalities of the magnetometers ($\sim$kHz bandwidth) and the optical system.

During data collection, all modalities—RGB, stereo IR, tactile—are time-synchronized. Human demonstration datasets ($D^H$) comprise trajectories $T_0, \ldots, T_{N-1}$, where each time frame $F_k$ consists of:

• $I^H_{rgb} \in \mathbb{R}^{H\times W \times 3}$ (RGB image)
• $$I^H_{IR,left}, I^H_{IR,right} \in \mathbb{R}^{H\times W}$$ (stereo IR)
• $g^H \in \mathbb{R}^{3 \times 2 \times 5}$ (fingertip $\Delta B$ values; five fingertips, two magnetometers, three axes)

Robot demonstration datasets ($D^R$) pair each frame with an extracted proprioceptive state $q^R \in \mathbb{R}^{13}$ (7 DoF Franka arm, 6 DoF Ability Hand) via inverse kinematics.

Policy learning employs a diffusion policy architecture conditioned on RGB, proprioceptive, and tactile ($\Delta B$) streams. The visual encoder is frozen DINOv2; proprioceptive and tactile streams are encoded by dedicated two-layer MLPs. The combined conditioning vector is modulated through FiLM, controlling a U-Net–style diffusion model. The policy models $p(a_{0:15} | F_k)$, predicting 16 future joint commands, deploying the first four in each step. Loss is computed by standard DDPM $l_2$ distance over 100 denoising steps. Inputs are normalized per Barreiros et al. (2025), with

$$y_i = \mathrm{clip}\left(2 \frac{x_i – x^{0.02}}{x^{0.98} – x^{0.02}}, –1.5, 1.5\right)$$

5. Robot Policy Execution and Evaluation

At deployment, the robot (Franka + Ability Hand) is instrumented with its own OSMO glove, ensuring that tactile input format, coordinate frames, and calibration are consistent with the human demonstration data—directly addressing embodiment and domain gaps without recourse to vision-based force inference techniques. The diffusion policy generates target joint angles, which are executed by a low-level joint position controller (Yin et al., 9 Dec 2025).

For empirical evaluation, the principal task involves wiping marker patterns from a whiteboard using a sponge affixed to the robotic fingertips. The success metric is the proportion of marker pixels erased after 90 s, as measured by binarized vision segmentation. Comparative experiments under matched conditions yield the following (mean ± std, $n=12$ rollouts, two stencil patterns):

Input Modalities	Success Rate (%)
Proprio only	$27.1 \pm 32.4$
Vision + proprio	$55.8 \pm 30.0$
Tactile + vision + proprio	$71.7 \pm 27.4$

The tactile-aware policy delivers a $\sim 72\%$ average success, exceeding vision-proprioceptive baselines by more than $16$ percentage points. Failure analysis identifies that, without tactile feedback, “light touch” and contact loss routinely cause incomplete wiping or object drops. In contrast, tactile-equipped execution achieves automatic pressure regulation; remaining error is attributed to kinematic drift and hand-pose estimation inaccuracies.

6. Open-Source Resources and Community Impact

The OSMO project provides extensive open-source artifacts to facilitate adoption and reproducibility:

• Hardware: Full CAD designs for the glove, sensor PCBs (Gerbers), MuMetal shields, and comprehensive bill of materials (magnetometer/IMU sourcing and elastomer formulation).
• Firmware: STM32Cube codebase, firmware routines for differential sensing, calibration, and ROS2/USB drivers.
• Assembly: Layered stepwise sensor instructions, wiring diagrams, and sewing patterns.
• Software: Python/ROS2 pipelines for multimodal synchronized data capture, hand-pose postprocessing (SAM2+HaMeR+stereo), IK retargeting (Mink), dataset encoding.
• Policy: Open training scripts, hyperparameter settings, pretrained DINOv2 weights, and quantitative evaluation notebooks.
• Datasets: Approximately 140 human demonstrations (≈2 hours) with RGB, stereo IR, tactile, and retargeted proprioceptive data, structured for direct use in imitation learning.

All documentation, fabrication files, and experimental data are available under permissive open licenses at https://jessicayin.github.io/osmo_tactile_glove/ (Yin et al., 9 Dec 2025). This enables replicability across research labs and paves the way for rapidly scaling in-the-wild, contact-rich demonstration corpora for robot skill acquisition.

7. Significance and Prospective Directions

By unifying high-fidelity tactile signal acquisition with robust hand-tracking compatibility and an imitation learning policy framework, OSMO addresses a critical limitation in video-based skill transfer: the absence of contact-specific feedback essential for manipulation. Direct tactile parity between human and robotic demonstrations eliminates the need for force-inference or domain adaptation. A plausible implication is that OSMO’s design supports future extensions to unsupervised domain adaptation, more complex manipulation tasks, and large-scale collection of human tactile behavior patterns, catalyzing tactile-aware policy development at scale (Yin et al., 9 Dec 2025).