Human video demonstrations provide abundant training data for learning robot policies, but video alone cannot capture the rich contact signals critical for mastering manipulation. We introduce OSMO: the Open Source tactile glove for huMan-to-robOt skill transfer. The glove features 12 three-axis tactile sensors across the fingertips and palm and is designed to be compatible with SOTA hand-tracking methods for in-the-wild data collection.
We demonstrate that a robot policy trained exclusively on human demonstrations collected with OSMO, without any real robot data, is capable of executing a challenging contact-rich manipulation task. By equipping both the human and the robot with the same glove, OSMO minimizes the visual and tactile embodiment gap, enabling the transfer of continuous shear and normal force feedback while avoiding the need for image inpainting or other vision-based force inference. On a real-world wiping task requiring sustained contact pressure, our tactile-aware policy achieves a 72% success rate, outperforming vision-only baselines by eliminating contact-related failure modes. We release complete hardware designs, firmware, and assembly instructions to support community adoption.
When the soft magnetic elastomers are deformed, the magnetometers measure changes in magnetic flux, which can be correlated to applied XYZ forces.
The 12 sensor PCBs communicate with an STM32 microcontroller via I2C, enabling modular addition or removal of sensors.
Data streams to a host computer via USB-C, with native Python and ROS2 interfaces supporting time-synchronized sensor packets.
The complete glove requires only USB-C for power and data. Alternatively, the microcontroller also accomodates a small LiPo battery and a microSD card slot for wireless operation.
The OSMO glove provide tactile coverage on both fingertips and palm while omitting the intermediate and proximal phalanges. This design choice accommodates varying hand sizes by allowing flexible finger lengths while ensuring consistent sensor placement at the primary contact regions.
The palm sensors are arranged in three sections aligned with the major planar regions of palm motion.
An outer beige glove provides uniform appearance and protects the electronics.
To minimize the embodiment gap, we design the OSMO glove with a stretchable base layer and sensor layout that can be worn by both human demonstrators and biomimetic robot hands, such as the Psyonic Ability Hand. A shared glove platform can help minimize the gap in tactile and visual feedback at data collection and deployment time.
The OSMO tactile glove features 12 three-axis magnetic tactile sensors across the fingertips and palm. We directly use the raw magnetic flux signals (μT) as tactile data.
The OSMO tactile glove is designed to be compatible with off-the-shelf hand-tracking devices for in-the-wild data collection. OSMO is worn on the right hand, and the device's native hand-tracking estimates are overlaid. The glove works seamlessly with:
We demonstrate that a robot policy trained exclusively on human demonstrations collected with OSMO, without any real robot data, is capable of executing a challenging contact-rich manipulation task. On a real-world wiping task requiring sustained contact pressure, our tactile-aware policy achieves a 72% success rate, outperforming vision-only baselines by eliminating contact-related failure modes.
We collect approximately 2 hours (140 demos) of OSMO glove fingertip tactile data and Realsense RGB + IR videos of wiping a whiteboard.
The estimated hand keypoints from HaMeR are overlaid on the RGB video. We omit the palm tactile data for this task because it is not in contact during demonstrations.
We equip the robot with the same OSMO glove and rollout the policy learned from human demonstrations.
The OSMO tactile glove provides two key advantages that typically require substantial preprocessing.
First, using the same glove on both the human and the robot greatly reduces the visual domain gap, enabling the use of an off-the-shelf vision encoder without the image editing, inpainting, or hand-masking steps commonly required to reconcile differences between human and robot embodiments.
Second, the glove enables rich, continuous force transfer directly from human to robot: the robot policy utilizes continuous shear and normal tactile signals at deployment time, despite being trained entirely on human tactile data.
Policies without tactile feedback in training and deployment frequently exhibit contact-related failures: inconsistent or insufficient applied pressure on the sponge or losing grasp of the sponge. The elimination of contact-related failures and improved success rate show that the OSMO tactile glove feedback is beneficial for this contact-rich task.
Scaling from single sensor implementations to 12 closely-spaced magnetic sensors introduces significant crosstalk. We introduce two new techniques to our sensor design for crosstalk mitigation:
Our combined approach reduces average noise by 57% compared to a single magnetometer configuration.
@article{yin2025osmo,
title={OSMO: Open-Source Tactile Glove for Human-to-Robot Skill Transfer},
author={Jessica Yin and Haozhi Qi and Youngsun Wi and Sayantan Kundu and Mike Lambeta and William Yang and Changhao Wang and Tingfan Wu and Jitendra Malik and Tess Hellebrekers},
journal={arXiv:2512.08920},
year={2025}
}
For researchers and developers interested in reproducing or extending this work, detailed technical information is available below:
All source code for the OSMO Tactile Glove is available in the repository:
PCB files for the OSMO Tactile Glove:
Step-by-step guide to building the OSMO Tactile Glove: