Sensor evidence / Magnetic tactile skin
ReSkin: magnetic touch with a replaceable skin
Understand ReSkin magnetic tactile sensing: replaceable elastomer, five-magnetometer design, 400 Hz research setup, Python data collection, and calibration evidence.
Published 2026-09-12 | Updated 2026-09-12 by RoboSkin.ai Editorial Team
- 5
- sections
- 4
- questions
- 5
- next routes
Short answer
What you need to know
- 1
ReSkin is a magnetic tactile skin introduced by Bhirangi, Hellebrekers, Majidi, and Gupta at CoRL 2021. A passive magnetized elastomer deforms during contact, and nearby magnetometers measure changes in the magnetic field. The electronics are separate from the replaceable contact surface.
- 2
The original study describes a 20 × 20 mm sensing area and five three-axis magnetometers. Its electronics stream temperature plus three magnetic values per chip: 20 values altogether, of which 15 are magnetic channels. Force and contact location are learned from those measurements; they are not 20 direct force readings.
- 3
The paper investigates models that remain useful across skin instances and can adapt using self-supervision. This supports studying replaceable tactile interfaces, while the performance of a new mounting, material batch, or robot task still needs evaluation. Our review does not include fabrication or an independent reproduction.
Topic 01
The original hardware and signal path
The quantities below refer to the original ReSkin sources. Related magnetic skins and later interfaces should be identified separately rather than inheriting these specifications.
| Quantity | Source-reported configuration | Scope |
|---|---|---|
| Contact surface | Magnetized elastomer separated from the board | Paper, sensor design; permits replacement of the passive interface |
| Sensing area | 20 × 20 mm | Original paper, Section 3 |
| Magnetometers | Five MLX90393 chips; four around a central chip | Three magnetic axes per chip |
| Streamed values | Temperature, Bx, By, Bz for each of five chips | 20 values including temperature; 15 magnetic channels |
| Sampling | Approximately 400 Hz in the described board/setup | Acquisition rate is separate from inference and robot-control latency |
| Skin thickness | 2–3 mm in the Meta research announcement | Passive skin figure; not total integrated sensor thickness |
Topic 02
Collecting data with the released Python library
The project links the reskin_sensor repository, maintained by an original author. Its README documents the reskin_sensor package, a five-sensor board connected through a microcontroller, firmware-upload instructions, and a computer-side sensor test.
ReSkinBase is documented for blocking, standalone collection. ReSkinProcess supports background collection while other code runs. Choose based on the acquisition architecture, then verify timestamps and synchronization against the robot state and reference instrumentation.
For a repeatable experiment, preserve unloaded field readings, temperature channels, skin identity, mounting orientation, firmware version, sampling configuration, and the learned model. The released library handles access to observations; it does not by itself validate a force calibration or a manipulation policy.
Topic 03
What calibration and transfer results mean
ReSkin tackles variation across fabricated skins and changes with wear. The study trains response models using multiple skins and evaluates generalization and self-supervised adaptation. An unseen-skin result is meaningful only with its training skins, held-out skins, reference measurements, and adaptation procedure specified.
The original experimental setup fixes the board and skin on a printed mount, uses a robot-driven hemispherical indenter, and measures reference normal force with an ATI Nano17. The paper describes quasi-static measurements unless stated otherwise. These conditions differ from arbitrary sliding, impact, or full-hand manipulation.
The widely quoted 1 mm localization at 90% accuracy is a study summary, not a universal tolerance for every installed skin. Likewise, reported durability over many contacts depends on the interaction and material conditions. Preserve the evaluation protocol whenever quoting those figures.
Topic 04
Integration decisions before mounting a skin
A magnetic sensor depends on the relationship between the elastomer, board, and surrounding field. Fix the mounting geometry and characterize unloaded signals before and after replacement. Test whether nearby magnetic or moving metallic components affect the measurements in your intended setup.
Separate three checks: whether contact is detectable, whether the calibrated quantity is accurate, and whether using that quantity improves robot behavior. A high acquisition rate can help capture changes, but it does not establish a low-latency control loop or robust slip recovery.
Compared with an optical fingertip, ReSkin supplies sparse magnetic measurements rather than a dense contact image. Its replaceable surface can suit different collection geometries, while the choice should follow the signals and validation resources required by the task.
Topic 05
Source access and practical limits
The official project links the paper and an author-maintained repository containing software and hardware-related resources. The reviewed reskin_sensor repository carries an MIT license. Check licenses separately for any external data, models, or components you add.
The review confirms public documentation and repository access on September 12, 2026. It does not confirm current commercial stock, fabrication yield, or plug-and-play compatibility with an arbitrary gripper. Use the project sources to identify the exact board, firmware, and elastomer process before building.
Common questions
FAQ for this topic
Does ReSkin use a camera?
The original ReSkin uses magnetometers beneath a magnetized elastomer. Its raw observations are magnetic-field and temperature readings, rather than tactile camera images.
How many channels does the original board provide?
The paper describes five chips, each sending temperature and three magnetic components. That is 20 streamed values, including 15 magnetic channels. Derived contact or force estimates are separate outputs of a model.
Does replacing ReSkin require no calibration?
The research evaluates cross-skin models and self-supervised adaptation. It does not establish that every replacement on every robot preserves a calibrated response without checks.
Is 400 Hz the robot control rate?
Approximately 400 Hz is the sampling rate in the paper setup. Transport, inference, filtering, controller timing, and actuation add separate delays and must be measured for the application.