ReSkin: versatile, replaceable, lasting tactile skins

1 Introduction

• Existing tactile solutions do not meet all practical needs, including compactness, low cost, longevity, and large-area coverage.The introduction cites bulky, expensive, or slow vision-based sensors, connection-heavy resistive and capacitive sensors, costly limited-form-factor commercial options, and rigid sensors without a deformable surface.Source: §1 · p. 1
• Soft tactile skins face a trade-off between conformal contact and durability, while models trained on one sensor generally fail to transfer to another.Replacing a skin can therefore require relearning, making experiments with a single sensor more practical than replacement.Source: §1 · p. 1
• ReSkin is a compact, inexpensive, replaceable magnetic tactile skin whose learned models estimate contact location and force from magnetic-field changes.The proposed sensor is under $30 and 2–3 mm thick, with reported operation beyond 50K interactions, up to 400 Hz temporal resolution, and 1 mm spatial resolution at 90% accuracy.Source: §1 · p. 2

2 Background

• Soft-sensor models have been used for grasping, proprioception, and object classification, while unpredictable material responses motivate data-driven approaches.Examples include resistive knit gloves, scalable capacitive skins, and multimodal skins.Source: §2 · pp. 2–3
• Magnetic sensing separates the elastomer from its electronics, facilitating replacement and compact, diverse form factors compared with camera-based optical sensing.The paper demonstrates magnetic skins on an arm sleeve, glove, dog shoe, and robot end-effector.Source: §2 · p. 3
• ReSkin combines multi-sensor training with a self-supervised calibration procedure using a couple of hundred pokes to adapt models to particular sensors.The paper reports that single-sensor models fail to generalize to other sensors or circuit-board designs, while multi-sensor training improves generalization but remains below same-sensor training.Source: §2 · p. 3

3 Design and Fabrication

• ReSkin maps deformation-induced magnetic-field changes to contact characteristics using embedded magnetic microparticles and a nearby magnetometer.The composite can be molded into many shapes and thicknesses.Source: §3 · pp. 3–4
• The sensing area is 20 mm x 20 mm, measured with five magnetometers: four MLX90393 sensors spaced 7 mm apart around a central magnetometer.The molds, circuit board files, bill of materials, and libraries were publicly released and open-sourced on the website.Source: §3 · p. 4
• The elastomer is made by mixing magnetic microparticles and two-part polymer, vacuuming out bubbles, and curing the sample at room temperature for at least three hours.The mold uses a 4x4 grid of cube magnets above and below the sample during curing.Source: §3 · p. 4
• Rigid and flexible circuit-board variants attach replaceable skins with screws or sewable snaps rather than permanent adhesive.The boards transmit measurements from five magnetometers through a microcontroller to a computer.Source: §3 · p. 4

4 Experimental Setup

• A Dobot Magician applies indentations while five magnetometers stream readings at 400 Hz and an ATI Nano17 force sensor records force at 1 kHz.The skin and circuit board are fixed to a 3D-printed mount.Source: §4 · p. 4
• The experiments use quasi-static contacts, labeling location with the robot's specified xy position and force magnitude with the Nano17 normal-force measurement.The world xy-plane is aligned with the robot base.Source: §4 · p. 4
• Indentations follow a snake-like path over 65 points on a 9x9 grid, with six depths from 0.2 mm to 1.2 mm for 390 indentations per iteration.Data are collected over multiple iterations.Source: §4 · pp. 4–5

5 Single Sensor Model – Decoding Magnetic Flux to Contact Characteristics

• 99.58 ± 0.34% contact localization accuracy, with location MSE of 0.037 ± 0.014 mm2 and force MSE of 0.005 ± 0.002 N2, was achieved on a random held-out test set.The model was trained on 45K of 50K samples.Source: §5 · p. 5
• 0.0011±0.0002 N2 Fxy MSE was achieved in dynamic shear sensing, alongside 0.003 ± 0.001 N2 normal-force MSE and 0.085 ± 0.006 mm2 contact-location MSE.The skin was indented and dragged along x and y directions; the model predicted (x, y, Fx, Fy, Fz).Source: §5 · p. 5
• More frequent no-load measurements can offset drift-related prediction-error growth, while a single initial measurement produces higher, faster-growing errors.Errors are also higher when no-load measurements are updated before every contact, possibly because of overfitting; later experiments benefit from updates just before contact.Source: §5 · pp. 5–6, Figure 5 · p. 6
• As interactions increase, the elastomer softens and the model overestimates applied force on average.The study evaluated a model trained on the first 5,000 indentations against later test sets, using measurements at different intervals.Source: §5 · pp. 5–6

6 Adapting to New Sensors – MultiSensor Model + Self-supervised Learning

• Multi-sensor training exposes the model to more diverse data, while feature-space triplet loss encourages nearby skin locations to have nearby representations.The triplet loss acts as a regularizer and supports self-supervised adaptation.Source: §6 · p. 6
• The self-supervised adaptation uses unlabeled, indexed indentations to form distance-ordered triplets, then fine-tunes the multi-sensor model alongside batches from the original labeled training data.The triplet loss does not require ground-truth contact locations or force readings.Source: §6 · p. 6

6.1 Results

• Multi-sensor models significantly outperform the single-sensor baseline, and self-supervised adaptation further improves prediction accuracy and force prediction.The single-sensor baseline fails to capture variability across sensors.Source: Table 2 · p. 7, §6.1 · p. 7
• Self-supervised adaptation improves performance with little adaptation data, plateaus beyond 780 points, and offers larger gains when fewer sensors train the model.It performs better than the other models across the tested training-sensor quantities.Source: §6.1 · p. 7, Figure 6 · p. 7
• After 325 manual indentations, accuracy rose from 79.86% to 84.84%, MSExy fell from 0.676 mm2 to 0.489 mm2, and MSEF fell from 0.268 N2 to 0.192 N2.The experiment was less controlled and closer to an end-user application setting.Source: §6.1 · p. 7
• On a flexible board adapted from rigid-board training, average contact localization accuracy was 75%, with location MSE of 0.72 mm2 and force MSE of 0.54 N2.The board change was harder because the skin-to-board distance was 80% smaller; the authors attribute larger force errors to overestimation from stronger signals.Source: §6.1 · p. 7

Force Sensitivity: Water in shot glass

• Less than 20 mL of water (< 0.2N) produces monotonically increasing measurements, illustrating ReSkin’s sensitivity to small forces.The demonstration uses a shot glass placed on the sensor as water is added.Source: §7 · p. 8

Robot Gripper

• Grasping blueberries and grapes works with ReSkin force feedback where the gripper’s built-in force sensing (30N minimum) is insufficient, and continues after skin replacement without tuning.Two sensors with flexible circuit boards are placed on either side of a parallel jaw gripper.Source: §7 · p. 8

Location Sensitivity: Poking

• Real-time contact-location estimation is demonstrated by poking a new sensor.The demonstration illustrates location sensitivity.Source: §7 · p. 8

Dog Shoe

• A sensor in a dog shoe tracks force magnitude and direction while a 17 lb dog rests, walks, and runs.The shoe uses a 1/16in urethane foam layer for comfort and logs data at 250 Hz.Source: §7 · p. 8

Glove

• A magnetic skin on an index finger records sensor output during dough sealing, demonstrating use during human-object interaction.Data is logged on-board at 250 Hz.Source: §7 · p. 8

Arm Sleeve

• Eight flexible boards form a 2in x 4in continuous skin, demonstrating contact localization across a larger surface area.The boards connect to a microcontroller sampling 40 magnetometers at 133 Hz.Source: §7 · p. 8

• ReSkin combines multi-sensor learning, self-supervised triplet loss, and an adaptation procedure to refine models for new skins.The conclusion describes replaceable skins as usable right away and likens replacement to peeling and applying a band-aid.Source: §8 · p. 8
• The authors demonstrate a compact sensor for diverse applications, including delicate-object grasping, dog-foot force measurement, and wide-coverage tactile sensing.Source: §8 · p. 8
• Experiments cover single-point contact; multi-point contact remains a future direction.The authors also identify external magnetic fields and metallic objects, and dynamic time-series modeling, as topics for further study.Source: §8 · p. 8