← News
Tech for Good Tech for Good Frontiers

A tactile skin robots can redesign for themselves

A tactile skin robots can redesign for themselves
1,024 Taxels in one skin
16 Sampled taxels per update
4 robot types tested
More Than 30 custom sensor designs
19.2X faster partial readout

A robot gripping a slippery object may need to know not just that it is touching something, but which way the object is starting to slide. OpenTac’s sensor skin is built to detect both: in a loaded grasping test, a robot using its three-dimensional tactile feedback held the object in 16 of 20 trials, compared with 9 of 20 using either a force/torque sensor or OpenTac’s normal-force-only mode. The result is an unusually concrete promise for tactile robotics: better contact sensing can turn a grip that often fails under load into one that succeeds more often.

The bigger idea is that the sensing surface itself can be redesigned for the robot. OpenTac combines a flexible circuit, pressure-sensitive film and a patterned, deformable cover with software that automatically turns a chosen shape into fabrication files. The team reports fabricating and testing 31 sensor layouts, from fingertips and gripper pads to a robot foot. If the approach proves robust beyond the paper’s demonstrations, it could make tactile sensing less like a bespoke research project and more like a component that can be fitted to many kinds of machines.

Figure 1: Tactile samples recorded with OpenTac skins. Heatmaps indicate normal load; cyan arrows denote per-bump shear; white arcs denote rotation estimates. (a) 64×64 skin, twisting a dome and clippers. (b) 9×9 fingertip, opposite thumb twists. (c) UMI gripper, 29×6 finger skins: can and ball rocked in the grasp. (d) Extrinsic contact: screwdriver moved down and up.
Figure 1: Tactile samples recorded with OpenTac skins. Heatmaps indicate normal load; cyan arrows denote per-bump shear; white arcs denote rotation estimates. (a) 64×64 skin, twisting a dome and clippers. (b) 9×9 fingertip, opposite thumb twists. (c) UMI gripper, 29×6 finger skins: can and ball rocked in the grasp. (d) Extrinsic contact: screwdriver moved down and up. Source: Ariel Slepyan, Geeta Chandra Raju Bethala

The Science

Touch is a rich source of information for robots, but it has been harder to deploy than cameras or joint sensors. A camera can often be bolted onto a standard mount. A tactile skin must match the shape of a fingertip, palm, gripper, or other surface; it must also route its many electrical connections and produce data in a format the robot’s software understands. A sensor that works beautifully on one robot may not fit another.

OpenTac, introduced by Ariel Slepyan and colleagues, takes on that integration problem as much as the sensor problem (Slepyan et al., 2026). The sensing elements, called taxels, are small pressure-sensitive cells arranged in a flexible printed circuit board, or FPCB. A resistive film sits above the circuit: pressure changes its electrical resistance, and the electronics convert that change into a signal. The key mechanical addition is a series of soft, structured bumps above groups of taxels.

Press straight down on a bump and the underlying cells respond in broadly similar ways. The sum of their signals estimates normal force—the force perpendicular to the surface. Push sideways and the bump tilts, redistributing pressure among its neighboring taxels. Differences in their readings reveal the direction of shear, or tangential force. In the simplest illustrated arrangement, four taxels under one bump provide a shared response for normal load and directional differences for two shear components. That turns a relatively simple resistive array into a sensor that can distinguish a push from a sideways drag.

The sensor is only one part of the system. OpenTac-CAD generates circuit layouts and fabrication files from a drawn outline or a surface selected on a 3D robot model. It can also accept a verbal design request through a language-model agent, which proposes dimensions and sensor settings before the CAD compiler checks and routes the design. The system writes a machine-readable “manifest” containing the sensor’s geometry, wiring, and bump groupings. That common description lets the other tools know what a particular custom sensor looks like.

OpenTac-Sim uses the manifest to create simulated tactile sensors, including a finite-element model of the deformable bumps. OpenTac-Transfer maps calibrated force readings between different sensor layouts using their physical positions, rather than requiring paired training data for every pair of sensors. OpenTac-Xampling changes how the array is read: instead of scanning every taxel in sequence, it adaptively samples likely contact regions, spending some samples searching for new contacts and others following contacts already detected.

Figure 2: OpenTac-CAD creates manufacturable tactile skins for 2D/3D shapes.
(A) Selected surfaces (orange) on YUBI, UMI, LEAP Hand, Robotiq 2F-85, and Unitree B2 models and the generated FPCBs; labels give taxel count, pitch, and row-plus-column readout channels.
(B) (i) Sensor stack: a printed bump lid places one bump over each ×2 taxel group above the piezoresistive film and two-layer FPCB. (ii) One generated taxel and (iii) a routed corner of the YUBI board; values in orange are compiler inputs, shown at the YUBI settings, that can be changed to generate a different sensor.
(C) Sensorized area versus the largest inscribed rectangular array at the same pitch, and generation time versus taxel count for all 31 fabricated boards, all with zero DRC violations.
Figure 2: OpenTac-CAD creates manufacturable tactile skins for 2D/3D shapes. (A) Selected surfaces (orange) on YUBI, UMI, LEAP Hand, Robotiq 2F-85, and Unitree B2 models and the generated FPCBs; labels give taxel count, pitch, and row-plus-column readout channels. (B) (i) Sensor stack: a printed bump lid places one bump over each ×2 taxel group above the piezoresistive film and two-layer FPCB. (ii) One generated taxel and (iii) a routed corner of the YUBI board; values in orange are compiler inputs, shown at the YUBI settings, that can be changed to generate a different sensor. (C) Sensorized area versus the largest inscribed rectangular array at the same pitch, and generation time versus taxel count for all 31 fabricated boards, all with zero DRC violations. Source: Ariel Slepyan, Geeta Chandra Raju Bethala

What They Found

The design pipeline produced 31 fabricated circuit boards with different shapes and sensing layouts. The paper reports that every board passed design-rule checks without violations and that all fabricated sensors worked. Five examples fitted to robot parts contained 153 to 467 taxels, while requiring 35 to 46 electrical channels. That matters because a large sensor need not demand one wire per taxel: the circuit addresses cells through row-and-column connections.

Shape-aware packing helped make better use of irregular surfaces. The authors compared the active sensor area with the largest rectangular grid that could fit at the same spacing. For five highlighted robot surfaces, OpenTac-CAD covered 75% of the YUBI finger region, 90% of the UMI gripper region, 89% of the LEAP Hand region, 87% of the Robotiq gripper region, and 93% of the Unitree B2 foot region. The rectangular-grid comparison reached 57%, 90%, 79%, 87%, and 93%, respectively. The advantage was largest for the tapered YUBI and LEAP shapes: 18 and 10 percentage points more coverage. On those designs, the computer generated fabrication files in 0.7 to 2.6 seconds.

Shape-aware layouts use more of irregular robot surfaces

Active taxel-cell coverage compared with the largest rectangular grid that fits at the same pitch for five selected robot regions.

Shape-aware layouts use more of irregular robot surfaces
LabelValue
YUBI75 % coverage
UMI90 % coverage
LEAP Hand89 % coverage
Robotiq 2F-8587 % coverage
Unitree B293 % coverage

The researchers also tested language-guided design. In nine agent sessions, a coding agent used short prompts and, where needed, web searches to assemble design inputs; the CAD system then performed the actual packing, routing, and rule checks. All nine generated layouts were design-rule clean, with 76 to 254 taxels. The median session took 6.5 minutes. These were design demonstrations, not nine additional fabricated sensors, an important distinction when judging what is already validated.

The highest-speed result addresses a bottleneck that grows with sensor size. Conventional raster scanning reads a 1,024-taxel array at 67 updates per second in the reported comparison. With adaptive sampling of 16 taxels per update, OpenTac-Xampling reached 1,286 updates per second—19.2 times the raster rate. That matters for brief contacts: the paper notes that fast interactions can last under 10 milliseconds, short enough to be missed by slower whole-array scans. Adaptive sampling assumes contacts are sparse and tend to persist near where they were detected; if contact spreads across the entire skin, the method falls back toward raster-like complexity.

Adaptive sampling reads contacts faster

Reported update rates for a 1,024-taxel skin, comparing adaptive partial scans with full-array raster scanning.

Adaptive sampling reads contacts faster
LabelValue
Adaptive: 16 taxels per update1,286 Hz
Full-array raster scan67 Hz

The most immediately legible downstream test is a grip under load. In the reported table, the robot held the object in 9 of 20 trials with force/torque feedback, and also 9 of 20 with OpenTac using only normal-force information. With OpenTac’s three-dimensional sensing, it held the object in 16 of 20 trials. Among trials where it held, average slip was 5.5 cm with the force/torque sensor, 5.8 cm with normal-only OpenTac, and 2.6 cm with three-dimensional OpenTac. The table also includes a no-load condition: three holds with proprioception alone, compared with 20 each for the force/torque sensor and both OpenTac modes. Under no load, the reported average slip was 0.15 cm for the force/torque sensor, 0.2 cm for normal-only OpenTac, and 0.11 cm for three-dimensional OpenTac.

Three-dimensional touch improved loaded grasp outcomes

Held-trial counts and mean slip distance under load for three feedback approaches. Slip values are shown as centimeters; held counts are out of 20 trials.

Three-dimensional touch improved loaded grasp outcomes
LabelValue
Force/torque sensor9 held / slip cm
OpenTac normal-only9 held / slip cm
OpenTac 3D16 held / slip cm

That comparison suggests the added shear signal is useful when the contact is challenged, rather than merely duplicating a conventional pressure measurement. But the result is a small experimental table, not a universal estimate of how much tactile feedback improves grasping. The supplied paper text also describes demonstrations of contact-rich control and manipulation learning on four robot embodiments, and a peg-insertion task; the complete numerical details of those experiments are not included in the available text. The evidence here supports feasibility and a promising controlled result, not a broad claim that all robots will improve by the same amount.

Figure 8: SO-101 peg insertion with deformable tactile pads.
Blue spheres and orange cubes mark training (n=100) and held-out
evaluation (n=50) start poses, respectively. The purple curve shows
an expert trajectory, with translucent gripper configurations at
multiple timestamps; the dashed gray curve shows an alternative
demonstration.
Inset shows close-ups of the pads during insertion. Arrows show per-bump tangential displacement.
Figure 8: SO-101 peg insertion with deformable tactile pads. Blue spheres and orange cubes mark training (n=100) and held-out evaluation (n=50) start poses, respectively. The purple curve shows an expert trajectory, with translucent gripper configurations at multiple timestamps; the dashed gray curve shows an alternative demonstration. Inset shows close-ups of the pads during insertion. Arrows show per-bump tangential displacement. Source: Ariel Slepyan, Geeta Chandra Raju Bethala

Why This Changes Things

The central contribution is not simply that a robot can feel sideways force. Other tactile systems have measured shear, and other groups have built flexible skins. OpenTac’s distinctive bet is that sensor design, fabrication, simulation, data acquisition, and software compatibility should be treated as one connected problem. A skin can only become useful in everyday robotic research if it fits the robot, can be built without a bespoke electronics effort, and produces observations the control system can interpret.

That full-stack approach could widen who can experiment with touch. A lab designing a new gripper might be able to specify a sensing region from the robot’s geometry instead of starting with a fixed commercial shape or developing a custom circuit from scratch. The paper’s 31 boards, spanning robot-derived and synthetic outlines, are evidence that the workflow can accommodate variety. The freely available hardware, software, and firmware may also let other teams modify and reproduce the designs rather than treating the sensor as a closed product.

The faster readout could change what tactile feedback is useful for. A 67 Hz scan gives a new reading about every 15 milliseconds; a 1,286 Hz adaptive stream corresponds to less than a millisecond between updates, if sustained under the reported sampling setup. Robots could potentially react to rapid slips, taps, or impacts instead of averaging them into a slower stream. The benefit is not only speed: because the method samples a small subset at a time, large skins may not have to sacrifice temporal resolution simply because they contain many taxels.

There is a practical bridge between custom hardware and reusable algorithms, too. OpenTac-Transfer maps contact measurements according to sensor cells’ physical areas and orientations. If a policy or analysis method is developed using one sensor geometry, that coordinate-based mapping may help carry observations to another. It is a different strategy from learning a shared representation from paired data: it uses calibration and geometry. That is potentially simpler when sensors can be described precisely, though it depends on knowing their alignment and calibration.

The work also makes an important tradeoff visible. The sensor uses a low-cost resistive stack and a mechanically structured cover, rather than a camera and complex optical assembly. This simplicity is attractive for wrapping robot surfaces, but the design still needs calibration, fabrication, and mechanical choices suited to the expected force range. A smaller taxel pitch gives finer spatial sampling, while larger sensing elements can provide a stronger signal; the system exposes those decisions but does not make them disappear.

What's Next

The next test is durability in the settings where robots actually work: repeated loading, abrasion, sweat or dust, changing temperatures, and the occasional hard impact. The paper demonstrates functional fabrication and robot deployments, but the available text does not report long-term wear results or a systematic cost comparison across the full ecosystem. Those questions will matter if a flexible skin is to move from a research bench into production environments.

Force estimation also deserves careful interpretation. The paper describes calibrating the combined normal-and-shear signals against a commercial force/torque reference and shows sensor characterization, but the supplied excerpt does not provide the numerical error values from those plots. Readers should therefore distinguish the existence of a three-axis calibration evaluation from a particular accuracy claim. Performance may vary with sensor materials, bump geometry, contact location, and saturation, especially as designers change the taxel pitch or force range.

The simulation tools have a related boundary. OpenTac-Sim models deformable bumps with finite elements and supports sensor geometries from the manifest, but the current release is described as stable for rigid-object interactions. Soft objects, complex friction, wear, and subtle stick-slip behavior are harder to simulate and are important in tasks such as handling food, cloth, or fragile goods. Cross-sensor transfer likewise assumes that calibration and contact-frame alignment are available; it cannot rescue a poorly calibrated or misaligned skin by itself.

And adaptive sampling’s speed is most useful when its assumptions hold. A small contact that moves gradually is a natural fit for searching and tracking active regions. Widespread contact or many simultaneous touch points can consume the sampling budget, so performance under dense, fast-changing contact deserves more testing. A robust system may need to combine adaptive tactile scans with other sensors and a dependable fallback mode.

Still, OpenTac offers a valuable direction: make touch customizable without making every new sensor a one-off invention. If independent teams can reproduce the hardware, extend the design compiler, and test the tools on new robots, the resulting ecosystem could turn tactile sensing from an expensive specialty into a more ordinary layer of robotic design. Robots will not become dexterous through sensors alone. But a skin that can be shaped for the machine, read quickly, and understood by its software gives researchers a more practical way to teach machines what contact feels like—and what to do when that contact begins to change.