Latest / Embodied AI 101 / DexSkin: A High-Coverage, Conformable "Electronic Skin" for Robot Fingers
Transcript
- DexSkin: A High-Coverage, Conformable “Electronic Skin” for Robot Fingers. Imagine trying to wrap a rubber band around a small box while wearing thick ski gloves, or blindly fishing a key from your pocket. Even if your muscles and vision are perfect, the lack of fingertip sensitivity makes these simple tasks almost impossible. In everyday life we take for granted the ability to feel an object’s position and motion through tiny pressures on our fingertips, sides, and even the backs of our fingers. Robots, by contrast, typically suffer from a permanent “thick glove” handicap: most grippers are rigid, and existing tactile sensors are often flat, limited to the fingertip, and costly (see Fig.1). They simply cannot cover the curved surfaces of a gripper. This lack of coverage – and the resulting blind spots – makes many contact-rich tasks (in-hand manipulation, object reorientation, tool use, etc.) extremely hard to learn or execute by a robot. In response to this gap, Wistreich et al. (Stanford/UAB) introduce DexSkin, a novel soft, stretchable robotic “skin” that wraps conformally around a robot gripper’s fingers. As the authors put it: DexSkin is “a soft, conformable capacitive electronic skin that enables sensitive, localized, and calibratable tactile sensing, and can be tailored to varying geometries”. In practice, this means each robot finger can be covered end-to-back with hundreds of tiny pressure sensors (“taxels”), instead of leaving most of the finger blind. With this rich tactile array, the robot can truly feel where and how much it touches an object anywhere on the finger – enabling dexterous actions that were previously too uncertain or unsuccessful for vision alone. The authors demonstrate DexSkin on challenging manipulation benchmarks (in-hand pen reorientation, elastic-band box wrapping, and delicate berry picking) and show that the tactile feedback dramatically improves learning-based controllers. Crucially, they also devise a quick calibration method so that models trained on one patch of DexSkin can be transferred to fresh new sensors – a key step for real-world deployability. In this article we dive into the details of DexSkin: its design and fabrication, how it’s integrated on a robot hand, its performance characteristics, and the experiments showing what new skills it enables. We’ll also touch on the calibration strategy that lets learned policies survive sensor replacement. The goal is to give you, the robotics or ML researcher, a complete technical picture: what DexSkin does, how it works, and why it matters for learnable contact-rich manipulation. The “Thick Glove” Problem of Robotic Manipulation. Robots today often rely heavily on vision. Cameras and deep learning have moved things forward tremendously in pick-and-place and simple grasping. But vision alone lacks information about contact. When two surfaces slip or bump, vision can’t directly tell a robot the subtle pressure or shear forces at the interaction. Humans use our sense of touch exactly to perceive such things and adjust our grip in milliseconds. Without it, even basic tasks become brittle. The DexSkin authors illustrate this point vividly: “Imagine wearing heavy, stiff winter gloves and trying to blindly fish a specific key from your pocket, or to loop a rubber band around a small box. Even if your muscles know what to do, the lack of tactile feedback makes it nearly impossible”. In those scenarios, we rely on feeling the key’s edges with the tips and sides and even the back of our fingers to guide it. We detect very small pressure changes as things slide. Standard robot grippers, however, are usually rigid with little coverage. Most existing sensors (flat-topped pads, small cameras like GelSight, magnetic pads, etc.) cover only a tiny patch of the fingertip. They never wrap the finger, and they often miss any side or back contact. The review explains:. “Most robot grippers are rigid, and existing tactile sensors are typically flat, costly pads covering only a small fingertip area. They lack the conformability to wrap curved robot fingers, and lack coverage of side or back-of-finger contacts”. This “blindness” is especially problematic for learning-based control. Suppose you train a neural policy for gentle reorientation of an object. If the sensor data can’t tell the controller exactly where on the finger the object is touching at each moment, the policy can’t learn precise in-hand adjustments. Moreover, these soft sensors must be repeatable and calibratable if we want to train on one sensor and then swap in another (no one wants to retrain a million rolling policies every time a skin wears out). DexSkin is explicitly designed to meet these challenges. It aims to provide large-area, high-resolution pressure sensing over nearly the whole finger surface, in a soft, stretchable material that can be manufactured quickly. The authors describe it as closed to human skin on fingers: “covering the entire finger surfaces” with taxels, giving high coverage and sensitivity for dexterous tasks. Below we unpack how DexSkin achieves this. DexSkin Hardware Design and Fabrication. At its core, DexSkin is a capacitive tactile array. That means it’s built like a sandwich: two conductive electrode layers separated by a compressible dielectric. Wherever the electrodes cross, you get a little capacitor whose value changes when pressed. In DexSkin’s design, the electrodes and dielectric are made from soft, stretchable materials. By arranging the electrode “wires” in a grid pattern on both layers, the sheet forms many individually-addressable touch pixels (“taxels”). The material choice is key: the electrodes are screen-printed onto a thin film of SEBS (styrene-ethylene-butylene-styrene), a highly stretchable thermoplastic elastomer. SEBS can stretch and flex thousands of times without tearing, which is crucial for a conformal skin that will wrap around joints and curves. The dielectric layer is then a 20-micron-thick spin-coated silicone (possibly PDMS) which is also very soft. When the robot presses the skin, the dielectric compresses slightly, changing each pixel’s capacitance. But making a planar sensor wrap onto a finger without wrinkles is a classic challenge. DexSkin addresses this with a clever 2D “flower-petal” electrode pattern. As the team describes it, they design the electrode layout to look like a flat flower. When this patterned sheet is folded around a 3D fingertip core, the “petals” come together to form a smooth dome. This avoids the creasing and gaps one would get by just bending a flat sheet around a sphere. The blog summary explains: “The electrode pattern looks like a flower. When folded to the 3D-printed finger core, the ‘petals’ converge to form a seamless dome”. In other words, they cut the flat electrode sheet into curved segments such that it transform into a near-hemisphere once wrapped. Practically, each gripper finger is built in three layers: a rigid inner core (to give shape), a soft compliant sleeve (for giving some squish), and the outer DexSkin layer (with electrodes). For the fingertip dome portion, the team’s design forms 12 distinct taxels on the spherical tip. Around the finger sides, additional flat strips are wrapped, forming 48 more taxels around the finger’s cylindrical body (covering about 294° around the circumference). In total, each finger ends up with 120 taxels. That means anywhere you press on the tip or most of the sides, at least one capacitive sensor picks it up. The large layout is summarized in the project summary: “The result is a robotic finger with 120 individually-addressable tactile pixels (taxels). Dome tip: 12 taxels for fine sensing at the tip. Cylinder body: 48 taxels covering 294°”. This is far more coverage than, say, a single rigid fingertip sensor. Material-wise, the electrodes are patterned on ~240μm SEBS films. According to the supplemental info, they spin-coat resin into micro-structured molds and then print silver traces for conductivity. The result is a very thin, stretchable film. The pieces are then assembled by hand: the inner electrode plate is attached to the finger core, the dielectric layer wrapped over it, and finally the outer electrode plate (the “flower” dome) is sealed on top. Because it is largely flat-piece manufacturing, this process is surprisingly quick: the authors note that DexSkin can be customized and built same-day at low cost. In other words, if you need a skin for a different gripper geometry, you could redesign the pattern and print it out in a few hours. All the electrodes are wired back to one central readout board mounted on the robot’s wrist. The group built a custom 72.6×40.6 mm PCB driven by an onboard microcontroller (an ESP32-S3). That controller sequentially scans all 120 taxels via analog multiplexers and a capacitance-to-digital chip, converting the tiny capacitance changes into digital values. The board communicates these readings at about 30 Hz, enough bandwidth for real-time control (effectively giving 30 tactile frames per second for all finger taxels). In short: “DexSkin is integrated into a tabletop manipulation system… on a 7-axis Franka Emika Panda arm. Each finger is a three-layer structure: the DexSkin sensor, a deformable sleeve for compliance, and a rigid inner core. A custom board scans all 120 taxels at 30 Hz”. This integration turned a standard parallel-jaw gripper into a high-coverage tactile sensing hand. In summary, DexSkin’s physical design innovations are:. Soft, stretchable construction (SEBS elastomer electrodes, ~20 μm silicone dielectric). Conformal coverage via a “flower petal” layout, yielding dozens of taxels wrapping the finger’s tip and sides. High taxel count (120 per finger) enabling near-full coverage of a gripper finger. Modular assembly, so new skins can be produced quickly on custom molds, enabling same-day fabrication. Sensor Performance: Range, Resolution, and Repeatability. Designing the skin is only half the battle – it must also perform well as a sensor array. DexSkin’s performance characterization is impressive. The capacitive taxels turn out to be very sensitive and consistent. In systematic tests, the team applied controlled forces to one taxel at a time and recorded the response. Here are some highlights:. Force sensitivity and range: Even a very small normal force produces a measurable signal. With a soft, low-density sleeve, DexSkin can sense as little as 27 milliNewtons (≈1.7 kPa of pressure) – roughly the weight of a small coin pressing on your finger. At the high end, with a stiff sleeve, it can handle up to 40 Newtons before saturating (≈2527 kPa). In other words, the same skin easily detects both a gentle poke and a firm grasp. This wide dynamic range (from ~30 mN to ~40 N) covers the sorts of forces one expects in manipulation. Spatial resolution (taxel size): The physical spacing of taxels is on the order of 0.5–0.6 mm between neighboring electrodes, giving sub-millimeter resolution. This exceeds many prior tactile arrays. For comparison, a researcher-compiled table shows DexSkin’s spatial resolution at about 0.60 mm, compared to ~1.0 mm for ReSkin, ~1.4 mm for BioTac, and much coarser values for others. Combined with the large area (up to ~6,471 mm² for the finger), DexSkin gives extremely high-resolution coverage. Force resolution and noise: Across taxels, the contact force estimates are precise. The authors report a root-mean-square (RMS) force estimation error of about 0.086 N over a 0–2.5 N range. So grasp forces just above 0.1 N can be distinguished. Notably, DexSkin’s capacitive signals are quite linear (in log-space) with pressure, making calibration straightforward. Hysteresis & repeatability: Soft sensors often suffer from hysteresis (lag) when compressing and releasing. DexSkin performs very well here: the average hysteresis was only about 6.5% of full-scale. In practice this means the sensor recovers quickly and outputs repeatably as you load/unload it. In durability tests, after 500 squeeze-and-release cycles at high pressure, the sensor’s baseline drifted by only ~1–2% and peak response by ~2%. These small drifts imply the robot can use the skin reliably over long routines without constant recalibration. Low crosstalk: Equally important for high resolution is minimal interference between taxels. DexSkin achieves excellent localization: pressing one taxel produces almost no ghost signals in neighbors. The team measured crosstalk by loading one taxel and checking others; the largest unwanted reading was under 3% of the main signal, with a mean around 1.5%. This level of isolation is much better than many soft sensors (some can see 20–30% ghost peaks). It means the robot can clearly tell when there are one vs. two separate contacts, which is vital for fine manipulation. In short, DexSkin behaves like a high-fidelity multi-pixel pressure camera on the robot finger. It can detect very light touches (<0.03 N) and heavy contacts (>10 N) while maintaining ~millimeter spacing and <0.1 N force accuracy. The capacitive mechanism and uniform fabrication give it predictability (good taxel uniformity) and low lag. These attributes make the data usable for learning algorithms. One striking comparison from the project data shows DexSkin’s advantage: sensor coverage. Most existing tactile solutions – like the well-known BioTac sensor (484 mm² area) or the GelSight finger (3,142 mm²) – cover only a fingertip patch. By contrast, DexSkin’s full-finger wrap spans up to ~6,500 mm² of surface. The force sensitivity around 0.09–0.1 N is competitive with or better than many, and the sub-millimeter spacing beats almost all (see table below for context). Aligned with human fingertip skin in coverage, DexSkin offers robot controllers a touch field unparalleled in earlier work. Table 1 (from DexSkin data): Comparison of tactile sensors. DexSkin’s area (~1944–6471 mm²) far exceeds others, with sub-millimeter spatial resolution. (DIGIT, GelTip, etc., for reference.). Control Experiments: Dexterous Tasks with and without DexSkin. With the hardware and sensing established, the core question is: What can a robot do with this skin that it couldn’t do before? The authors test DexSkin on three representative manipulation challenges requiring fine contact. In each case, they train a control policy (typically by imitation learning) and then compare performance with DexSkin input versus without. The results are striking: the added tactile feedback transforms marginal policies into reliable solutions. Task 1: In-Hand Pen Reorientation. In this benchmark, the robot must pick up a pen lying on the table and flip it to a vertical orientation in its grasp, so it’s ready for writing. This requires gently sliding and adjusting the pen’s orientation inside the fingers. Vision alone can tell the rough pose, but only touch can sense the small roll of the pen tip. Challenges: The robot needs to track the pen’s rotation angle via subtle contact, and adjust forces to twist it upright. If the pen slips out of contact or the robot misestimates pose, it fails. To test robustness, the authors even poke the pen by hand at random times, forcing the robot to recover. Method: They collected human demonstrations and trained a diffusion-based imitation policy (essentially learning from experts). They ran four variants: (a) Vision-only (no touch), (b) “Spatial average” touch (naively averaging all taxels into one signal), (c) DIGIT (a single-point fingertip camera sensor, as a partial-touch baseline), and (d) DexSkin. Results: With DexSkin, the robot succeeded in 19 of 20 trials, and likewise recovered from all 19 times it was poked by a human perturbation. In contrast, all the baseline policies without rich tactile input almost always failed to recover after a disturbance – the pen slid away and the task could not be completed. In one example, a vision-only policy oriented the pen somewhat, but once it was bumped the robot had no cue to the new pen angle and dropped it. The DexSkin-enabled policy, by sensing exactly where on the fingers the pen applied pressure, could detect the roll and continuously adjust. In short, high-resolution skin allowed the robot to “feel” the pen rolling and keep it under control, whereas coarse or no touch was insufficient. Quantitatively, the success gap is huge. DexSkin yielded ~95% success (19/20), whereas the baselines essentially 0% on the perturbed cases. The authors report: “DexSkin-equipped robots successfully reoriented pens (19/20) and recovered from human perturbations (19/20), while policies with no tactile, spatially pooled tactile, or DIGIT sensors failed to recover”. This dramatic difference highlights how full-coverage, fine-grained touch can supply the missing link for in-hand pose estimation. Task 2: Box Packaging with Elastic Bands. Next, the robot must secure a rubber band around a small box with a lid (imagine packaging a gift). A twist: sometimes the band is a “perforated” sacrificial one that will break if overstretched. The robot has to learn to wrap the band tightly, but if it feels the band tear (sensed as a sudden loss of tension), it should drop it and grab a new intact one. Challenges: This task needs contact on the back side of the fingers – human fingers often use their backs to stretch a band. If the gripper only has sensors on the gripping surface, once the band goes around the outside of the finger there is no feedback. The robot must also distinguish a band that has broken (tension loss) from one that is still intact. Method: Again, an imitation policy is trained. The environment gives the robot either an intact band or a pre-perforated band (which will snap if pulled). The goal is to loop the band around the box lid. DexSkin’s wrap-around coverage is crucial here: it can actually feel the band shifting onto the outer (dorsal) part of the fingers. Results: The DexSkin-augmented policy succeeds far more often. When given a perforated (bad) band, the robot senses the tension drop mid-stretch via the side/back sensors and promptly discards it. With a good band, it wraps tightly by using sense from the sides too, preventing slippage. Importantly, DexSkin was the only setup that reliably completed the task. Limited-coverage baselines (e.g. sensors only on the inner gripping surface) failed in most trials. Once the band slid to the outside, those policies lost track of it and could not finish wrapping. The authors summarize: “DexSkin: The robot identified fragile bands (via sensing low tension) and discarded them. It then used side- and back-of-finger sensors to maintain band stretch without slipping. Limited-coverage baselines failed at a high rate once the band moved to the outside of the gripper”. Numerically, the DexSkin policy correctly selected and wrapped the intact band in about 15–18 out of 20 trials, whereas the others rarely succeeded. One table of results shows DexSkin as the only condition that consistently picked the right band and completed wrapping. Figure 2 (hypothetical) would show example runs: DexSkin tip sensors brightly lit as the band goes around, whereas a non-DexSkin run loses contact. In any case, this task drives home that omnidirectional coverage – being able to feel on the sides of the finger – is not a luxury but a necessity for certain manipulations. DexSkin’s design directly enables the robot to “feel” the back of its own fingers, turning a nightmare of blind manipulation into a solvable problem. Task 3: Fragile Berry Picking with Residual RL. The final task is a classic “gentle grasp” challenge. Blueberries (or similarly delicate fruits) must be picked up and moved without crushing them. This is often cited as a need for tactile finesse: how strong is the grip? When does the contact force approach the berry’s limit? Vision alone won’t know if it’s too tight until the berry is squished. Challenges: A hard-coded or purely vision-based policy tends to squeeze too much, especially since model uncertainties and weight differences can vary. The robot must learn through trial and error the right amount of force. This is done with reinforcement learning (RL), rather than pure imitation. However, learning from scratch on real fragile berries would damage hundreds of them. Method: The team uses a clever two-stage “residual RL” approach. First, they collect demonstrations of the robot just picking and transporting a generic object (it might clamp berries too hard, but it roughly learns the motion) – call this the base (imitation) policy. Then they freeze that as a backbone and train a small residual policy network on top, which receives only the tactile readings from DexSkin (and perhaps small proprioceptive cues). The residual network learns an adjustment to the base controller’s actions to avoid high force. The reward is shaped specifically: it penalizes the squared amount by which any taxel’s force exceeds a threshold (essentially punishing high contact pressure). Results: This combination works well. The learning quickly converges to a policy that lifts berries “just enough” without crushing them. In experiments, the DexSkin-augmented residual policy achieves a high success rate in moving berries intact, whereas the baseline policy (the imitation-only pick) fails frequently by squeezing them too much. The authors show learning curves where damage incidents drop sharply. In summary: the system learned to grasp with just the right gentle force, guided by DexSkin’s real-time feedback. While exact numbers aren’t given in the writeup, the text notes a “significant reduction in damage compared to the baseline”. This task highlights another dimension: data-driven adaptation. The reinforcement policy relies on fast tactile feedback to continuously modulate grip force. Without DexSkin, there’s no direct observation of force beyond the safe baseline, so the best it can do is guess, leading to many crushed berries. With the skin’s signals, the policy knows exactly when it is beginning to overstretch the berry, and can immediately back off. In all three tasks, the story is the same: touch matters. In tasks requiring fine reorientation or dealing with uncertain elastic forces, DexSkin’s rich array allows the robot to sense critical contact information. In each case, the learned policy with DexSkin far outperforms policies that lack high-resolution, full-coverage touch. The authors report near-perfect success for DexSkin muscle strategies (e.g. 19/20 trials). Equally important, the baseline policies without it repeatedly failed in scenarios where sensing on the finger’s back or detailed pressure differences were needed. Calibration and Model Transfer. One unique contribution of the DexSkin work is addressing the “holy grail” problem in tactile robotics: how to reuse policies across different sensor instances. In practice, any handmade soft sensor will vary slightly, so a neural network tuned on Skin A may break when Skin B is put on (the response curves differ). For truly deployable learning, replacing a worn sensor should not force complete retraining. DexSkin tackles this through rapid per-sensor calibration. The authors built a simple pneumatic chamber that can uniformly inflate around the entire finger, applying a known pressure to all taxels at once. In practice, they calibrate a new skin by recording its raw readings under several known pressures. By comparing these to the “source” skin’s calibration (the one originally used for training), they derive a mathematical mapping between them. The actual formula involves exponentials and logs (it matches two exponential curves), but conceptually it’s like finding how to scale and shift the new sensor’s output to match the old one. The key is that this mapping can be computed in ordinary minutes, meaning a fresh DexSkin can be installed and calibrated to behave like your trained one in almost no time. The paper emphasizes: “This calibration only takes minutes, saving hours or days of retraining… the calibrated replacement performance is nearly as good as the original, achieving model transferability – the ‘holy grail’ of real-world tactile robots.”. In experiments, they show that after calibration, rolling out a policy on the calibrated new sensor yields success rates indistinguishable from the original. Without calibration, identical force leads to different readings and the policy would misinterpret the situation. This advance in calibration means DexSkin isn’t just a one-off lab gadget: it’s a repeatable platform. A practitioner can feasibly have spare sensors and swap them out, reusing learned controllers. “Sensor variability is a huge obstacle,” the authors note, so solving it is a big deal. DexSkin’s combination of careful manufacturing regularity plus this final calibration step provides a practical path toward widely applicable tactile skins. Comparisons and Context. How does DexSkin fit into the landscape of tactile sensing? Prior to this work, most roboticists had to choose between a few classes of touch sensors:. Camera-based (vision) sensors like GelSight or TacTip can achieve very high resolution, but are basically rigid cases with a camera observing a gel. They tend to be large, only flator slightly curved surfaces, and can’t easily wrap a finger. Resistive or capacitive sheets (e.g. OhMNI skin, ReSkin) can be made thin and flexible, but usually have limited stretch and resolution; often they still cover a small pad area. Bio-inspired fingertip sensors (like BioTac) mimic human fingertip with little lumps or pins, but only cover the tip and are not scalable to wrap. Optical or magnetic arrays exist for fingertips too, but again small coverage. None of those provide seamless, full-coverage wrapping of a robotic finger. DexSkin is unique in its combination of full-geometry coverage plus high taxel density and calibratability. In direct comparison metrics (see Table 1), its sensing area (thousands of mm²) is an order of magnitude beyond flat pads, and its spatial and force resolution is state-of-the-art. This is not to say DexSkin has infinite resolution – at ~120 taxels/finger it’s not quite as dense as human skin on a fingertip – but it’s far above anything previously mounted on robot manipulators. The authors candidly compare it to others: for example, DexSkin covers ~4–6 times the area of DIGIT360 and has about 10x more taxels than standard GelSight (see sensor comparison table). The outcome is clear: if you need to feel anywhere on your robot finger, DexSkin is in a class by itself. One nice bonus is that DexSkin is being open-sourced. The team has released all hardware and firmware designs, including the sensor geometry files, PCB layouts, and example interface code. This means researchers can build their own DexSkin or adapt it to new fingers. The posted website even includes extended FAQs, assembly instructions, and calibration footprints. By contrast, many commercial tactile skins keep designs proprietary, which has slowed broader adoption. DexSkin’s open-nature follows the growing trend (see OPEN AI Safety Gym, etc.) but for touch hardware. We can expect to see cloned or improved versions soon. Conclusions and Outlook. DexSkin represents a significant step forward in embodied AI research. By giving robots a sense of touch across nearly their entire fingertip geometry, it opens up manipulation skills that were impractical or impossible before. The suite of experiments – pen twisting, band wrapping, gentle picking – were carefully chosen to highlight precisely this new capability. In every case, the touch signals turned a failing strategy into a successful one. For professional roboticists and ML researchers, the implications are clear. If your manipulation task involves sliding, reorienting, dynamic adjustments, or any contact on the finger sides, DexSkin or skins like it should be on your radar. Tactile feedback is no longer an afterthought; it can be an essential affordance. In future work, we may well see more integration of sensors like this with powerful vision-based “foundation models” or world models, where vision plans the gross motion and skin refines the grips in real time. There are challenges ahead too. DexSkin in its current form is a finger-skin for a parallel jaw gripper. Extending it to anthropomorphic hands with more joints might need more sophisticated routing of the skin around knuckles. The calibration approach seems promising, but it would be nice to automate a bit further or to reduce the need for re-inflation if we could somehow self-calibrate through routine movements. Also, while 30 Hz is okay now, future dexterous tasks (like slip detection) might push for higher bandwidth or even add shear sensing. The authors mention that the sensor is calibratable, meaning in principle you could add more taxels for higher resolution at the expense of more channels. Another consideration is that training algorithms must now effectively leverage the tactile data. The success stories here used largely imitation learning with hand-crafted policies, plus a bit of shaped RL. There might be room for end-to-end RL or more advanced tactile perception (e.g. neural net encoders) now that raw touch is available. Researchers might also fuse DexSkin data with vision in a multi-modal way for even greater robustness. Finally, there is the question of robustness and materials. Soft robotics often deals with wear-and-tear. SEBS is quite durable, but real warehouses may still scuff the skin or damage taxels. The calibration step is a great mitigation, but long-term studies on sensor longevity would be valuable. The authors do note how stable the readings were after cycling, which is a good sign. In closing, DexSkin’s main message is that skin matters. As the authors conclude, they’ve demonstrated DexSkin’s “suitability and practicality for learning real-world, contact-rich manipulation”. In effect, we no longer have to imagine the world of robots with human-like touch; it’s already starting to happen. By open-sourcing the project, the team has set the stage for others to experiment, adapt, and build on this idea. For anyone working on robot manipulation, it’s worth asking: could tactile coverage like this unlock your next breakthrough?. References: The above description is drawn from DexSkin: High-Coverage Conformable Robotic Skin for Learning Contact-Rich Manipulation by Wistreich et al., CoRL 2025 (see project website and preprint).