OptoSkin: Optical Sensing in Soft Media

• OptoSkin is a design paradigm that uses continuous, skin-like media to transduce mechanical, physiological, and environmental variations into optical signals.
• Systems employ methodologies like OCT attenuation, edge-lit waveguides, and speckle imaging to achieve precise tactile and biomedical sensing.
• Research emphasizes continuous media sensing over dense taxel arrays, enabling simplified wiring and versatile applications in robotics, diagnostics, and wearables.

Searching arXiv for recent and relevant "OptoSkin" papers to ground the article. arXiv search: OptoSkin-related papers and adjacent optical-skin/tactile sensing work. OptoSkin denotes a family of optical-skin systems in which a biological skin layer, a compliant polymer, or a transparent skin-like sheet transduces mechanical, physiological, or environmental variation into optical observables. In the literature, these observables include attenuation coefficients in skin measured by OCT, light-path perturbations in elastomeric waveguides, coherent speckle fields, optical-fiber power loss, event-camera brightness changes, and direct Time-of-Flight returns (Martinelli et al., 2018, Piacenza et al., 2018, Shimadera et al., 2022, Koolani et al., 7 Jan 2026, Aulika et al., 30 Jul 2025). This breadth suggests that OptoSkin is best understood as a design paradigm rather than a single standardized device class.

1. Transduction principles and conceptual scope

Across the literature, OptoSkin systems share a common abstraction: a continuous medium replaces, or substantially reduces, dense taxel-level instrumentation, while the inverse map from optical signal to physical state is handled by regression, classification, triangulation, or explicit calibration. The major distinction is between systems that interrogate native human skin and systems that embed the optical transducer inside an engineered soft medium.

Mechanism	Representative observable	Representative work
OCT attenuation modulation	$I(z)=I_0\exp(-\mu z)$ and $\mu(t)$	vPPG through outer skin layers (Martinelli et al., 2018)
Edge-lit waveguide perturbation	64 ambient-subtracted LED/PD intensities	PDMS tactile localization and depth sensing (Piacenza et al., 2018)
Speckle interference	$I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$	multimodal force-position-temperature sensing (Shimadera et al., 2022)
Optical-fiber attenuation	$\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$	self-calibrated normal/shear tactile sensing (Chen et al., 2023)
Event-based brightness change	DVS events $e=(u,v,t,p)$	stereo opto-tactile skin (Koolani et al., 7 Jan 2026)
Direct ToF ranging	$d=ct/2$	tactile sensing with a 940 nm ToF head (Aulika et al., 30 Jul 2025)

A recurrent misconception is that optical skin necessarily implies either camera-only vision sensing or large-area integrated tactile arrays. Several OptoSkin variants explicitly avoid both. The 2018 edge-mounted PDMS sensor uses eight LEDs and eight photodiodes around a continuous elastomeric volume, while the 2022 speckle-based optical skin senses contact force, contact location, and temperature using a single soft material without requiring complex integration (Piacenza et al., 2018, Shimadera et al., 2022).

2. Human skin as the sensing medium

In biomedical OptoSkin, the skin itself is the optical transducer. A central result is the demonstration that arterial pulsation modulates the optical attenuation coefficient of outer, even non-vascularized, skin layers. In a homogeneous weakly scattering model, OCT backscatter amplitude follows

$I(z)=I_0\exp(-\mu z),$

with $\mu$ the total attenuation coefficient. The pulse-coupled model writes

$$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$

and the measurement model becomes

$$I_{\rm meas}(z,t)=I_0\exp\!\left[-(\mu_0+k\Delta P(t))z\right].$$

Using a swept-source OCT system (Thorlabs OCS1300SS) at 1300 nm, with 12 µm axial resolution, 25 µm lateral resolution, 10 B-scans/s over 30 s, and a $\mu(t)$0 ROI, attenuation coefficients were extracted from linear fits of $\mu(t)$1 over $\mu(t)$2. The reported mean attenuation coefficient for outer skin layers was $\mu(t)$3, the temporal standard deviation was $\mu(t)$4, and the spatial standard deviation across 2 mm lateral was $\mu(t)$5. Because $\mu(t)$6, the effect was interpreted as a true temporal modulation rather than motion artifact; FFT peaks of $\mu(t)$7 matched finger PPG reference peaks, including 72 bpm in one subject and 58–85 bpm across three volunteers (Martinelli et al., 2018).

The significance of this result is methodological. It shifts the explanatory basis of video plethysmographic contrast away from a purely blood-volume-change interpretation toward a model in which arterial transmural-pressure waves induce mechanical compression and decompression of extracellular matrix and capillaries, thereby altering scattering and attenuation in superficial skin. The same work states that depth sensitivity in the $\mu(t)$8–$\mu(t)$9 range implies that even low-penetration wavelengths such as blue and green can sense the effect, and that sampling rates of at least 10 Hz are sufficient for cardiac and low-frequency respiratory components. The proposed empirical pressure-to-attenuation coefficient $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$0 is on the order of 0.005–0.02 $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$1 for soft dermal tissue, but direct calibration remains open (Martinelli et al., 2018).

Human skin also appears in OptoSkin-adjacent work as an imaging target for objective computational diagnosis. A multi-beam swept-source OCT study at $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$2 acquired 1000 healthy images across ten anatomical sites and 242 pathologically confirmed tumor images, then computed 63 optical and textural features, including attenuation, first-order statistics, GLCM features, and GLRLM features. After z-score normalization and principal-component-based feature selection, a linear SVM achieved 80.9% accuracy, 81.9% sensitivity, 80.5% specificity, and ROC–AUC $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$3 for BCC versus healthy skin, while a quadratic SVM achieved 87.2% accuracy, 87.0% sensitivity, 87.3% specificity, and ROC–AUC $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$4 for SCC versus healthy skin (Adabi et al., 2017).

A complementary optoacoustic line of work addresses smooth epidermis segmentation in raster-scan optoacoustic mesoscopy. On 31 patient RSOM volumes with voxel size $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$5 and $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$6, a U-Net and a 7-layer FCN were trained with a binary cross-entropy term plus a shape-specific smoothness penalty. For the U-Net on unseen test data, the epidermis segmentation reached Dice $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$7 and IoU $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$8, while the downstream vessel-segmentation Dice improved from $I(p)=\left|\sum_j a_j \exp(i\phi_j)\right|^2$9 without the epidermis mask to $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$0 with it (Gerl et al., 2020). Together, these studies place OptoSkin within a broader biophotonic program of quantitative skin readout.

3. Edge-lit waveguides, tactile localization, and event-driven readout

A canonical robotic OptoSkin architecture uses a transparent elastomeric waveguide interrogated from its boundary. In the 2018 optics-based tactile sensor, the active sensing area is $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$1, embedded in a $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$2 cavity and formed from Sylgard 184 PDMS mixed at 1:20, with 8 mm of clear PDMS above a 1 mm PDMS-plus-carbon-black base layer. Eight SunLED XSCWD23MB LEDs and eight Osram SFH 206K photodiodes are edge-mounted in alternating fashion. The physical basis is two-mode light transport: shallow indentation perturbs the PDMS–air interface and destroys total internal reflection, while deeper indentation blocks direct and near-direct optical paths between emitter–receiver pairs. Ambient-subtracted intensities are defined as $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$3, yielding a 64-dimensional feature vector for each sample. A linear SVM performs touch/no-touch classification, and Laplacian-kernel ridge regression estimates $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$4 (Piacenza et al., 2018).

Quantitatively, that system reported reliable contact detection of at least 90% for depths $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$5, corresponding to approximately 2 N. Under ambient light, localization median error decreased from 2.19 mm at 0.1 mm depth to 0.31 mm at 5.0 mm depth, while depth-estimation median error was 0.05 mm at 0.5 mm, 0.040 mm at 2.0 mm, 0.037 mm at 3.0 mm, and 0.081 mm at 5.0 mm. The article also emphasized a systems-level advantage: wiring scales as $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$6 for $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$7 LEDs and $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$8 photodiodes, rather than $\Delta P_i=(P_{0,i}-P_i)/P_{0,i}$9 for an $e=(u,v,t,p)$0 taxel matrix (Piacenza et al., 2018).

The same waveguide logic has been pushed toward neuromorphic sensing. An event-based opto-tactile skin uses a $e=(u,v,t,p)$1, 4 mm thick PDMS sheet with 32 edge-mounted NIR LEDs and two Prophesee EVK1 VGA Gen 3.1 Dynamic Vision Sensors mounted in a stereo configuration, looking sideways through the silicone. DVS pixels emit events when log-intensity changes exceed a contrast threshold, and DBSCAN with $e=(u,v,t,p)$2 pixels and min_samples = 10 clusters press-induced event clouds. Triangulation of the two horizontal centroids localizes the press. On a 250-point meander raster scan repeated 10 times over a probed area of approximately $e=(u,v,t,p)$3, the system achieved RMSE $e=(u,v,t,p)$4 across 95% of valid presses visible to both cameras, with RMSE$e=(u,v,t,p)$5 mm and RMSE$e=(u,v,t,p)$6 mm. The detection latency distribution had a characteristic width of 31 ms. Under stochastic down-sampling to $e=(u,v,t,p)$7 of the original event rate, RMSE rose to 9.33 mm and the pass rate fell to approximately 85%, but the system remained functional (Koolani et al., 7 Jan 2026).

A third variant moves the photometric transducer outside the contact interface. In the externally attachable photoreflective tactile sensor, a pigmented Ecoflex-0030 silicone block of nominal dimensions $e=(u,v,t,p)$8, $e=(u,v,t,p)$9, and thickness $d=ct/2$0 is paired with a Pololu QTR-1A photoreflector operating at $d=ct/2$1. The undeformed air gap is $d=ct/2$2. Reflected intensity is modeled as $d=ct/2$3, voltage is linearized as $d=ct/2$4, and contact force is fit as $d=ct/2$5. Experimentally, the force range was 0–7 N over 0–3 mm indentation, the voltage change over this range was approximately 0.40 V, sensitivity was $d=ct/2$6, hysteresis was less than 5% full scale, repeatability drift was less than 0.5% over 100 cycles, and response lag was less than 0.01 s at speeds of at least 1 mm/s (Yamamoto et al., 9 Nov 2025).

These three systems make a useful contrast. The first embeds a low-cost multiplexed optical network in the skin, the second shifts to sparse asynchronous imaging and triangulation, and the third externalizes the sensor altogether while still reading skin deformation. This suggests that, within robotic tactile OptoSkin, the core design variable is not whether optics are used, but where the optical state is measured.

4. Interferometric, fiber-based, and ultrasensitive optical skins

A different branch of OptoSkin uses coherent interference rather than boundary waveguiding. In the 2022 multimodal “optical skin,” a He–Ne laser at $d=ct/2$7 illuminates a 5 mm thick transparent silicone elastomer, and microscopic scatterers generate a speckle field recorded by a CMOS camera with 5 ms exposure. The phase of path $d=ct/2$8 is

$d=ct/2$9

and the far-field intensity at pixel $I(z)=I_0\exp(-\mu z),$0 is

$I(z)=I_0\exp(-\mu z),$1

Raw $I(z)=I_0\exp(-\mu z),$2 speckle images are down-sampled to 30% and cropped to $I(z)=I_0\exp(-\mu z),$3 pixels. A convolutional network with a shared feature extractor and branched decoders is trained with mean-squared error using Adam and batch size 50. The reported sensing resolutions were $I(z)=I_0\exp(-\mu z),$4 in depth over a 0–212 µm range, corresponding to $I(z)=I_0\exp(-\mu z),$5 force resolution, $I(z)=I_0\exp(-\mu z),$6 in lateral position over a 0–1.12 mm contact line, and $I(z)=I_0\exp(-\mu z),$7 over 17.9–26 °C. Even when two or three stimuli changed concurrently, the model retained less than 4% relative error on each channel; three indenter shapes yielded 94% classification accuracy with 3.8% depth error, and a haptic interface with four buttons was read with 100% correctness (Shimadera et al., 2022).

Fiber-based OptoSkin reaches similar goals with different inversion machinery. A polymer-based self-calibrated optical-fiber tactile sensor embeds two anisotropic layers inside elastomer: four 1 mm-diameter U-shaped PU fibers in a shear-sensitive top layer, two straight vertically crossing PU fibers in a normal-sensitive lower layer, plus a bottom photodiode measuring leakage from all bottom fibers. Normalized attenuation is defined as

$I(z)=I_0\exp(-\mu z),$8

and the six sensing channels are written as

$I(z)=I_0\exp(-\mu z),$9

where $\mu$0 and $\mu$1 are force and indentation/size variables. A two-stage linear regression estimates $\mu$2, $\mu$3, and $\mu$4, then reconstructs indentation depth, indenter radius, and force vector through Moore–Penrose-style pseudo-inverse operations. On a robotic arm tip, the reported average accuracies were 0.15 N for normal force, 0.17 N for shear force in the X-axis, 0.18 N for shear force in the Y-axis, all within a 0–2 N sensing range, with average object-size error of 0.4 mm over a 5–12 mm indenter diameter range and indentation-depth error of 0.3 mm over 0–5 mm (Chen et al., 2023).

An earlier “O-skin” formulation pushes optical softness toward very high sensitivity by embedding glass micro/nanofibers in thin PDMS films. External pressure induces curvature, increases evanescent-field leakage, and reduces transmitted power according to

$\mu$5

For small pressures in the empirical linear regime, $\mu$6 with measured sensitivity $\mu$7. Reported performance included a 7 mPa detection limit, 10 microseconds response time, vibration sensing up to 20 kHz, and practical demonstrations in wrist-pulse monitoring, voice detection, a five-sensor optical data glove with angular resolution below $\mu$8, and a $\mu$9 tactile sensor (Zhang et al., 2018).

Taken together, these systems show that OptoSkin does not require a single inversion philosophy. Speckle-based devices trade hardware simplicity for high-dimensional learned decoding, fiber-based devices exploit anisotropy to obtain closed-form self-calibration, and micro/nanofiber skins emphasize analog sensitivity and bandwidth.

5. Transparent, visuotactile, and Time-of-Flight variants

The term also encompasses platforms in which the skin-like layer is intentionally semitransparent. The See-Through-your-Skin sensor uses a P-595 silicone gel with maximum sensing area $$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$0 and thickness approximately 5 mm, coated with a lightly layered Rust-Oleum mirror spray and a protective top coat. A ring of 32 addressable NeoPixel RGB LEDs is placed around the periphery, and imaging is performed by a fisheye Raspberry Pi camera module at $$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$1 pixels. By varying internal illumination, the device can operate in pure-tactile mode, pure-vision mode, or a combined visuotactile mode. Tactile images and visual images are processed by two parallel ResNet-50 streams, concatenated into a 4096-dimensional feature before classification (Hogan et al., 2020).

This dual-use optical skin was validated on both synthetic and physical tasks. In the simulated ShapeNet experiment with 10 categories, top-1 validation accuracy was 88.8% for vision only, 83.1% for tactile only, and 96.9% for visuotactile fusion. In real-world bottle classification with 10 instances and 800 visuotactile samples, average test accuracy was approximately 76% for vision only, approximately 82% for tactile only, and approximately 94% for fusion. For bottle-fullness metrology, the respective accuracies were 56.3%, 80.6%, and 94.4% (Hogan et al., 2020). The system is notable because it treats tactile and visual sensing as two illumination regimes of the same hardware architecture.

Direct Time-of-Flight sensing defines another recent OptoSkin branch. Aulika et al. use a single-point AMS-OSRAM TMF8828 direct ToF LiDAR head coupled to a transparent or translucent waveguide. A pulsed 940 nm emitter measures the time delay to the first strong scattering interface through

$$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$2

while attenuation inside the waveguide is described by a Beer–Lambert form $$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$3. Four materials were characterized at 940 nm: acrylic glass with $$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$4, multi-layer Formlabs resin with $$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$5, single-layer Formlabs resin with $$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$6, and Liqcreate Clear Impact resin with $$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$7. Across 100 ToF frames per material–target pair, the SNR trade-off depended strongly on target class. Sample A gave the best SNR for silicone contact at 9.3, while Sample D gave the best SNR for white and gray plastics at 15.9 and 13.4; black plastic produced the lowest SNR overall. The authors recommend a mid-range scattering coefficient of approximately $$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$8–$$\mu(t)=\mu_0+\left(\frac{\partial \mu}{\partial P}\right)_k \Delta P_{\rm pulse}(t),$$9 as a compromise, yielding SNR $$I_{\rm meas}(z,t)=I_0\exp\!\left[-(\mu_0+k\Delta P(t))z\right].$$0 for both silicone and white/gray plastics with depth jitter $$I_{\rm meas}(z,t)=I_0\exp\!\left[-(\mu_0+k\Delta P(t))z\right].$$1 mm (Aulika et al., 30 Jul 2025).

These systems broaden the meaning of tactile optical skin. In one case the same semitransparent layer alternates between camera-like and GelSight-like behavior; in the other, the measurement is not an image of deformation but a range-and-scattering response from a ToF head. The unifying element is still optical readout of a mechanically perturbed skin-like layer.

6. Broader extensions, limitations, and terminological ambiguity

The present literature also uses “skin” and “optical skin” language beyond tactile robotics and physiological sensing. A skin-compatible photodetector platform based on nitrogen-doped carbon dots and single-layer graphene on PET reports broadband response from 400–800 nm, responsivities of approximately 0.19 A/W at 406 nm, 0.32 A/W at 642 nm, and 0.18 A/W at 785 nm, operation below 1.5 V gate bias, reversible performance down to a bending radius of 0.8 cm, and non-cytotoxicity in SkinEthic testing (Loudhaief et al., 25 Aug 2025). In a different field entirely, optically-transparent electromagnetic skins are implemented as dual-layer copper-mesh metasurfaces laminated on standard “4-10-4” insulating-glass panels; at 26 GHz the optimized meta-atom provides approximately $$I_{\rm meas}(z,t)=I_0\exp\!\left[-(\mu_0+k\Delta P(t))z\right].$$2 phase swing, worst-case transmission magnitude $$I_{\rm meas}(z,t)=I_0\exp\!\left[-(\mu_0+k\Delta P(t))z\right].$$3, and optical transparency above 80–90% (Oliveri et al., 2023). These usages suggest that the term is polysemous: it can denote a tactile sensor, a biophotonic readout of human skin, a wearable optoelectronic component, or an optically transparent electromagnetic surface.

Several limitations recur across otherwise dissimilar OptoSkin systems. Calibration is frequently device- or subject-specific: the OCT attenuation model for cuffless blood-pressure inference still requires direct calibration of the pressure-to-attenuation coefficient $$I_{\rm meas}(z,t)=I_0\exp\!\left[-(\mu_0+k\Delta P(t))z\right].$$4 (Martinelli et al., 2018), the edge-lit PDMS tactile sensor requires data-driven calibration per device (Piacenza et al., 2018), and the external photoreflective module depends on elastomer-specific force–indentation and voltage–indentation fits (Yamamoto et al., 9 Nov 2025). Material selection is likewise central: speckle-based sensing depends on scatterer distribution and optical loss control (Shimadera et al., 2022), direct ToF sensing depends on the scattering coefficient of the waveguide (Aulika et al., 30 Jul 2025), and event-based stereo localization is currently formulated for single-point contacts, with multi-touch left as a future problem of stereo correspondence among multiple clusters (Koolani et al., 7 Jan 2026).

A final methodological point is that OptoSkin research repeatedly favors continuous media over discretized taxelization. Whether the medium is epidermis and dermis, a PDMS slab, a semitransparent gel, a silicone waveguide, or a fiber-embedded elastomer, information is distributed across an optical field rather than localized to independently wired pressure pixels. This suggests that the most stable definition of OptoSkin is not a specific sensor topology, but a class of systems in which optical propagation through skin or skin-like matter is the primary information carrier.