Biologically Inspired Layered Sensing

• Anatomically Inspired Layered Multimodal Sensing is a design paradigm that mimics biological tissue layers by organizing sensing elements across stratified materials to capture multiple physical signals.
• It assigns distinct sensing roles to different layers, integrating modalities such as vision, touch, and physiological monitoring through carefully engineered material couplings.
• Recent implementations in robotics and wearables have demonstrated enhanced accuracy in object recognition, force estimation, and multisensor fusion, underscoring its practical impact.

Anatomically inspired layered multimodal sensing is a class of sensing architectures in which material strata, geometry, and transduction pathways are organized to emulate biological tissues—most often skin, fingertips, the ear canal, or related mechanosensory interfaces—so that a common compliant body can encode multiple signal types. Across recent work, this paradigm appears in visuotactile robot fingertips, selectively transmissive membranes, optical skins, in-ear physiological monitors, metamaterial neck interfaces, exoskeleton sleeves, and robotic palpation probes. The unifying premise is that biological function is not reproduced by adding modalities in parallel alone, but by assigning different sensing roles to different layers and coupling them through shared mechanics, optics, or both (Hogan et al., 2020, Goverdovsky et al., 2016, Shimadera et al., 2022, Fan et al., 2024, Lambeta et al., 2024, Xu et al., 26 Nov 2025, Tang et al., 16 Aug 2025).

1. Biological templates and anatomical mapping

A recurrent design pattern is the explicit mapping of engineered layers onto epidermal, dermal, and subcutaneous analogues. In ViTacTip, a transparent outer “epidermal” skin of Agilus30 Clear, an underlying silicone-based gel, and an array of biomimetic pyramidal pins are presented as analogues of transparent epidermis, soft dermis, and mechanoreceptive dermal papillae; the black pin tips act as visual markers that amplify local deformation (Fan et al., 2024). The later benchmarked ViTacTip formulation sharpens this analogy by describing the transparent outer skin as stratum-corneum-like, the intermediate gel as a dermal or subcutaneous coupling layer, and the pin-shaped markers as dermal papillae that amplify shear and normal deformation (Zhang et al., 4 Jan 2025).

The same anatomical logic appears in artificial fingertips. The multimodal fingertip of “Digitizing Touch with an Artificial Multimodal Fingertip” is organized as an ultrathin reflective epidermis-analogue, a compliant gel dermis-analogue, and a rigid substrate carrying optics, thermal, acoustic, chemical, and edge-AI electronics (Lambeta et al., 2024). HumanFT likewise adopts a thermochromic and reflective coating over a bulk PDMS body and a sensor-bearing PCB, explicitly using the soft outer coating for optical and thermal functions and the thicker inner elastomer for force propagation (Wu et al., 2024).

Outside robotic fingertips, anatomical inspiration is more site-specific. Hearables exploit the ear canal’s approximately 24–26 mm length, tapered geometry, compliant cartilage, and proximity to peri-auricular vessels; the earpiece uses a deformable memory-foam cylinder and a layered stack consisting of plastic film, conductive cloth electrode, embedded microphone diaphragm, and foam substrate (Goverdovsky et al., 2016). The leg sleeve for exoskeleton interaction defines skeletal, muscular, and cutaneous sensing layers rather than literal material skin layers: IMUs are assigned to skeletal-level kinematics, textile sEMG to muscular activity, and textile strain sensors to skin deformation (Tang et al., 16 Aug 2025). BMMI extends the biological analogy to skin microrelief and mechanoreceptor diversity by using star-shaped plateau-and-furrow metamaterial geometry to emulate microrelief, high-strain FS-2 regions to emulate Pacinian-like vibration sensing, and relaxed FS-1 regions to emulate Ruffini-like stretch sensing (Xu et al., 26 Nov 2025).

Taken together, these systems suggest that “layered” does not denote packaging alone. A plausible implication is that the central design variable is functional segregation by depth or region: optical access near the surface, deformation coupling in the bulk, and more stable or higher-bandwidth transduction deeper in the stack.

2. Layered architectures and physical transduction

Several distinct architectural families recur in the literature.

Family	Representative stack	Encoded modalities
See-through visuotactile skins	Transparent or semitransparent membrane, internal illumination, embedded or reflective layer, camera beneath	Vision, touch, proximity
Marker-amplified fingertips	Transparent skin, soft gel, biomimetic pins or markers, internal camera and LEDs	Contact geometry, force, pose, texture
Physiological wearables	Conformal foam/textile/elastomer layers aligned to tissue zones	EEG, mechanical pulse, respiration, sEMG, strain, kinematics
Metamaterial and optical skins	Engineered scatterers or auxetic sublayers with optical or piezoresistive readout	Force, location, temperature, vibration, posture

In STS, the stack is a compliant gel of approximately 5 mm thickness, a half-silvered coating, a protective silicone overlay much thinner than 0.1 mm, and a rigid acrylic backing with the camera mounted approximately 5 mm underneath. Its defining mechanism is illumination-controlled duality. In camera-mode, when internal LEDs are off or dim, the membrane acts as a window and the image is modeled as

$$I_{\mathrm{vis}}(x,y)=\int_\lambda T(\lambda)\,E_{\mathrm{ext}}(\lambda,x,y)\,S_{\mathrm{cam}}(\lambda)\,d\lambda \;\;\circledast\;\; h(x,y)+\eta .$$

In tactile-mode, bright internal LEDs make the same half-silvered layer behave as an opaque diffuse mirror, while contact-induced deformation modulates reflected intensity through a Phong reflectance model. The force model is written as $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$ and $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$ (Hogan et al., 2020).

A related but not identical strategy appears in the selectively transmissive soft membrane for simultaneous visuotactile and proximity sensing. Here, three Ecoflex-based silicone layers produce high transmission at 860 nm for the ToF channel, near-opacity in the visible band, and phosphorescent fiducials for visuotactile imaging. The membrane is approximately 1.0 mm thick, inflated to approximately 0.02 PSI, and paired with an internal RGB camera, a MEMS pressure sensor, and an infrared ToF depth camera (Yin et al., 2022). CompdVision also preserves optical transparency but avoids mode switching: a transparent elastomer with an embedded marker layer is placed over a compound-eye imaging subsystem with dual-focus microlenses, so that far-focus stereo units observe external surfaces while near-focus tactile units observe marker motion simultaneously at 25 Hz (Luo et al., 2023).

Other systems encode multiple variables without discrete sensor integration at the surface. Optical skin uses a single 5 mm-thick transparent silicone slab containing microscopic impurities and entrapped air bubbles as scattering centers. A He–Ne laser illuminates the slab, a camera images the underside speckle, and deformation or thermo-optic variation modulates the interference pattern through

$I(r)=\left|\sum_k A_k e^{i\phi_k}\right|^2 .$

This architecture supports simultaneous decoding of contact force, contact location, and temperature from a single speckle image (Shimadera et al., 2022).

BMMI replaces optical transparency with geometry-enabled strain engineering. It combines a high-modulus PDMS FS-1 auxetic backbone, low-modulus PDMS FS-2 infill, Ecoflex FS-3 contact layer, a graphene nanoplatelet sensing film, and polyimide isolation frames. The auxetic mechanism is parameterized through rotating star-shaped unit cells, with negative Poisson’s ratio emerging as the modulus ratio $E_1/E_2$ increases; finite-element analysis reports a transition from $\nu\approx+0.47$ to $\nu\approx-0.05$ as $E_1/E_2$ rises from 1:1 to 144:1 (Xu et al., 26 Nov 2025).

The physiological wearables rely on a comparable logic of layered coupling. Hearables use a 50 μm plastic barrier, conductive cloth electrode, microphone diaphragm, and viscoelastic foam, whereas the exoskeleton sleeve combines ultra-thin textile electrodes, screen-printed graphene strain sensors, and sewn IMU modules in a compression sleeve (Goverdovsky et al., 2016, Tang et al., 16 Aug 2025).

3. Fusion, switching, and decoding algorithms

The algorithmic layer is as central as the physical layer. In STS, visual and tactile frames are resized to 224×224 px and processed by two independent ResNet-50 backbones pretrained on ImageNet, producing $f_{\mathrm{vis}}\in\mathbb{R}^{4096}$ and $f_{\mathrm{tac}}\in\mathbb{R}^{4096}$. These are concatenated and passed through a multilayer perceptron:

$p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$0

Training uses cross-entropy, Adam, learning rate $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$1, batch size 100, and approximately 100 epochs (Hogan et al., 2020).

Hearables employ a different fusion philosophy. The physiological estimate is formulated as probabilistic weighted least squares, $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$2, with an equivalent Bayesian expression using modality-specific covariance matrices. Artifact rejection is explicitly cross-modal: the mechanical channel acts as an adaptive LMS reference for ear-EEG denoising, followed by multivariate empirical mode decomposition and band-pass filtering tailored to EEG, PPG, or respiration (Goverdovsky et al., 2016).

ViTacTip and its 2025 benchmark combine physical multimodality with computational modality conversion. DenseNet121 is used for image-based regression and classification tasks, while Pix2Pix conditional GANs perform marker removal and light removal. The two generators map ViTacTip images either to a marker-free ViTac style or to a TacTip-like opaque tactile style, using adversarial and $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$3 reconstruction losses; the 2025 benchmark further reports optional perceptual loss and typical weights $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$4 and $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$5 (Fan et al., 2024, Zhang et al., 4 Jan 2025). This is a notable shift from illumination-only mode control toward learned cross-modality interpretation.

Optical skin uses a shared feature extractor on a single 64×64 greyscale speckle image, followed by one decoder branch per modality and an MSE loss over the three regression targets. CompdVision demultiplexes a single frame into stereo and tactile tiles, with the stereo path using rectification, SGBM, WLS filtering, and $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$6, while the tactile path uses marker tracking, cubic interpolation to a 6×6 dense flow field, and a CNN regressor for $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$7 (Shimadera et al., 2022, Luo et al., 2023).

More recent systems broaden the decoder repertoire. BMMI’s CA-Net uses four parallel 1D-convolution branches, concatenation, two Transformer layers without positional embeddings, and three parallel softmax heads for emotion, attitude, and attention (Xu et al., 26 Nov 2025). The human-inspired soft hand converts force, acceleration, stretch, and temperature into spike trains using Izhikevich, Poisson, and thermoreceptor-inspired encoders, then fuses 145 channels in a five-layer feed-forward SNN trained with SLAYER (Wang et al., 2 Sep 2025). The artificial multimodal fingertip integrates MobileNetV2, ResNet-18, and an MLP on a GAP9 accelerator, reporting total latency of approximately 1.2 ms on-device and bandwidth reduction from 148 MB/s peak video to approximately 0.1 MB/s through local feature-vector transmission (Lambeta et al., 2024).

A plausible implication is that layered multimodal sensing has produced three distinct computational regimes: explicit analytic fusion, deep shared-latent fusion, and learned modality translation.

4. Robotic manipulation, metrology, and tactile perception

The empirical record in robotic sensing is broad and quantitatively strong.

System	Tasks	Representative reported results
STS	Object recognition, textures, fullness metrology	96.9% visuotactile accuracy in simulation; fusion achieved ≥ 91% recall on every bottle class; ≈95%+ bottle fullness correctness (Hogan et al., 2020)
ViTacTip	Grating, pose, contact localization, force	99.72% grating accuracy; minimum error 0.08 mm and 0.03 N; contact and force errors 60% and 52% better than TacTip (Fan et al., 2024)
ViTacTip benchmark	Object recognition, proximity, multi-task classification	99.91% object recognition; hardness 97.47%, material 98.81%, texture 97.78%; absolute proximity error ~0.5 mm within 0–18 mm (Zhang et al., 4 Jan 2025)
CompdVision	Near-field 3D vision and tactile force	Median depth error < 0.2 mm; temporal noise < 0.15 mm RMS; force RMSE 0.17 N, 0.18 N, 0.26 N (Luo et al., 2023)
PVFT	Subsurface feature detection under palpation	Tactile-only F1 1.00, force-only F1 0.78, combined F1 0.98; depth-estimation error $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$8 mm vs $p(x,y)=k_{\mathrm{pixel}}\delta(x,y)$9 mm from force alone (Ren et al., 24 Dec 2025)
HumanFT / artificial fingertip	Force, vibration, temperature, tactile imaging	HumanFT RMSE 0.4 N, 0.6 N, 0.8 N; artificial fingertip resolved 7 μm spatial features and 1.01 mN normal-force resolution (Wu et al., 2024, Lambeta et al., 2024)

STS established the shared-aperture visuotactile template. In simulation on 10 ShapeNet categories with 1,200 images per category, vision only reached approximately 88.8% accuracy, tactile only approximately 83.1%, and visuotactile fusion approximately 96.9%. On real bottles, the paper reports confusion in the unimodal baselines and at least 91% recall on every bottle class for fusion. On 3D-printed black texture samples, vision alone was near chance at approximately 17%–20%, tactile reached at least 80% on 5/6 samples, and fusion matched tactile performance. For bottle fullness, vision alone was approximately 60% correct, tactile approximately 93%, and fusion approximately 95%+ with mis-predictions only to adjacent levels (Hogan et al., 2020).

ViTacTip extends this line by embedding biomimetic pins within a transparent fingertip. Its initial verification reports 99.72% accuracy in grating identification, pose regression errors down to 0.08 mm, and force estimation error down to 0.03 N, while the 2025 benchmark broadens evaluation to 21 object classes, 25,000 hardness-material-texture samples, and proximity estimation with mean MSE(d) approximately $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$0 for textured objects and absolute error around 0.5 mm within 0–18 mm (Fan et al., 2024, Zhang et al., 4 Jan 2025).

CompdVision emphasizes compactness and simultaneity rather than switching, reporting depth range 11–70 mm, median depth error below 0.2 mm, and force RMSE of 0.17 N, 0.18 N, and 0.26 N on a 16,000-indentation dataset (Luo et al., 2023). The selectively transmissive soft membrane adds dynamic manipulation use-cases: approach-and-contact, catching, and throwing, with best mean depth error of approximately 1 mm at 50 mm, worst mean error approximately 31 mm at 100 mm, and average size error approximately 4.3% on an object dataset (Yin et al., 2022).

At the high-performance end, the artificial multimodal fingertip reports approximately 8.3 million taxels, 7 μm spatial resolution, 1.01 mN normal-force resolution, 1.27 mN shear-force resolution, vibration sensing up to 10 kHz, odor classification at 91% across six household substances within 90 s, and 1.2 ms reflex-loop latency on-device (Lambeta et al., 2024). HumanFT, by contrast, emphasizes compact humanoid-finger form factor, with pressure-sensor-based force reconstruction yielding $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$1 and RMSE of 0.4 N in $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$2, 0.6 N in $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$3, and 0.8 N in $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$4, plus a binary overtemperature alert latency below 1.2 s for $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$5 (Wu et al., 2024).

PVFT demonstrates that multimodality remains important even when one modality is already tactile. In robotic physiotherapy palpation, force profiles alone were frequently ambiguous, whereas tactile images revealed clear structural differences in presence, diameter, depth, crossings, and multiplicity; combined sensing maintained force-tracking RMSE of 7.04% at 25 N during sliding (Ren et al., 24 Dec 2025).

5. Physiological and body-interfaced sensing

In physiological sensing, anatomical layering is tied directly to tissue coupling and artifact rejection. Hearables use the ear canal as a mechanically stable, socially unobtrusive site for electrical and mechanical sensing. Reported results include ear EEG ASSR SNR of approximately 6.2 dB versus 4.8 dB uncorrected, heart-rate RMSE below 1.2 bpm versus 2.8 bpm from unfiltered MMS, sleep-stage Cohen’s $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$6 with sensitivity and specificity above 85% for Wake versus Sleep, Pearson correlation $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$7 between mechanical plethysmography and finger PPG, and 75–85% attenuation of jaw-clench artifacts using adaptive removal with the mechanical channel as reference (Goverdovsky et al., 2016). The paper also presents optical PPG as an extension rather than as an implemented core modality.

BMMI addresses a different physiological regime: heterogeneous mechanodermal activity over adjacent skin regions. By using dual-modulus auxetic sublayers, it separates a high-gain vibration channel from a low-gain stretch channel. For the D3 design with modulus ratio $F_{\mathrm{total}}=\iint_A p(x,y)\,dA$8, the measured gauge factors are approximately 405 in FS-2 and approximately 30 in FS-1 under 1% strain; detection limits are 0.01% strain for BMMI-S and 0.1% strain for BMMI-L; electrical bandwidth reaches 300 Hz; and durability exceeds 10,000 stretch-release cycles at 10% strain with less than 5% drift in gauge factor. In a dataset of 1,200 recordings from 5 subjects, CA-Net achieved 95% emotion, 100% attitude, and 99% attention classification accuracy, while single-channel or reduced-modality inputs dropped performance by 5–15% (Xu et al., 26 Nov 2025).

The smart leg sleeve applies layered sensing to human–exoskeleton interaction. Two MPU-9250 IMUs, three dry textile sEMG electrodes over each of tibialis anterior, fibularis brevis, and medial gastrocnemius, and graphene-based textile strain sensors along the posterolateral heel are embedded into a compression sleeve weighing less than 20 g including sensor and electronics. Leave-one-subject-out evaluation reports ankle joint moment estimation RMSE of 0.133 ± 0.015 Nm/kg, metabolic trend classification accuracy of 97.1%, and injury-risk detection recall of 96.4%, with latency below 100 ms in 95% of cases (Tang et al., 16 Aug 2025).

Optical skin demonstrates that large-area body-interfaced sensing need not rely on dense wired arrays. Using a single soft slab and speckle decoding, it reports indentation-depth resolution of ±3.95 μm, lateral resolution of ±37.3 μm, temperature resolution of ±0.23 °C, latency of approximately 200 ms per frame, shape classification of approximately 94%, and 100% classification accuracy for four touch zones in a haptic human–machine interface (Shimadera et al., 2022).

These studies indicate that the same layered principle can support both external-environment perception and internal-state inference, depending on whether the interface is optimized for contact mechanics, electrophysiology, vascular coupling, or strain localization.

6. Design tensions, recurring misconceptions, and open directions

A recurring misconception is that multimodal sensing is equivalent to placing several independent sensors in one package. The literature shows multiple alternatives. Optical skin explicitly avoids large-scale sensor integration by encoding several physical quantities into a spatial optical interference pattern in a single soft material (Shimadera et al., 2022). STS uses one camera, one field of view, one pixel grid, and illumination control to obtain both vision and touch (Hogan et al., 2020). CompdVision achieves simultaneous near-field 3D vision and tactile sensing through focal diversification within one compound-eye optical stack (Luo et al., 2023). BMMI separates signal classes by local strain engineering inside one metamaterial substrate rather than by adding isolated transducers (Xu et al., 26 Nov 2025).

A second misconception is that transparency or optical access necessarily weakens tactile sensing. The available evidence does not support such a blanket claim. ViTacTip reports 99.72% grating identification and lower pose and force errors than its baseline counterparts, while STS reports systematic improvements over vision-only and tactile-only baselines in simulation and experiments (Fan et al., 2024, Hogan et al., 2020). At the same time, the literature also shows that no single strategy dominates. Some systems rely on illumination-controlled switching, some on GAN-mediated interpretation, and some on simultaneous optical multiplexing (Zhang et al., 4 Jan 2025, Yin et al., 2022, Luo et al., 2023).

A third misconception is that force alone is an adequate descriptor of interaction in deformable environments. PVFT provides a clear counterexample: force signals alone frequently produced ambiguous responses in soft-tissue phantoms, whereas tactile images revealed depth, multiplicity, and crossing geometry; fusion then preserved controlled palpation while retaining structural discrimination (Ren et al., 24 Dec 2025).

The major unresolved design tensions concern calibration burden, data requirements, robustness, and embodiment. Optical skin notes the need for hundreds to thousands of labelled speckle images and sensitivity to environmental drift (Shimadera et al., 2022). The exoskeleton sleeve notes the need for expert anatomical placement and points to long-term robustness under heavy wash cycles as future work (Tang et al., 16 Aug 2025). STS proposes curved or dynamically shaped surfaces, active display layers, and multiphoton imaging through the gel for sub-surface sensing (Hogan et al., 2020). ViTacTip proposes deeper “hypodermis” layers, tunable pin geometry, and cycle-consistent GANs (Fan et al., 2024). CompdVision suggests more lens types, including a third “proximity” focus, while the artificial fingertip foregrounds embedded neural accelerators and reflex-like local control loops (Luo et al., 2023, Lambeta et al., 2024).

Taken together, these directions suggest that the field is moving from multimodal co-location toward anatomically stratified, computation-aware interfaces in which geometry, mechanics, optics, materials, and learning are jointly designed.