A robotic sensory system with high spatiotemporal resolution for texture recognition

Robotic technologies increasingly require tactile sensation for safe and dexterous interaction. Existing flexible tactile sensors can precisely detect pressure, shear, and strain, but often lack the ability to perceive and recognize objects during touch. In contrast, human skin employs slow-adapting receptors for static pressure and fast-adapting receptors for high-frequency vibrations, enabling object identification during gentle sliding (e.g., reading Braille or distinguishing textiles). Achieving both ultrahigh sensitivity (to interact with micron-scale features) and rapid response-relaxation speed (to resolve close spacings and high-frequency vibrations) in a single flexible sensor remains challenging. Many artificial sensory systems therefore rely on two sensors and separate circuits for static and dynamic detection. Furthermore, the link between sensing performance and recognition capability has not been fully clarified, with frequency range often emphasized over temporal (frequency) resolution, which is crucial for discrimination. This study reports a single-sensor, real-time sensory system for texture recognition and introduces spatiotemporal resolution as a criterion correlating sensing performance with recognition ability.

Prior work on flexible tactile sensors has focused on precise detection of physical stimuli (pressure, shear force, strain) to improve robotic manipulation and feedback. Biological insights highlight the complementary roles of slow- and fast-adapting mechanoreceptors for static and vibrational cues in haptic perception. Existing fingertip-like sensing platforms struggle to simultaneously achieve ultrahigh sensitivity and rapid response needed for discriminating fine textures; consequently, systems frequently combine two different sensor types (e.g., piezoresistive with piezoelectric or triboelectric sensors) and dual acquisition circuits to emulate SA and FA sensing. Earlier studies emphasize wide frequency ranges, but the importance of high temporal (frequency) resolution for discriminating closely spaced surface features has been under-discussed. The present work positions a single iontronic slip-sensor and a spatiotemporal resolution criterion to address these gaps, contrasting with multi-sensor approaches that increase system complexity and integration burden.

Sensor design and materials: The slip-sensor comprises an artificial PDMS fingerprint, an ionic gel layer of PVA–H3PO4 (8.3 wt%), two Au-on-PET flexible electrodes, and a flat PDMS encapsulation. The fingerprint uses concentric elliptical ridges (triangular ridge height ~260 μm, spacing 350 μm), with a tip width set to 13 μm to effectively couple with fine textures. The ionic gel features graded microstructures: periodic domes (~200 μm diameter, ~55 μm height) densely covered with finer protrusions. Mechanism and properties: The sensor leverages tunable electric double layers (EDLs) with nanoscale charge separation for capacitive readout, yielding ultrahigh capacitance-to-pressure sensitivity (up to 519 kPa−1), high spatial resolution (detecting 15 μm spacing and 6 μm height features), and fast response to vibrations up to 400 Hz with high frequency resolution (~0.02 Hz at 400 Hz). Rapid response-relaxation arises from materials and structure: a stiff ionic gel (E ≈ 5.5 MPa) lowers viscosity and interfacial adhesion; graded microdomes with fine protrusions reduce contact area, lowering adhesion energy and increasing energy release rate; small ionic radii (H+ and inorganic anions) facilitate fast ion migration. Interfacial adhesion tests (loading 50 kPa, hold 3 min, then release) show negligible adhesion for graded PVA–H3PO4 (E ~5.5 MPa), versus 1.8 kPa for flat PVA–H3PO4 (E ~5.5 MPa), 27 kPa for flat PVA–H3PO4 (E ~2.0 MPa), and 122 kPa for flat PDMS (E ~1.5 MPa). Spatial and temporal resolution characterization: With a 13 μm ridge tip, the sensor detects surface textures with spacings ≥ tip size (e.g., 15 μm, 50 μm) but not 10 μm; increasing tip width to 25 μm prevents detection of 20 μm spacing while allowing 30 μm. Signal amplitude increases with contact pressure and structure height; at 50 μm spacing, amplitudes decrease as height reduces from 50 to 30 to 10 μm. The sensor detects features with 15 μm spacing and 6 μm height. Temporal resolution was quantified by applying vibrations at 400.0, 400.1, and 400.2 Hz and resolving distinct peaks with FWHM ~0.02 Hz (~0.005% at 400 Hz). Texture-frequency relation: The characteristic vibration frequency during sliding follows f = v/l, validated on a textile with l ≈ 275 μm at v = 2, 20, 100 mm s−1 yielding ~7, 73, and 365 Hz, respectively. A base pressure of ~50 kPa during sliding maintains intimate contact and enables high-frequency capture at high sliding rates (≥100 mm s−1). Electronics and signal acquisition: A custom readout board includes power, STM32 microcontroller (central processing), I/O interfaces, signal conditioning, and 24-bit ADC. A 12-bit DAC in the STM32 generates a stable sine excitation; a reactance-to-voltage conversion circuit and low-pass filter produce a DC voltage proportional to capacitance, which is digitized by the ADC for capacitance computation. Board size: ~5.5 cm × 3.5 cm. Machine learning workflow: For offline classification, a Bagging ensemble (base learners: KNN, random forest, logistic regression, decision tree) was used. Features were extracted from time-series signals using tsfresh (statistical, frequency-domain, autoregressive, wavelet/time-domain features). t-SNE visualized high-dimensional embeddings. Data splitting used block-wise cross-validation (each category with 100 sets, divided into five blocks; one as test, others as train, iterated). Real-time system and interface: A portable system integrates the slip-sensor on a prosthetic or human fingertip, the circuit board for acquisition, and a PC-based visual interface. For real-time inference, features are extracted (tsfresh) and random forest classifiers applied for immediate predictions; total preprocessing and inference time <20 ms per data point. The UI displays real-time predicted labels, confidence, waveform overlays, and microscopic images. Experimental datasets: Fixed-rate dataset: 20 textiles (various materials including wool, cotton, blends) with similar structure periods; sliding distance 40 mm at 2 mm s−1; 2000 data sets collected with ~50 kPa base pressure. Higher-rate tests at 40 mm s−1 also conducted. Random-rate dataset: slip-sensor mounted on index fingers of three subjects; unconstrained sliding with unknown pressures and rates (estimated 0–30 mm s−1), collecting before-touch, touch, slide, and withdrawal phases; IMU verified chaotic acceleration indicating variable rates. Each textile category: 400 sets (combined across subjects, 2:1:1), totaling 8000 sets for 20 categories; 40% reserved for testing. Fabrication details: Fingerprint molds and other microstructures were 3D printed (NanoArch S130). PDMS (Sylgard 184, 5:1 base:curing agent) was cast and cured at 80 °C for 30 min to form ~350 μm thick fingerprints. Ionic gel: PVA (2 g in 20 g DI water) dissolved at 90 °C for 2 h, cooled to 50 °C; H3PO4 (1.5 mL, ≥85%) added and stirred 1 h; cast on inverse-graded resin molds and cured 24 h at 24 °C, 43% RH; films ~120 μm thick, die-cut to 7 mm diameter. Electrodes: 100 nm Au sputtered on 40 μm PET, die-cut to 7 mm diameter. Device stack: PDMS fingerprint / PET–Au top electrode / PVA–H3PO4 graded microdome layer / PET–Au bottom electrode / PDMS encapsulation. Additional characterizations: Vibration generator applied constant-frequency vibrations under ~10 kPa normal load; responses measured by LCR meter. Electrochemical etching (1.5 V, 10 s; pressures 10–100 kPa) assessed contact area evolution using a PET–Au bottom electrode, Cu top electrode, and the ionic gel electrolyte.

- Introduced spatiotemporal resolution as a key criterion linking sensing performance to recognition capability in tactile systems. - Single iontronic slip-sensor responds to both static and dynamic stimuli (0–400 Hz), eliminating the need for dual sensors and circuits. - High spatial resolution: reliably detects surface features with 15 μm spacing and 6 μm height; spatial resolution surpasses human fingertips (human subjects averaged 24.3% accuracy with Kappa 0.12 for microstructures ≤50 μm, accuracy ≤44% in all such cases). - High temporal (frequency) resolution: resolves 400.0, 400.1, 400.2 Hz with FWHM ≈ 0.02 Hz (~0.005% at 400 Hz). - Sensitivity: ultrahigh capacitance-to-pressure sensitivity up to 519 kPa−1 via EDL-based iontronic mechanism and graded microstructures. - Texture recognition performance: 20 textiles classified with 100.0% accuracy at a fixed sliding rate of 2 mm s−1; 99.5% at 40 mm s−1; 98.9% accuracy under random sliding rates and contact pressures (8000 data sets across 20 categories). Real-time system achieved 100.0% average accuracy at fixed rate and 98.6% at variable rates with <20 ms inference time. - Robust high-speed operation: maintains intimate contact and captures high-frequency vibrations at sliding rates up to at least 100 mm s−1 (e.g., frequencies ~7, 73, 365 Hz for v = 2, 20, 100 mm s−1 on l ≈ 275 μm). - Electronics performance: high SNR (86.79 dB) and effective number of bits (ENOB ≈ 14.12) support precise signal capture. - Fingerprint’s necessity: removing the artificial fingerprint reduced recognition accuracy to 54.5% at 2 mm s−1. - Amplitude dependence: signal amplitude increases with contact pressure and feature height, and also depends on material stiffness; thus, frequency alone or amplitude alone is insufficient for identification, motivating machine learning feature fusion.

The study addresses the core challenge of texture perception—simultaneously achieving ultrahigh sensitivity and rapid response for both static and dynamic cues—by engineering a single iontronic slip-sensor with graded microstructures and a bio-inspired fingerprint interface. The proposed spatiotemporal resolution criterion captures how fine spatial detectability (15 μm spacing, 6 μm height) and high frequency resolution (~0.02 Hz at 400 Hz) jointly enable discrimination among textures with very close characteristic spacings and frequencies. Integrating the sensor on a prosthetic fingertip and combining it with feature-rich machine learning yields near-perfect recognition at fixed sliding rates and robust performance under random, human-like sliding conditions. These results demonstrate that a single-sensor architecture can match and even surpass human fingertips in resolving fine textures, while simplifying system integration compared with dual-sensor approaches. The artificial fingerprint is shown to be critical for coupling microfeatures at high speed, and the electronics deliver high SNR and ENOB to preserve subtle tactile information. Overall, the findings validate that optimizing spatiotemporal resolution in tandem with robust feature extraction/classification is key to reliable touch-based object recognition in robotics and prosthetics.

This work presents a single iontronic slip-sensor–based artificial sensory system capable of capturing both static pressure and high-frequency vibrations with exceptional spatiotemporal resolution. The sensor’s graded microstructures and bio-inspired fingerprint enable detection of 15 μm spacing and 6 μm height features and resolve frequency differences of ~0.02 Hz at 400 Hz. Integrated with a compact readout circuit and machine-learning pipeline, the system recognizes 20 textiles with 100.0% accuracy at fixed sliding rates and ~99% under random human-like conditions; real-time inference (<20 ms) supports immediate, visual feedback. The approach simplifies hardware (single sensor and circuit) while improving accuracy and robustness relative to prior dual-sensor systems. Future directions include expanding to broader material classes and contact conditions, integrating with complex robotic hands and prosthetic systems for closed-loop manipulation, leveraging onset-of-slip and frictional features for improved classification, and extending applications to haptics-based VR/AR and sensory restoration.

- Spatial detectability depends on fingerprint tip size: with a 13 μm tip, structures with 10 μm spacing were not resolved; increasing tip width to 25 μm further limits detection of smaller spacings (e.g., 20 μm), indicating a trade-off between robustness and spatial resolution. - Performance depends on contact mechanics: maintaining moderate normal pressure (~50 kPa) was used to ensure intimate contact at higher sliding rates; variations in contact pressure and texture stiffness affect signal amplitude. - Recognition relies on the artificial fingerprint: removing the fingerprint reduced accuracy to 54.5%, indicating system dependence on structured coupling. - Dataset scope: validation focused on 20 commercial textiles with relatively close structure periods; generalization to broader materials, environmental conditions (humidity, temperature), and long-term durability were not detailed in the main text. - Frequency range characterized up to 400 Hz; behavior beyond this range was not reported.