Spinal-inspired artificial tactile interneuron with high-order burst spiking for intelligent edge interfaces Authors: Fanfan Li, Zhanglu Yan, Jiayi Mao, Guolei Liu, Huihui Ren, Bangbang Qin, Zhongfang Zhang, Haiyue Zhang, Yiyang Shen, Zeqi Zheng, Weilong Feng, Dingwei Li, Yingjie Tang, Saisai Wang, Yaochu Jin, Tao Luo, Weng-fai Wong, Hong Wang, Bowen Zhu Published: 2026 ABSTRACT AbstractHuman tactile perception relies on hierarchical processing, where inputs entering the nervous system are fused by interneurons for sparse multimodal encoding, and the integrated signals are sent to the brain to generate perception. Replicating this pathway from primary sensory inputs to higher-order neural processing, which efficiently transforms signals into coherent representations of the external environment, is essential for artificial tactile systems. Here we present artificial multimodal interneuron (AMINs) by integrating strain, pressure, and temperature sensors with NbOx memristor neurons on a hybrid integrated platform, enabling hierarchical neural encoding and the generation of high-information-density temporal spike patterns. AMIN-based processing generates a unified burst spike train with wide temporal dynamics that encode object size, hardness, and temperature, serving directly as input to SNNs. In a 20-class tactile object recognition task, a hardware-characterized AMIN encoding model combined with a software SNN achieves 90.5% accuracy, demonstrating the potential of the proposed tactile encoding strategy for compact and low-power multimodal tactile intelligence.Subject terms: Electrical and electronic engineering, Electronic devices Introduction The human tactile system integrates multimodal sensory information to enable efficient and robust perception in complex dynamic environments, supporting accurate recognition of object properties and states1,2. This capability arises from the hierarchical processing of external stimuli by the sensory nervous system3. In the tactile perception process, primary sensory neurons encode tactile stimuli into periodic electric spikes, while spinal interneurons fuse and integrate this information into sparse, structured high-order spiking patterns that abstract tactile information4,5. These integrated signals are ultimately relayed to the brain to form unified perceptual representations for object recognition and classification6,7. Replicating this hierarchical integration and feature abstraction is crucial for artificial tactile systems aiming to match biological perception8,9. Many attempts have been made to emulate the multimodal functionality of biological tactile neurons. Existing artificial tactile systems integrate independent sensing elements and employ first-order spike encoding to convert analog inputs into separate spike trains10–16. This approach, however, requires multiple external measurement and readout pathways, resulting in increased systemic complexity and higher readout power consumption17–19. Although some progress has been made toward integrating multimodal sensing within a single primary sensory neuron, the prevalent strategies still rely on using spike frequency and spike amplitude to encode different tactile modalities (Supplementary Tables 1 and 2)20–22. This reliance on two distinct decoding pathways increases processing complexity and further deviates from the time-based signaling of artificial neurons23. To emulate high-order interneuron sensory processing in biological systems, we develop artificial multimodal interneurons (AMINs) that reproduce the integration and sparse coding mechanisms of spinal interneurons, enabling high-order tactile feature extraction encoded in a single sparse spike train. Specifically, the system integrates vertically aligned gold nanowires (v-AuNWs) strain sensors, polypyrrole (PPy) based pressure sensors, temperature sensors, and NbOx memristor neurons on a hybrid integrated platform, achieving modality sensing and fusion without analog-to-digital conversion (ADC). To ensure signal separation, three independent spike encoding channels provide broadband output from 2.9 kHz to 3.98 MHz with clear spectral separation between modalities to eliminate signal crosstalk. As a result, the fused signals (Vfusion) increase information density and are converted by the AMIN into sparse burst-spike patterns with excitatory or inhibitory polarity. The multi-timescale temporal features embedded in these bursts enable discrimination of object size, hardness, and temperature. To demonstrate system-level capability, a memristor-based decision module combined with a spiking neural network (SNN) incorporating brain-inspired temporal processing is employed for complex tactile feature recognition. In a 20-class object grasping task, the AMINs-SNN system achieves 90.5% accuracy and exhibits effective modality decoupling and robust operation. Importantly, by combining biological interpretability with hardware compactness, AMINs provides a practical route for neuromorphic robotics, virtual reality, and wearable tactile interfaces. Results Design of AMINs In the human tactile system, higher-order interneurons integrate spike timing patterns from primary sensory neurons to generate tactile features that support object recognition and classification. As illustrated in Fig. 1a, proprioceptors sense muscle strain and convey information about finger angles, while mechanoreceptors and thermoreceptors transmit information about pressure and temperature during object interactions. During this process, primary sensory neurons convert external stimuli into bioelectrical signals, generating electric spikes at different rates based on the intensity of the external stimulus24. Subsequently, interneurons integrate these signals and generate sparse, structured temporal spike patterns that convey higher-order tactile features for object recognition25. Hierarchical processing and artificial multimodal interneuron (AMIN) for tactile feature encoding. a Schematic of tactile encoding in the biological sensory pathway. Afferent neurons convert strain, pressure, and temperature stimuli into modality-specific spikes, which spinal interneurons integrate into burst patterns representing higher-order tactile attributes, such as size, stiffness, thermal state, and grasp-related condition for subsequent recognition in the brain. b Schematic of the AMIN system that integrates v-AuNWs strain, PPy pressure and temperature sensors with NbOx memristor-based neurons on a flexible printed circuit board. Multimodal tactile signals are fused on-chip without analog-to-digital conversion, enabling hierarchical tactile feature encoding for object recognition. c AMINs fuse frequency- and amplitude-modulated spikes from strain, pressure, and temperature sensors into sparse burst patterns whose temporal features are used by a SNN for object recognition. To mimic the functions of biological tactile interneurons, we develop AMINs that integrate v-AuNW strain sensors, PPy-based pressure sensors, and temperature sensors with NbOx Mott memristor-based spiking neurons on a hybrid integrated platform. This architecture enables hierarchical tactile encoding within a compact hardware system (Fig. 1b). The sensors’ outputs are simultaneously converted into spiking signals by the NbOx neuron devices and fused into a compound spike train with combined frequency and amplitude modulation, which is subsequently transformed into sparse burst patterns (Fig. 1c). These burst timing data are then decoded as direct input to a SNN for object recognition. Multimodal spiking encoding with frequency separation in Mott neuron circuits To replicate the functional role of biological spinal interneurons, the first-order AMIN integrates flexible strain, pressure, and temperature sensors with NbOx Mott memristors to fuse heterogeneous tactile inputs (Fig. 2a). The NbOx memristors provide device-level scalability and exhibit rich neural dynamics, enabling hardware realization of spiking sensory neuron behavior and higher-order biomimetic functions26–31. At the core of the system, the NbOx memristor has a layered Pt/Ti/NbOx/Pt/Ti structure, confirmed by the cross‑sectional TEM image (Supplementary Fig. 1), and shows volatile TS behavior with high cycle-to-cycle stability over 1000 sweeps (Fig. 2b). Although the exact switching mechanism remains under debate, the observed stability is commonly linked to the Mott transition in crystalline NbO2 formed during the electrical formation process (Supplementary Fig. 2)32,33. The device shows Joule heating-induced insulator-to-metal transition at Vth ≈ 1.60 V with a compliance current of 500 µA and resets to the insulating state when the voltage drops below Vhold ≈ 1.20 V. The switching endurance of the devices was examined over 1010 pulses (Supplementary Fig. 3). Symmetric TS behavior is observed under both polarities with S-shaped negative differential resistance (Supplementary Figs. 4 and 5)34–36. The NDR behavior is attributed to a positive feedback mechanism in NbO2, where Joule heating increases the local temperature and thereby promotes nonlinear thermally activated transport (Supplementary Fig. 6). Moreover, the memristor has a rapid transient switching speed, switching from a high resistance state (HRS) to a low resistance state (LRS) in ~60 ns and from LRS to HRS in less than 50 ns (Supplementary Fig. 7). The device-to-device variability across 40 devices, with the coefficient of variation (CV) calculated as the ratio of the standard deviation (σ) to the mean (μ) for Vth, Vhold, HRS, and LRS, was 2.40%, 2.36%, 9.59%, and 2.40%, respectively, indicating high uniformity and reliability across the devices (Supplementary Fig. 8). Multimodal spiking encoding and modality-specific frequency separation using Mott neuron circuits. a Schematic of a multimodal tactile platform integrating strain, pressure, and temperature sensors with NbOx Mott memristors. b I–V curves of the memristor showing reproducible threshold switching across 1000 cycles. c Spiking encoder circuit where each sensory modality drives an individual memristor neuron with distinct capacitance. d Output spike frequencies under single-modal stimuli: temperature inputs from 20 to 50 °C produce low-frequency spikes of 2.90–10.00 kHz, pressure inputs from 3.2 to 8.7 kPa produce mid-frequency spikes of 16.70–83.30 kHz, and strain inputs from 70 to 115% produce high-frequency spikes of 1.08–3.88 MHz. Based on this memristor property, we constructed a spiking neuron circuit by combining the NbOx device with series resistive tactile sensors and parallel capacitors to emulate the periodic firing behavior of biological sensory neurons (Fig. 2c). Under a constant bias, the sensor input modulates the charging rate of the node capacitor, triggering the memristor to switch on and generate spikes when the voltage reaches Vth, and resetting to the high-resistance state for discharge when the voltage falls below Vhold, thereby producing periodic output spikes. Moreover, the oscillator maintains stable spiking over 106 cycles with a 5 kΩ series resistor, indicating negligible variability in Vth and Vhold (Supplementary Fig. 9). To assess the effect of device-to-device variation on frequency consistency, 64 NbOx memristors were measured and showed a frequency coefficient of variation of 8.68%, indicating sufficiently consistent frequency output (Supplementary Fig. 10). The time constant τ = RC in each channel defines the spike frequency, mapping input intensity to firing rate (Supplementary Fig. 11). Here, the effective resistance R is directly tuned by the tactile sensors: the micro‑pyramidal PPy film responds to static pressure via resistance changes induced by variations in the contact area between the PPy layer and interdigitated electrodes, operating over 2.5–12.5 kPa; the v‑AuNWs sensor mimics proprioceptors and exhibits resistance changes under 10–120% strain; and the temperature sensor shows resistance variations across 10–70 °C (Supplementary Figs. 12–16)37,38. To avoid cross‑modal interference, the modalities are separated in the frequency domain before fusion (Fig. 2d). This separation is implemented through NbOx memristors, which enable broad frequency‑range encoding. In this scheme, the temperature channel includes a 47 nF capacitor and generates low‑frequency spikes (2.90–10.00 kHz), the pressure channel uses a 4.7 nF capacitor and produces mid‑frequency spikes (16.70–83.30 kHz), and the strain channel relies on parasitic capacitance to deliver high‑frequency spikes (1.08–3.88 MHz). This frequency‑selective multimodal encoding achieves spectral separation of modalities and converts heterogeneous tactile inputs into modality‑specific spike patterns for downstream neural computation. Modality fusion of the spike-based AMIN Spike-level modality fusion integrates sensory inputs early and generates clear temporal features for downstream neural encoding39. As shown in Fig. 3a, a fusion encoding scheme combining frequency and amplitude modulation is implemented to realize this multimodal representation. In this scheme, high-frequency spikes generated by strain and mid-frequency spikes modulated by pressure are merged into the fused output Vfusion to form a compound spike train carrying information from multiple modalities. Multimodal spike output with frequency–amplitude modulation for high-level tactile representation. a Time-domain outputs of the neuron circuit showing high-frequency (VTS1), low-frequency (VTS2), and fused (Vfusion) spike signals under multimodal input (VS = VP = 2.8 V, RS = RP = 10 kΩ, CP = 4.7 nF, CT = 47 nF, R0 = 50 Ω). FFT analysis of Vfusion under varying RS at fixed pressure (b) and varying RP at fixed strain (c), illustrating independent spike-frequency tuning by different modalities. d Robotic grasping sequence and spike outputs. As the gripper approaches, contacts, and senses temperature, the spike train shows distinct changes in frequency and amplitude, reflecting different tactile stages. Spectral analysis shows that the frequency components are independently tunable (Fig. 3b, c). When pressure is constant, increasing strain decreases the high-frequency component while leaving the mid-frequency band stable. Conversely, at fixed strain, raising pressure predominantly modulates the mid-frequency component with minimal influence on the high-frequency band. This separation in the frequency domain preserves the individual signatures of each modality after fusion. To further quantify crosstalk, we compared the frequencies of the three individual channels before and after tri-modal fusion. The relative frequency deviations are only 1.02%, 0.47%, and 5.61% for the strain, pressure, and temperature channels, respectively, confirming preserved modality-specific frequency signatures with minimal band leakage (Supplementary Fig. 17). To clarify the signal pathway in robotic grasping (Supplementary Fig. 18), the corresponding test platform and circuit implementation are shown in Supplementary Fig. 19. During grasping, finger bending, contact pressure, and contact temperature are converted by the integrated sensors into resistance changes, which regulate the firing dynamics of the first-order NbOx encoding memristors and converge at Vfusion. Dynamic responses during a robotic grasping task are illustrated in Fig. 3d. With the gripper open, no spikes are generated. As it approaches the cup rim, strain induces high-frequency spikes with stable amplitude. Upon contact, the pressure channel adds a mid-frequency component and raises spike amplitude, producing a bimodal waveform. When the liquid surface is touched, the temperature channel introduces a low-frequency component and further increases amplitude. The resulting tri-modal spike train consists of high-frequency (1.09–3.98 MHz) strain-driven, mid-frequency (16.7–83.3 kHz) pressure-driven, and low-frequency (2.9–10.0 kHz) temperature-driven components, allowing discrimination of sequential tactile phases “approach–contact–thermal sensing” (Supplementary Fig. 20). This fusion scheme increases information density and broadens tactile representation, but the fused spikes require separate frequency and amplitude readout paths, which limits low-power and real-time transmission (Supplementary Fig. 21). The addition of amplitude modulation also imposes stricter requirements on analog acquisition accuracy and energy use, highlighting the trade-off between sparse coding and representational capacity. AMIN‑based sparse burst spiking for sensory information encoding Biological interneurons utilize sparse bursts to efficiently compress and transform sensory information, transmitting it to the brain to generate real tactile sensations40. This capability arises from burst spikes conveying diverse information within a single spike train across multiple dimensions, including burst intensity (spikes per burst), intra‑burst spike density, and precise spike intervals41,42. Additionally, burst spikes can sparsely encode various stimulus intensities, efficiently transmitting sensory modality of information with low noise. Notably, the AMIN generates input-induced burst patterns driven by the fused multimodal signal Vfusion. Thus, the proposed AMIN is intended to mimic the functional role of hierarchical sensory processing, rather than the biophysical mechanism of intrinsic bursting. As shown in Fig. 4a, we designed a highly efficient NbOx memristor-based AMIN for device‑level integration and high‑order encoding of complex tactile inputs. Strain (RS), pressure (RP), and temperature (RT) channels are connected in series with first-order NbOx memristors and parallel capacitors, with the capacitors setting modality‑dependent spiking rates. The three channels are merged into a compound spike train (Vfusion) and processed by a second-order NbOx interneuron circuit, generating sparse output spikes (Vout). The DC supply polarity (VS+ and VS−) switches the circuit between excitatory and inhibitory operation, with VP and VT kept at positive bias to support compact and time‑efficient neural coding. AMINs for hierarchical burst spike encoding. a Circuit schematic of AMINs integrating strain (RS), pressure (RP), and temperature (RT) inputs via NbOx Mott memristors. Vfusion is amplified to drive a burst-spiking circuit generating output spikes (Vout), with configurable polarity for excitatory or inhibitory operation. Under a positive DC supply (VS = VP = VT = 3 V, RS = RP = RT = 5 kΩ, RL = 10 kΩ, Av = 10, R0 = 50 Ω, Cm = 0.47 nF), the neuron operates in excitatory mode and generates more spikes, whereas under a negative supply (VS− = − 3 V), it switches to inhibitory mode with fewer spikes. b Low-, middle- and high-frequency components together with amplitude variations in Vfusion modulate the MBI, IBI, and ISI, respectively, implementing a sparse burst-encoding scheme. c Experimentally measured input–output characteristics of the excitatory AMIN. The bottom, middle, and top traces correspond to the Vfusion, Vmem, and Vout, respectively. As the input amplitude increases, the membrane voltage accumulates and triggers burst firing once the threshold is exceeded. The output spikes form structured burst patterns with spike timing varying according to the input modality and amplitude. d ISI increases with both RS and RP inputs, reflecting additional spikes within each burst. e IBI increases under pressure input, indicating modulation of the burst-internal structure. f MBI is prolonged by RT, resulting in expanded spacing between bursts. As shown in Fig. 4b, the AMIN encodes signals from different modalities into temporal characteristics defined by the multi-burst interval (MBI), intra-burst interval (IBI), and inter-spike interval (ISI). As shown in Fig. 4b, the AMIN encodes signals from different modalities into temporal characteristics defined by the MBI, IBI, and ISI. To clarify the operating and encoding mechanisms of the AMIN, we performed LTspice circuit simulations based on the NbOx memristor model (Supplementary Table 3), illustrating the generation of burst-spike features from fused multimodal inputs (Supplementary Fig. 22). Activation of the low-frequency pathway alone causes the first-order NbOx memristor to switch to a low-resistance state at threshold, followed by slow discharge. This process generates Vfusion pulses featuring rapid rise and gradual decay, which, after amplification, trigger spike bursts with gradually increasing intervals over the burst duration (Tburst). The low-frequency pulse rate (Flow) thus determines the MBI. With simultaneous low- and middle-frequency inputs, middle-frequency pulses superimpose onto low-frequency signals, locally raising node voltage amplitudes. This local voltage elevation accelerates AMIN firing, producing multiple high-density spike clusters within each burst. The middle-frequency pulse rate (Fmiddle) therefore determines the IBI (Supplementary Fig. 23). Adding high-frequency input elevates the overall baseline of Vfusion, uniformly increasing spike numbers and shortening ISIs. The high-frequency pulse rate (Fhigh) therefore directly defines the ISI. These three modalities jointly regulate the AMIN burst structure, enabling hierarchical encoding of multimodal tactile inputs with clearly distinguishable temporal characteristics. The experimentally measured waveforms in Fig. 4c illustrate the relationship among the multimodal fused input Vfusion (bottom, blue), the interneuron membrane potential Vmem (middle, orange), and the resulting output spikes Vout (top, red). As multimodal inputs integrate at the node Vfusion, combining signals of varying amplitude and frequency, the membrane potential Vmem periodically accumulates until reaching the firing threshold, subsequently triggering sparse burst spike patterns. Figure 4d–f illustrate that individual input resistances (RS, RP, and RT) modulate distinct temporal features of burst spike patterns. The ISI (blue) increases markedly with rising RS and RP; the IBI (red) is governed mainly by RP and remains almost unchanged with variations in RS or RT; the MBI (yellow) is controlled solely by RT, becoming substantially longer as RT increases while remaining insensitive to RS and RP. Notably, the ISI also shows a weak dependence on RP because RP modifies the burst structure and introduces an IBI-related contribution to the extracted inter-spike timing. This coupling can be suppressed by using a smaller membrane capacitance Cm, which increases Nspikes within each burst and reduces the relative influence of IBI on ISI (Supplementary Figs. 24 and 25). Each tactile modality therefore regulates a separate temporal parameter of the burst output (Supplementary Fig. 26). The extracted ISI, IBI and MBI also remain stable over repeated burst cycles, confirming the cycle-to-cycle robustness of the AMIN temporal features against variability in Vth and Vhold (Supplementary Fig. 27). The capacitances in the AMIN circuit further provide circuit-level tunability of these temporal features, with Cm, CP and CT regulating ISI, IBI and MBI, respectively (Supplementary Figs. 28–30). LTspice simulations using the same NbOx memristor model further show that reversing the DC bias polarity VS switches the AMIN between excitatory and inhibitory modes, increasing or suppressing Nspikes, respectively, while IBI and MBI remain nearly constant (Supplementary Fig. 31). Comparing ISI values under the two polarities thus removes the contribution from IBI and isolates the RS-dependent ISI component, providing a means to decouple modality-specific temporal features (Supplementary Fig. 32). The mapping between sensory modalities and frequency bands can be tuned to match task requirements. For instance, strain and pressure signals essential for accurate grasping can be assigned to low- and mid-frequency channels, while temperature can be placed in the high-frequency band to provide supplementary information. Using this biologically inspired burst‑encoding scheme, the AMIN encodes strain, pressure, and temperature inputs into sparse spikes, supporting multimodal feature extraction and integration with neural networks. AMINs-SNN for multimodal object recognition To assess the performance of the AMIN coding strategy in multimodal tactile perception tasks, we integrated it with a SNN and built a multimodal object recognition system (Fig. 5a). It should be noted that the object recognition task was not performed as an end-to-end measured hardware demonstration. Instead, the physical sensors and NbOx circuits were experimentally characterized to obtain the sensing responses, spike-encoding behaviors, and AMIN temporal-feature relationships. These measured characteristics were then used to calibrate an equation-driven circuit model that converts tactile resistance signals into AMIN burst timing features (Supplementary Note 1). AMINs-SNN for multimodal object recognition. a Schematic of a multimodal recognition system integrating AMINs and SNN. Three input modalities- RS, RP, and RT are encoded into modality-specific burst spike features: ISI, IBI, and MBI, which are then fed into a SNN for object classification. b Heatmaps of AMINs-encoded ISI, IBI, and MBI features across time for six representative objects. MBI reflects object size (grapes compared with massage balls), IBI captures stiffness (corn and cotton pads), and ISI differentiates thermal properties (hot light and cold pepper). c Validation accuracy curves across training epochs under different modality combinations. Multimodal input (A&P&T) achieves the fastest convergence and highest validation accuracy. d Final testing accuracy for four input settings. The full AMIN input (A&P&T) achieves 90.5% accuracy, outperforming partial inputs. e Confusion matrix on the 20-class object dataset. Most objects are classified with ≥90% accuracy, confirming the discriminability of AMIN-coded burst features. We validated the system using a public dataset comprising 20 object classes, with tactile signals collected from natural interactions between a soft robotic hand and target objects43. Each sample consists of nine-dimensional tactile signals with 50 timesteps, acquired from three fingers and covering three sensory modalities: pressure, strain (finger bending angle), and temperature, with three independent channels for each modality. The temperature data were additionally collected under cold, room-temperature, and warm conditions. Based on the calibrated model, the recorded tactile signals were mapped to predefined resistance ranges for strain (4–10 kΩ), pressure (5–15 kΩ), and temperature (5 kΩ at 45 °C, 6 kΩ at 25 °C, and 8 kΩ at 10 °C), and then converted into sparse burst timing features, including ISI, IBI, and MBI (Supplementary Note 2 and Supplementary Table 4). The resulting temporal features were subsequently used as inputs to a software SNN for object classification. The SNN comprises an input spike layer, four hidden layers (384, 512, 768, and 128 neurons, respectively), and an output classification layer. Network architecture search with a design space of 4–10 layers, with each layer’s dimension varying from 64 to 1024, is used to explore as many possibilities as possible to design a high-accuracy but low-parameter architecture (Supplementary Note 3)44. After 25 iterations of mutation-based search, the final network structure provided a balance of performance and compactness. The dataset (2000 samples) comprising 20 objects ranging from corn (index 0) to cucumber (index 19), including representative items, such as plastic bottles and Rubik’s cubes, was separated into three parts, including 70% for training, 20% for validation, and the last 10% for testing. To demonstrate the discriminative capability of AMIN-encoded burst features for distinguishing multimodal physical attributes, we selected six representative objects (Fig. 5b) and analyzed their temporal spike patterns (MBI, IBI, and ISI). Results indicate that MBI captures object-size variations: small grapes induced rapid bending, rapidly reducing strain resistance and shortening the MBI; conversely, larger massage balls produced slower finger bending, leading to slower MBI changes. The IBI was sensitive primarily to object stiffness: harder objects (e.g., corn) triggered rapid pressure changes and substantial IBI reductions, while softer objects (e.g., cotton pads) showed less salient variations due to slower deformation. ISI, influenced jointly by temperature and pressure, provided further discriminability. A warm object (e.g., heated bulb) caused noticeable ISI reductions due to decreasing resistance, whereas cooler objects (e.g., cold peppers) yielded relatively stable ISI. Thus, AMIN-based burst spike encoding effectively represented multimodal tactile information, clearly separating modality-specific features. We further evaluated the AMIN coding scheme in a multi-class tactile classification task. Training involved a cosine annealing schedule over 310 epochs, with validation accuracy monitored continuously (Fig. 5c)45. Under four modality-combination settings, the combined angle (strain), pressure, and temperature modality (A&P&T) configuration provided the highest validation accuracy, achieving stable convergence within 280–310 epochs. Final test accuracy reached 90.5% under the A&P&T condition, substantially outperforming other modality combinations and confirming the advantage of integrated multimodal AMIN encoding (Fig. 5d). The confusion matrix (Fig. 5e) revealed that most object classes were identified with accuracies exceeding 90%, with nearly half reaching 100%. Collectively, these results illustrate the capability of AMIN burst spike encoding to efficiently and distinctly represent multimodal tactile information, enhancing classification accuracy in neuromorphic perception tasks (Supplementary Note 4 and Supplementary Table 5). Discussion In this study, we present AMINs inspired by the integration and higher-order coding schemes of biological spinal interneurons. AMINs extract multimodal tactile temporal encoding within a single spike train. The system integrates distributed tactile sensing and hierarchical encoding circuits on a hybrid integrated platform, combining v-AuNW strain sensors, PPy-based pressure sensors, temperature sensors, and an NbOx memristor array. AMINs achieve a broadband spiking range of 2.9 kHz to 3.98 MHz, with the three modalities spectrally separated to enable independent modulation of output features. The fused spikes, enriched in information density, drive polarity-tunable AMIN neurons to produce excitatory or inhibitory burst firing for sparse burst coding. The resulting temporal structures of the burst spikes enable discrimination of object size, stiffness, and temperature. Additionally, by combining the experimentally characterized AMIN encoding behavior with a software SNN, the proposed framework evaluates multimodal tactile classification using sparse burst timing features and achieves 90.5% accuracy in a 20-class object recognition task. This hardware-level approach improves data efficiency, reduces readout power, and lowers system complexity, making AMINs suitable for low-power edge AI applications and next-generation human–machine interfaces. Methods Preparation of NbOx memristors The NbOx memristors were fabricated as follows: first, the bottom electrodes, consisting of Ti (5 nm)/Pt (35 nm), were patterned on a Si/SiO2 substrate using photolithography, deposited via e-beam evaporation, and followed by a lift-off process. Next, 35 nm thick NbOx active layers were deposited using magnetron sputtering at a power of 50 W in an argon atmosphere at 0.5 Pa. These active layers were then patterned at room temperature using photolithography and lift-off techniques. The top electrodes, made of Ti (5 nm)/Pt (35 nm), were similarly deposited by e-beam evaporation and patterned using photolithography and lift-off methods. The resulting two-terminal crossbar structure of the Mott memristors features an active area of 1 × 1 μm2. Device characterization The cross-sectional TEM images were acquired and analyzed using a transmission electron microscope (FEI Tecnai TF-20, UK). Room-temperature electrical measurements were performed with a Summit 1100B-M Probe Station. The DC mode was measured using an Agilent B1500 semiconductor parameter analyzer. For the transient switching response of NbOx, a B1530A fast measurement unit module was utilized. For testing artificial neurons, the input signals were generated by a series of arbitrary function generator (Tektronix AFG31000), and the output spike signals were captured and recorded with an oscilloscope (Tektronix MSO46 4-BW-500), ensuring accurate and precise measurements. ANN-to-SNN conversion To address the difficulty of training SNNs due to their discrete, binary spike-based communication, which is incompatible with conventional backpropagation, we adapt a quantized ANN for conversion into an SNN. During ANN training, we implement a specialized activation function based on clamp and quantization (CQ+) operations. This approach enables effective network training using a modified backpropagation algorithm. The forward pass is defined as: Here, xl represents the output of layer l, Wl is the weight matrix, and C and Q are the clamp and quantization functions, respectively. The gradients for backpropagation are then calculated as: The derivatives of the clamp (C’) and quantization (Q’) functions are either 0 or 1. Consequently, the backpropagation process in our training simplifies to the conventional method used for standard ANNs. This allows for direct and efficient training of the network parameters before converting the trained weights to the SNN architecture for deployment. The spike generation model of the SNN used in this paper is the average IF model, which is a more robust and higher-accuracy spiking neuron compared to the simple IF model45. Our model first computes a constant, time-averaged input, Ali, for each neuron i in layer l. This input is calculated by integrating weighted presynaptic spikes over a time window of duration T: where wi,jl are the synaptic weights, sil−1(t) are presynaptic spikes, and bil is a bias term. This static input Ail then drives a standard integrate-and-fire neuron. At each time step t, the membrane potential Uil(t) accumulates the averaged input: When the potential Uil(t) reaches a threshold θ the neuron fires an output spike, sil(t)=1. The potential is then reset by subtraction (Uil(t)←Uil(t)−θ).The firing condition is: Supplementary information Supplementary Information Transparent Peer Review File Source data Source Data