NeuroTac: Neuromorphic Tactile Sensor

• NeuroTac is a neuromorphic optical tactile sensor that mimics human fingertip morphology and transduces contact via asynchronous spike-based outputs.
• It encodes dynamic tactile information for applications like texture recognition, incipient slip detection, and contact-angle estimation with high accuracy.
• The system leverages low-power, low-latency neuromorphic hardware and spiking neural networks for efficient on-board processing in robotics.

The NeuroTac sensor is a neuromorphic optical tactile sensor engineered to provide biomimetic, event-driven tactile perception. It fuses a soft, human-fingertip-inspired morphology with an embedded event-based vision system, producing asynchronous spatiotemporal tactile signals analogous to those found in neural afferent pathways. This architecture supports efficient, low-latency tactile sensing, robust dynamic contact encoding, and low-power processing, with demonstrated applications in robotic texture recognition, incipient slip detection, and high-frequency contact-angle estimation. NeuroTac’s output is inherently spike-like, enabling direct interfacing with spiking neural networks (SNNs) and neuromorphic hardware for energy-efficient on-board perception (Ward-Cherrier et al., 2020, Xu et al., 18 Sep 2025, Lu et al., 11 Sep 2025, Lu et al., 2 Jun 2026).

1. Sensor Architecture and Neuromorphic Transduction

The NeuroTac consists of a compliant hemispherical tactile tip—modeled after the papillae-membrane structure of human glabrous skin—and a sealed internal optical chamber. The tip is structured with either internal white-tipped pins (standard TacTip-heritage implementations) or externally protruding papillae (for advanced slip detection tasks). Deformation of the outer membranous skin under contact conditions leads to displacements of these markers or papillae, which are optically tracked by a centrally mounted event-based vision sensor such as the iniVation DAVIS240 or DVXplorer (Ward-Cherrier et al., 2020, Lu et al., 11 Sep 2025, Xu et al., 18 Sep 2025).

Event cameras within NeuroTac output Address-Event Representation (AER) streams: asynchronous pixel-level events triggered by brightness changes caused by marker or papilla motion. These events are spatially and temporally filtered and binned to form high-dimensional spiking representations. For example, the original design outputs 49 taxel spike trains corresponding to 49 pins, while slip-detection and high-density event-matrix approaches pool events over a $20 \times 20$ spatial grid (Ward-Cherrier et al., 2020, Lu et al., 11 Sep 2025, Xu et al., 18 Sep 2025).

The neuromorphic paradigm is central: Instead of frame-based tactile images or analog signals, information is encoded as spike times and spike rates, naturally supporting SNNs and hardware with event-driven computation (Ward-Cherrier et al., 2020, Lu et al., 2 Jun 2026).

2. Event-Based Sensing Principle and Information Encoding

NeuroTac's signal transduction leverages the sensitivity of event cameras to changes in visual appearance generated by mechanical deformation of the soft tip. During contact, either the internal markers displace or the external papillae are deformed and sheared, resulting in localized optical changes internal to the sensor, which in turn trigger asynchronous events in the embedded neuromorphic sensor.

Processing pipelines typically perform the following steps:

• Spatial filtering: pixel events are assigned to receptive fields corresponding to taxels (marker-centered or papilla-centered regions).
• Noise filtering: isolated events lacking spatiotemporal neighbors are discarded (e.g., require another event within 5ms in a neighboring pixel).
• Temporal pooling: events are pooled within short intervals (e.g., 20 ms) or binned at high frequencies (e.g., 1 ms bins for high-rate processing).
• Spike encoding: events are grouped per taxel per interval, constructing multi-channel spike trains.
• Dynamic/static representation: in contact-angle tasks, decaying accumulators or sliding windows reconstruct either persistent contact contours or short-term activity (Lu et al., 11 Sep 2025, Lu et al., 2 Jun 2026).

The output is thus a structured, sparse, and temporally-resolved spike train array or spatiotemporal event tensor (e.g., dimensions $(1000, 20, 20)$ for 1000 time bins and $20 \times 20$ spatial pooling (Xu et al., 18 Sep 2025)).

3. Sensor Variants and Hardware Implementations

Two main morphological variants have emerged:

Design	Tactile Encoding	Contact Structure	Typical Application
Internal pin array	Marker-displacement	Densely connected dome	Texture perception, general touch
Externally protruding papillae	Papillae deformation	Segmented, radially layered tip	Slip detection, spatiotemporal slip onset

• Pin-based configuration: 49 (or 91) internal 3D-printed white markers in a soft membrane; marker motion optically detected and encoded as spike trains. Suitable for fine spatial coding and biomimetic temporal patterns (Ward-Cherrier et al., 2020, Xu et al., 18 Sep 2025).
• Papillae-based configuration: Extruding silicone papillae arranged in three concentric layers with graded height. Supports spatially resolved, temporally progressive incipient slip encoding, crucial for safety-critical manipulation (Lu et al., 11 Sep 2025).

All implementations use custom-molded silicone membranes, internal or external marker mechanics, event-camera mounting, and sealing for stability. The typical sensor diameter is 20 mm or less, enabling integration on robotic fingertips (Ward-Cherrier et al., 2020, Xu et al., 18 Sep 2025, Lu et al., 11 Sep 2025).

4. Signal Processing, Learning Algorithms, and Event Representations

NeuroTac’s event-driven outputs require specialized spike-based processing. Research has focused on encoding strategies and SNN architectures:

• Coding strategies:
  • Intensive (mean spike rate per sample)
  • Spatial (per-taxel spike count vector)
  • Temporal (rolling-window mean rates)
  • Spatiotemporal (multi-neuron Van Rossum distance, kernelized spike trains) (Ward-Cherrier et al., 2020)
  • Decaying-accumulator static vs. dynamic contour (for contact-angle estimation) (Lu et al., 2 Jun 2026)
• Learning and inference:
  • K-nearest neighbor (KNN) classification with specialized spike-train distances (e.g., Van Rossum metric) (Ward-Cherrier et al., 2020)
  • Spiking convolutional neural networks (SCNNs) with integrate-and-fire neurons, temporal binning, and spiking activation (e.g., Sinabs/Speck2f deployment) (Xu et al., 18 Sep 2025, Lu et al., 11 Sep 2025)
  • Multi-layer perceptrons (MLPs) for regression on event-derived binary contours (Lu et al., 2 Jun 2026)

Temporal coding is consistently the most effective, especially for dynamic tasks such as texture classification (up to 98.3% LOOCV on artificial textures (Ward-Cherrier et al., 2020)), with spatiotemporal metrics providing fine discrimination between complex patterns.

Static decaying-accumulator representations outperform pure dynamic representations for regressing persistent contact quantities, achieving MAE of 0.160° in real-time tactile angle estimation with sub-10 ms P99 processing latency (Lu et al., 2 Jun 2026).

5. Application Domains: Texture Recognition, Incipient Slip, and Angle Estimation

Texture Discrimination

• Under robot-controlled sliding, NeuroTac achieves 98.3% accuracy on artificial textures and 93.0% on natural textures, with timing-based (temporal) and spatiotemporal coding strategies performing best (Ward-Cherrier et al., 2020).
• With human-inspired exploratory movements, sliding and sliding+rotating are optimal. Under variable conditions, sliding+rotating achieves the highest generalized accuracy of 87.33% on 10 natural textures at 8.04 mW board power (Xu et al., 18 Sep 2025).

Incipient Slip Detection

• The papillae-based NeuroTac variant enables spatial segregation of stick-slip transitions; event spike rate surges as outer papillae enter incipient slip, followed by central papillae.
• A SCNN, operating on event tensors ($30 \times 1 \times 20 \times 20$), attains 94.33% accuracy across three slip states and detects incipient slip at least 360 ms before gross slip (across all tested gravity-induced slip trials) (Lu et al., 11 Sep 2025).

High-Frequency Contact-Angle Estimation

• Event-derived static (state-accumulator) representation combined with an MLP yields mean absolute error of 0.160° in rolling contact for contact-angle regression, robust across speed and indentation depth, with consistent sub-10 ms P99 end-to-end latency (Lu et al., 2 Jun 2026).
• The system rapidly infers angle even with motion interruptions and varying deformation rates, demonstrating suitability for closed-loop robotic control.

6. Energy Efficiency, Latency, Robustness, and Practical Deployment

NeuroTac’s event-driven encoding allows neuromorphic architectures to operate with orders of magnitude lower power than frame-based analogs:

• Demonstrated Speck2f board operation at 7.01–8.04 mW for inference, over 10,000× and 2,500× lower power than conventional CPU and GPU baselines, respectively (Xu et al., 18 Sep 2025).
• Sub-10 ms P99 processing latency for all tested contact state estimation pipelines, compatible with high-frequency tactile feedback requirements (Lu et al., 2 Jun 2026).
• Robustness to contact speed, indentation, and exploratory movement variability, although system performance is strongly dependent on maintaining sufficient dynamic deformation for event generation, with shallow indentations leading to perceptual degradation (Xu et al., 18 Sep 2025, Lu et al., 11 Sep 2025, Lu et al., 2 Jun 2026).
• Complications may arise from silicone tip deformation nonlinearity, marker/papilla variability due to fabrication tolerances, and the need for careful contact depth and motion control to prevent hardware damage (notably above 3 mm indentation in some configurations) (Xu et al., 18 Sep 2025).

7. Technological Positioning, Comparative Context, and Limitations

NeuroTac defines a distinctive class within tactile sensing:

Sensor Type	Modality	Output Representation	Biologically Inspired	Dynamic Suitability
NeuroTac	Optical/event	Spiking event streams	Skin and neural code	Dynamic texture, slip, angle
Frame-based optical (GelSight, MiniTac)	Optical/frame	Dense images (RGB or HSV/deformation)	Skin/topography only	Suited for static and slow changes
Capacitive/Resistive arrays	Electrical	Analog taxel array	Taxel topology	Variable

Relative to sensors such as OmniTact, NeuroTac is unique in providing asynchronous spatiotemporal outputs and native compatibility with neuromorphic computation (Padmanabha et al., 2020). In contrast to vision-based tactile imagers like MiniTac, NeuroTac does not reconstruct dense deformation fields but instead encodes dynamic contact structure for tasks where compression and event sparsity are advantageous (Li et al., 2024). NeuroTac is less suited to tasks demanding absolute static force mapping or high-precision 3D pressure fields, and performance declines if contact is purely static.

NeuroTac’s strengths are event-driven tactile encoding, low power, low latency, robust slippage/warning (with co-designed morphology), and compatibility with edge-computation and adaptive active exploration. Limitations include performance dependence on motion/dynamics, static force insensitivity, practical demands on contact management, and variation due to fabrication tolerances.

• NeuroTac: A Neuromorphic Optical Tactile Sensor applied to Texture Recognition (Ward-Cherrier et al., 2020).
• Exploratory Movement Strategies for Texture Discrimination with a Neuromorphic Tactile Sensor (Xu et al., 18 Sep 2025).
• A Neuromorphic Incipient Slip Detection System using Papillae Morphology (Lu et al., 11 Sep 2025).
• Static and Dynamic Representations for Tactile Contact-Angle Estimation with Event-Based Sensors (Lu et al., 2 Jun 2026).
• MiniTac: An Ultra-Compact 8 mm Vision-Based Tactile Sensor (Li et al., 2024).
• OmniTact: A Multi-Directional High Resolution Touch Sensor (Padmanabha et al., 2020).