When a Shoe Knows How to Think: How Wearable Pressure Sensors Are Teaching Neural Networks to Walk

A physical therapist watches a patient cross the room. Not counting steps, she reads the subtle weight shifts, the hesitation before the heel strike, the slight drag of one foot against the floor. Rhythm is what she is reading. Now consider that a network of artificial neurons embedded in a simulation running on a GPU attempts the same kind of reading. The neurons do not see the patient. They feel the floor through sensors tucked inside a pair of insoles, and they learn.
That is the core idea behind a study published at the 2026 International Russian Smart Industry Conference, where researchers at Kazan Federal University’s Institute of Information Technologies and Intelligent Systems presented a framework for training a spiking neural network directly from wearable insole data. The work sits at the intersection of neuromorphic computing, biomechanics, and embedded sensing, and it raises a question worth asking carefully. What does it mean for a machine to learn how to walk by listening to pressure?
The Power Budget That Changes Everything
Before getting into the mechanics of how this system works, it is worth understanding why spiking neural networks exist in the first place, and why the answer is almost always about energy.
Always-On Versus Event-Driven Compute
A conventional deep learning model runs continuously. Whether the sensor detects anything interesting or not, the network processes, multiplies matrices, updates activations, and consumes power at a steady rate. It is like leaving every light in a building switched on regardless of whether anyone is in the room. For a device that lives inside a shoe and must run on a coin cell or a small rechargeable battery, that model fails on first principles.
A spiking neural network operates on a fundamentally different budget. Meaningful compute only happens when an event occurs, when a sensor crosses a threshold, when a neuron fires. During the swing phase of a gait cycle, when the foot is airborne and pressure readings are flat, the spiking network is essentially silent. No spikes, no synaptic updates, no wasted current. The shoe is listening, but not burning. That distinction separates a wearable sensor that needs nightly charging from one that runs for weeks or months between charges, which is the only version of the product that works in the real world of patients, athletes, and clinical deployment.
This engineering case shapes every design decision that follows.
The Problem With Continuous Signals in a Spiking World
Sensors are honest but noisy. The insoles used in this study contained arrays of pressure elements that recorded changes in ground contact force throughout walking, standing, and turning. Data streamed continuously to a host computer through a wired USB interface, producing multivariate time series for each foot. On the surface, this sounds ideal. In practice, it creates a substantial engineering problem.
Spiking neural networks do not operate on continuous values. They operate on events, discrete moments in time when a neuron fires. The biological brain does not process a smooth voltage curve representing foot pressure. Rather, it processes a sequence of spikes, sharp electrical events that carry information through their timing rather than their amplitude. To bridge the gap between the sensor world and the spiking world, researchers must convert analog signals into spike trains, and that conversion is where things can go wrong.
Transient jumps in the raw pressure recordings did not reflect genuine gait events. These artifacts arose from sensor noise, cable interference, and sudden uncharacteristic movements. Under frequency-based spike encoding, a spike jump in the signal becomes a burst of spikes in the network. That burst, if left uncorrected, dominates the learning process and distorts the synaptic weights the network works to build. The researchers therefore treated preprocessing not as a formality but as a foundational design decision.
Smoothing the Signal Before It Becomes a Spike
The engineers evaluated three smoothing methods on representative sensor streams. They include moving average, exponential smoothing, and median filtering. The comparison was not merely cosmetic. Each method behaves differently when it encounters an abrupt transient.
Moving average and exponential smoothing both introduce temporal distortions during rapid pressure changes. Because these methods weight recent values continuously, they either lag behind a sudden rise or overshoot after a sudden fall. Either outcome corrupts the timing of the resulting spikes, and timing is precisely what a spiking network encodes.
Why the Median Filter Wins
Median filtering behaves differently. It replaces each point in the signal with the median of a surrounding window, which means a single outlier cannot pull the output toward an incorrect value. The median is inherently resistant to impulsive artifacts. For gait data, where the underlying pressure waveform has a consistent shape but is occasionally interrupted by sensor noise, median filtering preserves the overall phase and amplitude of the signal far more faithfully than the alternatives. The researchers selected it as the default preprocessing method for all subsequent encoding.
After filtering, per-sensor normalization rescaled the signals to a common range. This step was necessary because sensors at different locations on the insole experience different pressure magnitudes during the same gait cycle. Without normalization, sensors with higher peak values would generate higher spike frequencies simply because of where they happened to sit on the insole, not because they carried more meaningful information. Normalization decoupled signal content from sensor placement.
From Pressure to Spikes: The Frequency Encoding Architecture
With the signals cleaned and normalized, the researchers built an encoding algorithm that converts pressure values into spike generator activations. The design uses two thresholds. Crossing 0.2 activates the first spike generator. Crossing 0.4 brings additional generators online. Each generator tracks which sensors have already fired during a given step and avoids re-triggering them, which keeps the spike-frequency profile stable and prevents runaway activation.
Each time window produces an instantaneous firing rate mapped to a bounded interval. Spikes are then emitted either periodically, with an interspike interval equal to the inverse of that rate, or stochastically through a Poisson process at the same rate. The choice between these modes reflects a classic tradeoff in neural coding. Periodic emission is more deterministic and easier to analyze. Poisson emission introduces biological variability and can improve generalization by preventing the network from over-fitting to a fixed temporal pattern.
The result of the encoding is a spike-frequency histogram that rises and falls in correspondence with the actual gait phases recorded by the insoles. When the foot is flat on the ground, pressure is high, generators are active, and spikes are dense. When the foot is in the air, pressure drops, generators reset, and spikes become sparse. The network’s input layer receives not just the fact that someone is walking, but the temporal texture of how they are walking, encoded in the relative timing and density of spikes across channels.
STDP: The Learning Rule That Wires Itself
The learning algorithm at the center of this work is spike-timing-dependent plasticity, or STDP. Among the most studied synaptic learning rules in computational neuroscience, its basic principle is straightforward. When a presynaptic neuron fires before a postsynaptic neuron within a short time window, the synapse between them strengthens. Reversed timing weakens it. The rule is local, meaning each synapse updates based only on the activity of its own pre- and postsynaptic partners, without any global error signal propagating backward through the network.
The Intelligence Stays at the Edge
Because STDP is a local rule, the intelligence stays at the edge. No server farm needs to backpropagate errors across a cloud connection. Learning happens in the shoe, at the synapse, in real time. For engineers thinking about the Internet of Bodies, that architectural property is not a minor implementation detail. It is the technical moat. A device that learns on-chip, without phoning home, can operate in a hospital ward, on a hiking trail, or in a developing market where reliable connectivity cannot be assumed. Gradient-based backpropagation requires the full network to reside in memory simultaneously. STDP does not, and that asymmetry scales in ways that centralized training never will.
How the GRAS Simulator Implements STDP
In the GRAS simulator, STDP operates at the level of synaptic connections. Update magnitudes decay with the absolute time difference between pre- and postsynaptic spikes, and weight bounds keep values within a biologically plausible range. The researchers initialized all synaptic weights to 1500, the minimum threshold required to generate a spike in the ESRN neuron model. That value may sound surprisingly large compared to the 0.0 to 1.0 floating-point weights familiar from standard deep learning, but the difference is a matter of representation rather than magnitude. Many neuromorphic simulators, including GRAS, encode synaptic strengths as high-precision integers in a fixed-point arithmetic scheme, where 1500 corresponds to the lowest value capable of crossing the neuron’s firing threshold, roughly equivalent to a weight just above zero in a conventional network.
Why Early Activity Matters
That initialization guaranteed early network activity even before STDP had time to sculpt the weights, giving the learning rule enough signal to observe from the first training steps.
Closing the Loop: Biological Feedback as a Teaching Signal
Here is where the framework departs from standard unsupervised STDP and introduces something more structured. Pure STDP with no external guidance will learn whatever correlations exist in its input, but nothing in the rule itself evaluates whether what it has learned serves the intended function. The researchers incorporated an external feedback signal that compared the evolving activity of a group of extensor motor neurons against a biologically realistic reference configuration developed in the Laboratory of Neuromorphic Computing and Neurosimulations.
The comparison worked as follows. At each learning snapshot, the research engineers divided the normalized voltage trace across the extensor motor neuron group into 50 millisecond slices. Within each slice, they identified the first minimum following the monosynaptic response and fit a linear regression line through these minima to capture the overall trend of post-response activity. Cosine similarity then compared the slope of this line to the target slope from the reference model.
Cosine similarity measures the angle between two vectors. A cosine of 1.0 means the vectors point in exactly the same direction. In the reported experiment, cosine similarity between the learned slope vector and the target slope vector reached 0.99971, corresponding to an angular separation of approximately 1.36 degrees. Through STDP and feedback, the network had learned to reproduce the expected trajectory of extensor motor neuron activity almost exactly.
What the Network Actually Learned
Precision matters here. Nobody trained the network to classify gait events into categories. No labels entered the process. No one showed the network examples of correct and incorrect outputs in the way a supervised classifier requires. Instead, the researchers guided it to reproduce a pattern of neural activity meaningful from a motor control perspective, the kind of activity profile that a healthy nervous system produces in response to the same pressure inputs.
This framing carries weight for several reasons. First, neurophysiological constraints shape the learned representation rather than arbitrary classification boundaries. The network learns to resemble a motor circuit, not just to predict a label. Second, dependence on annotated data drops considerably. In clinical and rehabilitation settings, labeled data is often scarce, expensive to collect, or ethically complex to obtain. A reference pattern from physiology recordings or clinical protocols can substitute for labeled training sets in many motor control scenarios. Third, because the learned dynamics must match a known biological trajectory, the internal representations may prove more interpretable than those produced by gradient-based training on abstract objectives.
From the Lab to the Foot: Practical Implications
The researchers grounded this framework in data from actual wearable insoles worn by real subjects during walking, standing, and turning. That grounding matters for anyone thinking about where this technology eventually lands.
Neuromorphic computing has long carried the reputation of being a research curiosity, compelling in theory but distant from practical deployment. Studies like this one close that distance incrementally. The encoding scheme here, frequency coding with threshold-based generator activation, fits directly into the event-driven input formats of neuromorphic chips such as Intel’s Loihi 2 and BrainChip’s Akida, both of which are specifically engineered for asynchronous, spike-based input and on-chip local learning. Because STDP requires no global gradient computation, the rule can run on-chip without offloading to a host processor. An insole that encodes its own pressure data into spikes and updates its own synaptic weights through local plasticity could, in principle, operate for extended periods on a small battery.
Signal to Spike: A Gait Cycle Decoded
To understand concretely how the spiking representation maps onto a real gait cycle, the table below traces what happens to the signal and the network response at each phase of a single step.
| Phase of Gait | Analog Pressure Signal | Spiking Network Response |
|---|---|---|
| Heel Strike | Sharp voltage spike | High-frequency burst of spikes |
| Mid-Stance | Sustained high pressure | Steady, rhythmic spike train |
| Toe-Off | Rapidly dropping pressure | Thinning spike density |
| Swing Phase | Zero or baseline pressure | Silence, near-zero power draw |
Each row in that table is also a row in the power ledger. The network earns its silence during the swing phase and spends its compute precisely where the biomechanical information is richest.
Where the Applications Land
The potential applications span a range of clinical and consumer contexts. Neurorehabilitation stands out. A networked insole could learn an individual patient’s gait pattern and flag deviations that might indicate fatigue, injury, or neurological changes. Prosthetics research offers another venue, where the same framework could feed into control systems that adapt to a user’s walking style over time. Sports performance monitoring benefits as well, because the event-driven encoding preserves temporal detail that aggregate metrics discard.
Where the Researchers Go From Here
The authors of this study are candid about what remains to be done. Learning quality is sensitive to the choice of STDP parameters, including learning rates, time constants, weight bounds, and homeostatic terms. Different sensor layouts, different subjects, and different movement types may all require recalibration of these settings. The reference activity profile used for feedback is also not unique. Multiple patterns of motor neuron activity can produce the same behavioral outcome, and a system that matches one reference too rigidly might penalize valid compensatory strategies in rehabilitation contexts where patients have learned to move differently because of injury or neurological condition.
Future work from the group will pursue automated parameter selection, including Bayesian search and stability-driven tuning rules that adjust STDP settings based on measured firing-rate statistics during training. Validation on larger and more diverse datasets, covering different sensor configurations, movement classes, and inter-subject variability, is also planned. Most consequentially, the researchers intend to move toward closed-loop real-time experiments where the network’s output can influence stimulation patterns or assistive control policies directly. In that setting, the feedback signal is no longer just a teaching mechanism. It becomes part of an adaptive controller that personalizes its behavior over time while maintaining physiological plausibility and safety.
The Larger Significance
A spiking neural network that learns from pressure underfoot and calibrates its behavior against a biological reference is not a standalone curiosity. It represents a coherent approach to a genuine problem in wearable computing. How do you build a device that learns from the body without requiring the computational infrastructure of a data center?
The answer, as this research suggests, is to let the learning rule do the work locally, to encode information in the timing of events rather than in continuous values, and to evaluate learning not against an arbitrary loss function but against a reference grounded in biology. None of these ideas are new in isolation. What the researchers at Kazan Federal University have done is assemble them into a working pipeline, run it on real sensor data, and show that the resulting network converges to a biologically meaningful solution.
The patient crossing the room may not yet feel anything different underfoot. But the insoles are paying attention, and they are learning.
Analytic Note: The Market Shift to Vertical Intelligence
The following section is written for readers tracking the neuromorphic computing sector from a capital markets or technology strategy perspective.
The current AI investment cycle has concentrated almost entirely on centralized, large-scale compute. data centers, GPU clusters, and the cloud infrastructure that connects them to end users. That architecture makes sense for general-purpose language and vision models, where training requires exposure to enormous datasets and inference demands are bursty and unpredictable. For the Internet of Bodies, however, centralized compute is the wrong unit of analysis entirely.
Wearable medical devices, prosthetics, and continuous health monitors operate under constraints that no cloud subscription can solve. Battery life is finite. Connectivity is intermittent. Latency in a motor-control loop is not an inconvenience but a safety variable. The relevant question for investors and product strategists is not which cloud provider will host the model but which chip and learning rule will run inside the device itself.
This is the frontier the Kazan Federal University researchers are working on, and it maps directly onto what equity analysts sometimes call the “vertical intelligence” thesis which is the idea that the next durable moats in AI will not belong to the companies with the largest models but to those with the most defensible on-device learning stacks. On-chip learning, or OCL, refers specifically to the ability of a device to update its own parameters from local sensor data without phoning home to a training cluster. STDP is one of the most hardware-compatible OCL rules known, and the intellectual property surrounding its implementation in fixed-point, event-driven silicon is still relatively open terrain.
What to Watch
The hardware layer is beginning to consolidate. Intel’s Loihi 2 and BrainChip’s Akida both support asynchronous spike-based input and local synaptic updates, and both companies have positioned themselves explicitly for edge deployment in medical and industrial sensing. What the research community is now producing, including work like this insole study, are the software and algorithmic foundations that determine which learning rules, encodings, and feedback architectures actually work on real sensor data. Those foundations, once validated and published, tend to inform the next generation of chip-level design decisions. Early readers of that literature tend to arrive at informed positions well before they appear in product announcements.
For anyone building a position in the neuromorphic space or evaluating wearable health companies, the strategic question worth tracking is not whether spiking neural networks outperform transformers on a benchmark. It is whether a given company’s device can learn from the body, adapt over time, and do both without a cellular connection and without a nightly charge. The research reviewed in this article suggests that the algorithmic pieces for that product are closer to ready than the market currently credits.
Citation
Toschev, A., & Mikhailova, Y. (2026). Learning of a Spiking Neural Network with Feedback from the External Environment on Bio Realistic Patterns. In 2026 International Russian Smart Industry Conference (SmartIndustryCon) (pp. 86–90). IEEE. https://doi.org/10.1109/SmartIndustryCon68821.2026.11493007

Responses