The core insight of the architecture is that streetlight clusters behave like a collective electrical fingerprint. Instead of attaching dedicated sensing and communication hardware to every lamp post, the system groups ten streetlights under the supervision of one ESP32 microcontroller node. That node continuously samples voltage, incoming and outgoing current, and other electrical parameters, then transmits aggregated diagnostic data over a multi-hop wireless mesh to a centralized monitoring station. Because the sensing burden is shared across a cluster, the number of deployed devices—and the cost, installation labor, and embodied carbon that come with them—drops dramatically.

What distinguishes the system from earlier smart streetlight schemes is the breadth of faults it can catch. Most existing monitoring platforms focus narrowly on whether a lamp is on or off. The proposed architecture goes further, identifying faulty bulbs through near-zero current draw at nominal supply voltage, detecting leakage currents as a steady differential current exceeding safe limits, and flagging ground faults through a severe transient current spike paired with an instantaneous voltage drop. These last two categories matter for public safety: leakage currents and electrocution hazards at lamp posts are rarely addressed by commercial systems, yet they represent a genuine risk to pedestrians and maintenance crews.

The detection engine is built on statistics rather than fixed thresholds. During commissioning, the network establishes a baseline from ten fully operational streetlights, recording seven-dimensional feature vectors that include RMS voltage and current, total harmonic distortion up to the tenth harmonic, power factor, and RMS residual current. Each feature is computed over observation windows of 500 samples taken at 5 kilohertz, spanning five mains cycles, which suppresses single-sample analog-to-digital converter noise and transient switching spikes. From this baseline, the central station computes a mean vector and a covariance matrix, and each node then evaluates the Mahalanobis distance—a statistical measure that accounts for how the electrical parameters vary together—between live readings and the healthy reference.

When that distance exceeds a calibrated threshold, a fault flag is raised. The threshold is not chosen by trial and error: under a multivariate Gaussian assumption, the squared Mahalanobis distance follows a chi-squared distribution, so the team derives the cutoff analytically to hit a target false-alarm rate. For seven features and a 1 percent false-alarm target, the threshold works out to 4.93. The covariance matrix is regularized using the Ledoit–Wolf shrinkage method to guarantee it remains invertible even when features such as voltage and current become highly correlated under stable grid conditions. Crucially, the heavy matrix inversion is done once, offline, at the central station; each ESP32 receives a precomputed 196-byte inverse matrix, so the online computation reduces to a single matrix-vector product that executes in well under a millisecond on the microcontroller’s 240 MHz dual-core processor.

Communication reliability is handled by a self-healing mesh protocol. Every node acts as both a sensor and a relay, aggregating its own diagnostic payloads with data from downstream neighbors before forwarding them toward the sink. Routing paths are chosen by a dynamic cost function that weighs link quality and each neighbor’s residual energy, so traffic flows over the strongest, most sustainable routes. If an intermediate node fails, a beaconing mechanism detects the loss within 15 seconds—three missed hello beacons at 5-second intervals—and the affected node prunes the dead relay from its routing table and recalculates an alternate path. The authors’ probabilistic analysis shows that as node density increases, the probability of successful data delivery approaches one, meaning the network becomes more resilient, not less, as it grows.

Energy efficiency is engineered into every layer. Nodes spend most of their lives in deep-sleep mode, waking to sample electrical states on a low-duty-cycle TDMA schedule and transmitting only on fault detection or scheduled heartbeats. The average node power, including the radio, comes to 184.6 milliwatts, and a backup battery sustains operation for 7.5 days under typical urban channel conditions. Solar panels recharge the nodes, eliminating the need for additional grid power infrastructure and shrinking the carbon footprint of the monitoring network itself. The mesh topology also extends coverage across large urban areas, since nodes relay information through neighbors rather than requiring direct lines to a central station.

In simulation studies emulating a real-world urban deployment, the architecture delivered fault detection accuracy exceeding 97 percent with a low false-alarm rate, and the mesh network achieved packet delivery ratios close to 99 percent with low latency. The self-healing routing mechanism rerouted diagnostic data around failed intermediate nodes without interrupting transmission, and reliability analysis confirmed that overall network availability climbs with node density. A correlation analysis further revealed that voltage and current stability strongly predict fault detection accuracy, while mesh connectivity significantly influences packet delivery and fault localization performance—evidence that the electrical and communication layers reinforce one another.

The system’s most visible feature for city operators is its geographic intelligence. A centralized GIS-based mapping interface visualizes faults in real time, pinpointing the location of a defective streetlight cluster so maintenance crews can be dispatched directly to the problem. The authors argue this capability should shorten maintenance response times and improve operational efficiency compared with the prevailing model of manual inspection rounds and reactive public reporting, which delays fault detection and inflates operational costs. A comparison table in the paper shows that prior systems—from early 2007 wireless sensor network prototypes through ZigBee-based controllers and IoT platforms—generally lack leakage current detection, GIS mapping, mesh networking, solar power, and multi-light monitoring in combination, whereas the proposed architecture integrates all of them.

The work remains at the simulation and prototype stage; the hardware validation described includes an ESP32-based sensor node with a switching relay module and LED array simulating streetlight loads, alongside an Arduino gateway with a Bolt IoT wireless module for data logging. The authors outline future directions including full-scale field deployment, machine learning for predictive maintenance and fault forecasting, and integration with digital twin platforms and edge computing. If those steps succeed, the cluster-based monitoring strategy could offer municipalities a practical template for scaling intelligent infrastructure: fewer sensors, longer battery life, self-healing communications, and a map that shows exactly where the next dark street—or hidden electrical hazard—awaits repair.

Subject of Research: A centralized wireless sensor mesh architecture for automated streetlight fault detection and GIS-based fault localization

Article Title: A centralized mapping architecture for Automated Streetlight Fault Detection via Wireless Sensor Mesh Networks

Article References: Bhattacharya, S., Mandal, B., & Samanta, B. (2026). A centralized mapping architecture for Automated Streetlight Fault Detection via Wireless Sensor Mesh Networks. Frontiers in Sensors, 7, Article 1905827. https://doi.org/10.3389/fsens.2026.1905827

Image Credits: AI Generated

DOI: 10.3389/fsens.2026.1905827

Keywords: streetlight fault detection, wireless sensor networks, mesh networking, smart cities, ESP32, Mahalanobis distance, leakage current, GIS mapping, solar power, Internet of Things, fault localization, energy efficiency

The answer, according to researchers Rahul Nayak of VIT-AP University and Gondhi Navabharat Reddy of Symbiosis Institute of Technology, is a qualified yes, delivered with unusual statistical honesty. Their approach builds on federated learning, a training scheme in which each device trains a shared model on its own local data and transmits only parameter updates, never packets. That solves the bandwidth and privacy problems of centralised training, and it removes the single point of failure that a central model represents. But federated learning has two weaknesses that line up badly with the IoT setting. First, the canonical aggregation rule, FedAvg, assumes all clients draw data from a common distribution, while a camera, a thermostat and a gateway see utterly different traffic. Second, FedAvg takes an unweighted average of client updates with no scrutiny at all, so a handful of Byzantine participants, whether faulty, hijacked or openly adversarial, can quietly steer the shared model wherever they like.

The architecture at the heart of the study is a compact CNN–BiLSTM detector, a small hybrid of convolutional and bidirectional recurrent layers that treats each network flow’s feature vector as a short sequence. The edge configuration uses just sixteen convolutional channels and a sixteen-unit recurrent layer, amounting to roughly 6,300 parameters, about 24.7 kilobytes in 32-bit precision. The researchers trained this detector federally across twenty clients whose data was deliberately skewed using a Dirichlet partition, the standard way to simulate the non-IID reality of edge networks, and then pitted five aggregation rules against one another under three poisoning strategies: sign-flipping, in which adversaries return scaled, sign-reversed honest updates; label-flipping, in which benign and attack labels are swapped locally; and model-replacement, in which adversaries attempt to overwrite the global model with a crafted update.

The scale of the evaluation is what sets the work apart. Across 641 completed training runs, spanning 187 distinct configurations with five random seeds per headline configuration and three for the sweep and privacy tiers, every result is reported as a mean with a 95 percent confidence interval, and every claim of superiority is tested with a paired t-test corrected by the Holm procedure across all 64 comparisons in the family. The Matthews correlation coefficient, rather than F1 score, serves as the primary metric, and the reason is telling. A poisoned model collapses to a constant prediction, and on a balanced task that constant can be all-benign, scoring F1 of zero, or all-malicious, scoring F1 of 0.667. Two failures of identical severity would earn radically different F1 scores; MCC assigns both a deserved zero.

The headline result is stark. On the CICIoT2023 benchmark, every aggregation rule reaches an attack-free F1 between 0.989 and 0.990, confirming that neither the compact model nor the skewed federation is a bottleneck. But introduce a 30 percent sign-flipping adversary and plain FedAvg is annihilated: its Matthews correlation collapses from 0.979 to exactly 0.000 as the global model degenerates into a constant prediction. Three of five seeds produced a model that calls everything benign; two produced one that calls everything malicious. By contrast, trust-weighted aggregation, known as FLTrust, and the multi-Krum rule held Matthews correlations of 0.981 and 0.980 respectively, statistically indistinguishable from each other and from their own attack-free performance. The choice of aggregation rule, the authors conclude, is a security decision, not a tuning detail.

Perhaps the most practically valuable contribution is a sweep of the adversarial fraction from 5 to 45 percent, which locates each rule’s empirical breakdown point, defined as the smallest adversarial fraction at which mean F1 falls below 95 percent of its attack-free value. FedAvg fails at the smallest fraction tested. The coordinate-wise rules, median and trimmed-mean, both break at 20 percent, well inside the range their theoretical tolerance conditions permit, because those conditions are per-round, per-coordinate statements that say nothing about how residual bias accumulates over forty rounds. Multi-Krum breaks at 0.45, precisely the theoretical bound of (K−2)/2K for twenty clients, a rare case of practice landing exactly on theory. FLTrust showed no breakdown anywhere in the tested range, an observation the authors carefully qualify rather than inflate into an unrestricted guarantee.

Not every finding flatters the field. On the harder TON_IoT and Edge-IIoTset benchmarks, the strongest clean configuration was undefended FedAvg itself, at Matthews correlations of 0.838 and 0.802, while every robust rule paid a measurable accuracy premium. Robustness, the authors write, is insurance, and insurance has a premium that only pays off when an adversary actually appears. The study also corrects its own earlier version: a sampling flaw had drawn a contiguous prefix of the source archives, yielding a 91.4 percent attack-heavy sample on which a majority-class predictor scored an apparently excellent F1 while operating at a false-positive rate of 1.0. The corrected class-aware reservoir sampler scanned over 4.2 million rows to build a genuinely balanced 200,000-row sample, and the earlier perfect scores vanished.

The privacy results are equally uncomfortable. A formally accounted, client-level differentially private variant, composed in the Rényi differential privacy framework over forty rounds, failed at every meaningful budget. With twenty clients and full participation there is no privacy amplification by subsampling, and a closed-form noise-to-signal argument explains why: the injected noise has expected norm proportional to z times the square root of the 6,322-dimensional parameter space, roughly 79.5 times the clipped signal norm, so the update survives only at noise levels so small that the composed privacy budget becomes numerically vacuous. Every operating point measured sat at a Matthews correlation at or near zero. The authors also note a subtle interaction: FLTrust rescales updates to the root magnitude, which amplifies rather than attenuates differential-privacy noise, making the most robust aggregator the worst partner for a privacy layer.

The deployment claim, unusually, is a measurement rather than an extrapolation. The researchers ported the trained detector to a physical ESP32 microcontroller, bypassing a microcontroller interpreter that choked on the bidirectional recurrent layer, and instead implemented the forward pass directly in portable C, exploiting the fact that temporal average pooling means recurrent states need only be accumulated, never stored. The result runs at 23.5 milliseconds per inference, about 42.5 flows per second on a single core, using 24.7 kilobytes of flash for weights and just 5.0 kilobytes of scratch RAM, reproducing the host logits to within six parts in a hundred million. That is a per-flow rather than per-packet budget, suited to gateway-side classification, but it demonstrates that a Byzantine-resilient, federated intrusion detector can genuinely live on a two-dollar chip. The full training and evaluation pipeline, including per-round checkpointing so that multi-hour sweeps survive interruption, has been released for reproduction, and the authors point toward adaptive adversaries, partial participation and energy measurement as the next frontiers.

Subject of Research: Byzantine-resilient federated learning for intrusion detection on resource-constrained IoT edge devices under non-IID data conditions

Article Title: Byzantine-resilient federated edge intrusion detection for heterogeneous IoT networks under non-IID conditions

Article References: Nayak, R., & Reddy, G. N. (2026). Byzantine-resilient federated edge intrusion detection for heterogeneous IoT networks under non-IID conditions. Discover Artificial Intelligence, 6(1), Article 1409. https://doi.org/10.1007/s44163-026-02405-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02405-7

Keywords: federated learning, intrusion detection, Internet of Things, Byzantine robustness, non-IID data, edge computing, differential privacy, FLTrust, multi-Krum, ESP32, model poisoning, CICIoT2023

The core insight behind Tiny-STM is that object permanence, the cognitive ability to understand that things continue to exist when hidden from view, does not require a learned memory mechanism of the kind used in large transformer-based systems. Instead, the researchers, Swarnajit Bhattacharya of National Yang Ming Chiao Tung University in Taiwan and Ian Chao of the University of Washington, paired a standard four-gate Long Short-Term Memory network with a deterministic four-slot coordinate bank. Each slot stores the most recently observed three-dimensional position of one task object. When a cube disappears behind a stack of other blocks, its stored coordinate simply holds steady until the camera sees it again, a scheme the authors describe as a zero-order-hold update rule.

The architecture itself is deliberately conventional. The LSTM takes a seven-dimensional input vector combining the vector from the gripper tip to the target object’s stored position, the current tip position, and a one-dimensional infrared range reading from a Sharp GP2Y0A41SK0F sensor mounted on the gripper. A hidden state of 128 dimensions feeds a linear output layer that predicts an incremental three-dimensional displacement of the end effector. With 70,019 parameters in total, the network is tiny by modern standards, and after INT8 quantization through the TensorFlow Lite Micro pipeline it occupies just 68.4 KB, a small fraction of the ESP32’s 520 KB of static RAM.

Training relied on remarkably little data. The authors manually guided a Waveshare 5-DOF RoArm-M3 robotic arm through exactly 30 pick-and-place demonstrations, recording joint angles, cube centroids from a side-mounted camera, and infrared range measurements at each control step. Backpropagation through time with mean-squared-error loss and the Adam optimizer at a learning rate of 0.001 produced the deployed checkpoint. This data efficiency stands in sharp contrast to imitation-learning platforms such as ALOHA, which achieve 80 to 90 percent success on fine-grained bimanual tasks but require roughly ten minutes of teleoperated demonstrations and substantially more powerful compute.

The division of labor between hardware components is a key part of the design. A host PC or single-board computer captures camera frames and performs deterministic color-thresholding segmentation, converting detected pixel centroids into three-dimensional workspace coordinates through a calibrated pinhole camera model. The camera was calibrated with 25 checkerboard images using Zhang’s method, yielding a mean reprojection error of 0.3 pixels, while camera-to-base extrinsic calibration over 12 known points achieved a root-mean-square alignment error of 1.18 millimeters. The ESP32 then handles everything on the control side: the recurrent forward pass, analytical inverse kinematics, and PWM servo command generation, all in a measured 2.80 milliseconds.

The inverse kinematics routine reduces the five-degree-of-freedom arm problem to a planar three-degree-of-freedom calculation, exploiting the fact that the task uses a top-down grasp with a vertically oriented gripper. Base rotation comes from an arctangent of the target coordinates, while the shoulder and elbow angles follow from two-link geometry via the law of cosines. The authors handle three degenerate cases explicitly: base singularity when the target lies on the rotation axis, arm singularity near full extension where the Jacobian becomes ill-conditioned, and unreachable targets outside the geometric annulus defined by the link lengths, which are simply rejected so the arm retains its previous pose.

Physical evaluation consisted of 20 trials in which the arm sorted four labeled cubes from randomized positions into an ordered arrangement, including a vertical stacking sequence in which the fourth block becomes completely hidden behind the first three before its final grasp. The single deployed INT8 checkpoint completed 17 of 20 trials, an observed success rate of 85.0 percent with a two-sided 95 percent exact binomial confidence interval of 62.1 to 96.8 percent. The authors are careful to note that this interval describes uncertainty in the 20 observed outcomes of one checkpoint, not variability across independently trained models, since the original training seed and stopping rule were not retained.

The three failures are instructive. Mechanical backlash in the servo gearboxes introduced joint-level positional uncertainty of roughly 1.5 to 2.5 millimeters, which can accumulate across joints. Fixed color thresholds proved vulnerable to strong lighting changes and shadows that shifted pixel values out of range, degrading the segmentation masks. And when an occluded object was physically moved while hidden, the stored coordinate became stale, causing the controller to reach toward a position where the object no longer was; the gripper tolerance is approximately 18 millimeters. The authors formalize this limitation in a corollary showing that if an occluded object is displaced, the memory error grows at least in proportion to the displacement.

On latency, the measured ESP32 computation of 2.80 milliseconds combines with the nominal 33.3-millisecond camera period at 30 frames per second, profiled segmentation and Wi-Fi communication times, and a datasheet-based actuator response allowance to yield an aggregated end-to-end latency estimate of 92.6 milliseconds, equivalent to roughly 10.8 hertz. The authors emphasize that this figure is an estimate assembled from components rather than a single timed end-to-end measurement. Peak SRAM usage on the microcontroller was 94.8 KB, or 18.23 percent of capacity, leaving substantial headroom. Offline comparisons suggest the FP32 model would occupy 273.5 KB with an estimated 11.20 milliseconds of inference, underscoring the value of INT8 quantization for embedded deployment.

The study is notable as much for its honesty about boundaries as for its results. The authors explicitly state that no ablation was performed to separate the contribution of the coordinate memory from that of the recurrent state, that no numerical comparison with RT-2, OpenVLA, TinyVLA, or Diffusion Policy is possible given different tasks and hardware, and that the color-thresholding front end was never validated against ground-truth masks. Future work, they write, should include matched ablation studies with multiple training seeds, documented reproducibility protocols, perception validation, tests of behavior when occluded objects are displaced, and scaling experiments beyond four tracked objects. Even so, Tiny-STM offers a compelling demonstration that a form of object permanence can live in 68 kilobytes, bringing occlusion-aware manipulation within reach of the cheapest embedded hardware.

Subject of Research: Lightweight recurrent neural network controllers with external coordinate memory for occlusion-aware robotic manipulation on microcontrollers

Article Title: Lightweight LSTM-based memory architecture for occlusion-aware robotic manipulation on microcontrollers

Article References: Bhattacharya, S., & Chao, I. (2026). Lightweight LSTM-based memory architecture for occlusion-aware robotic manipulation on microcontrollers. Discover Informatics, 1(1), Article 23. https://doi.org/10.1007/s44564-026-00023-0

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00023-0

Keywords: TinyML, LSTM, robotic manipulation, object permanence, occlusion, ESP32, microcontroller, imitation learning, inverse kinematics, INT8 quantization, embedded robotics, recurrent neural networks

The stakes are considerable. Nitrogen, phosphorus, and potassium regulate chlorophyll synthesis, root development, photosynthesis, enzyme activation, and osmotic regulation in crops, making them the backbone of any fertilizer program. Yet across much of the world, fertility assessment still depends on labor-intensive laboratory chemistry. In India, the Soil Health Card (SHC) program provides standardized laboratory testing and fertilizer recommendations, but sample transport, processing, and delayed reporting make it poorly suited to frequent monitoring, particularly for resource-constrained farmers. Meanwhile, intensive cultivation and indiscriminate fertilizer use have degraded soils in many agricultural regions, sharpening the need for timely, site-specific nutrient information that precision agriculture demands.

Low-cost electrochemical NPK sensors have long promised a solution, but they carry a well-known flaw: their readings drift with the very conditions they are embedded in. Sensor responses are influenced not only by nutrient concentrations but also by soil texture, moisture content, salinity, pH, ionic mobility, and electrical conductivity. Identical nutrient levels can produce different electrical signatures in a wet clay soil and a dry sandy one. As a result, raw sensor outputs deviate nonlinearly from laboratory measurements, and models trained on one dataset often fail when applied to heterogeneous fields. The Dharwad team’s central insight was that this physicochemical interference should not be treated as noise to be minimized, but as information to be modeled explicitly.

To build that model, the researchers assembled a dataset grounded in real agricultural variability. They collected 150 representative soil samples from upland, midland, and lowland zones of Dharwad district in Karnataka, India—fifty from each physiographic region—drawing from the surface 0–20 centimeter layer that constitutes the active root zone for most crops. Within each field, multiple subsamples were gathered in a zigzag pattern and mixed into composite samples, following standard agronomic practice. Each sample was then measured twice: once in the field with the sensor platform, and once at an authorized Soil Health Card laboratory, creating synchronized pairs of raw sensor observations and laboratory reference values for supervised learning.

The sensing hardware itself is deliberately modest. A commercial 7-in-1 electrochemical probe measures nitrogen, phosphorus, potassium, pH, electrical conductivity, moisture, and temperature simultaneously. It communicates over an RS485 serial link, via a MAX485 transceiver running the Modbus RTU protocol, to an ESP32 microcontroller programmed through the Arduino IDE. In the laboratory the system ran on a regulated 12-volt supply; in the field it drew power from portable 12-volt rechargeable batteries, allowing continuous acquisition without external infrastructure. Sensor outputs, reported in milligrams per kilogram, were converted to kilograms per hectare using a soil-mass conversion factor of 2.24, corresponding to an assumed bulk density of 1.12 grams per cubic centimeter, so that predictions would align with SHC reporting conventions.

The calibration engine is a MultiOutput XGBoost regression model, an ensemble method built on gradient-boosted decision trees. Because standard XGBoost handles only a single target, the team used Scikit-learn’s MultiOutputRegressor wrapper to train three independent XGBoost regressors—one each for nitrogen, phosphorus, and potassium—sharing the same engineered feature matrix and identical hyperparameters. The final configuration, selected through five-fold cross-validation on the training data, used 200 trees, a learning rate of 0.05, a maximum tree depth of 6, a minimum child weight of 3, and subsampling and column-sampling ratios of 0.80, with L2 regularization set to 1.0. Regularization and subsampling are what allow the model to capture complex nonlinear feature interactions while resisting overfitting on a modest dataset.

The most conceptually elegant contribution is a single engineered feature: the product of soil pH and electrical conductivity. The rationale is rooted in soil chemistry. pH governs nutrient solubility, ionic speciation, and phosphorus fixation, while electrical conductivity reflects the dissolved ion concentration that drives current flow during electrochemical measurement. Because the sensor operates by measuring ionic responses, the combined pH–EC effect carries information that neither variable reveals alone. An ablation study confirmed the value of this domain knowledge: adding the pH × EC interaction raised the coefficient of determination from 0.903 to 0.916 and reduced root mean square error from 8.92 to 8.37 kilograms per hectare, a 6.17 percent error reduction under identical training conditions.

On the held-out test set of 30 samples, the calibrated model delivered laboratory-equivalent estimates with R² values of 0.916 for nitrogen, 0.887 for phosphorus, and 0.903 for potassium, with root mean square errors of 27.11, 4.01, and 22.84 kilograms per hectare respectively. Nitrogen performed best, likely because its electrochemical response was more consistent, while phosphorus lagged slightly, consistent with its low mobility and strong adsorption–desorption behavior in soils. The framework outperformed linear regression, decision trees, random forests, and support vector regression on the same data, achieving an average R² of 0.902 against 0.881 for random forest and 0.766 for linear regression. Five-fold cross-validation showed remarkable stability, with standard deviations of only about 0.003 in R², and 95 percent confidence intervals remained narrow across all three nutrients.

Equally important is how the team interrogated their own model. Residual analysis found mean errors close to zero—0.18, −0.05, and 0.24 kilograms per hectare for nitrogen, phosphorus, and potassium—with both positive and negative deviations, indicating no systematic over- or underestimation. Bland–Altman agreement analysis, a technique borrowed from clinical measurement science, showed that differences between calibrated predictions and laboratory values clustered around zero, with most observations falling within the 95 percent limits of agreement. A region-wise breakdown confirmed that performance held up across upland, midland, and lowland soils, with nitrogen R² ranging only from 0.899 to 0.928 across zones—evidence that the calibration generalizes across the physiographic heterogeneity it was designed to capture.

The authors are candid about limitations. The study did not quantify long-term sensor drift from electrode aging and environmental exposure, a known problem for low-cost electrochemical probes; prior work cited in the paper shows that inter- and intra-sensor variability can be substantial and that continuous, sensor-specific calibration reduces measurement error. The laboratory reference values themselves lacked an explicit uncertainty budget, and the model was trained on soils from a single district, so validation across diverse agro-climatic zones will be needed before large-scale deployment. Future work will extend the framework to organic carbon, micronutrients, and cation exchange capacity, and will explore adaptive drift compensation, explainable AI, and edge deployment on embedded IoT platforms. Even so, the demonstration stands: for a few dollars of hardware and a well-designed learning algorithm, the gap between instant field readings and trusted laboratory numbers has been narrowed to a level that could put real-time soil intelligence within reach of every farmer.

Subject of Research: Machine learning calibration of low-cost electrochemical sensors for real-time soil macronutrient estimation

Article Title: Real-time soil nutrient estimation using low-cost NPK sensors and soil-aware machine learning calibration

Article References: Kabbur, A. M., & Pawar, P. (2026). Real-time soil nutrient estimation using low-cost NPK sensors and soil-aware machine learning calibration. Smart Agricultural Technology, 15, Article 102582. https://doi.org/10.1016/j.atech.2026.102582

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102582

Keywords: soil sensors, NPK nutrients, machine learning, XGBoost, precision agriculture, IoT, ESP32, Soil Health Card, soil fertility, sensor calibration, electrochemical sensing, India

The choice of a hexacopter rather than the more common quadcopter is not cosmetic. With six brushless motors driving six propellers in a Hexa-X layout, the aircraft retains controlled flight even if one rotor fails, a margin of safety that matters enormously when the vehicle is dispatched over rubble, smoke or unstable terrain where a crash means losing both the airframe and the mapping data it carries. The redundant rotor configuration also increases total payload capacity, which is what allows the platform to carry multiple sensing systems at once. In disaster scenarios the aircraft can be deployed quickly without runways or significant ground infrastructure, and its maneuverability lets it thread through confined spaces and reach isolated zones that ground vehicles cannot access. The authors also emphasize the operating economics: small multirotors cost a fraction of larger aerial platforms, making repeated reconnaissance sorties financially realistic for agencies that must monitor an affected area many times over.

The thermal imaging payload centers on the FLIR Lepton 3.5, a remarkably small radiometric thermal camera that measures the actual temperature of every pixel rather than merely rendering a heat picture. The sensor is mounted on a tCam-Mini rev4 breakout board built around an ESP32-WROVER microcontroller with an onboard antenna, an open-source hardware and software design that streams the radiometric data over WiFi to a ground computer. Before flight, the camera’s housing is produced by 3D printing, with four STL components and two copies of a vertical holder printed to form an enclosure that shields the delicate sensor from vibration and mechanical stress during operation. Firmware is flashed to the ESP32 board using Espressif’s serial download tool, loading a bootloader, a partition table and the main tCam firmware in sequence; a dual-color LED confirms a successful boot by briefly turning red and then blinking yellow as the camera begins advertising its own WiFi access point. The companion desktop application then connects to the camera’s default address and displays the live thermal feed, allowing operators to detect human body heat and other thermal anomalies in real time even through smoke, darkness or visual clutter that would defeat an ordinary camera.

The mapping side of the system uses the Garmin LiDAR-Lite V3, a compact and lightweight laser rangefinder chosen for its accurate distance measurements and wide field of view. Because a single fixed sensor sees only one direction at a time, the team mounted it on a pan-and-tilt assembly built from two HS-422 servo motors and an aluminium bracket kit, letting the laser sweep across the terrain below. An Arduino Uno orchestrates the scanning: it toggles the sensor’s enable pin, manages the I2C clock and data lines used for communication, and arbitrates the mode-control pin through parallel resistors to prevent bus contention, all while driving the two servos through PWM signals. As the sensor sweeps, distance readings are transmitted over serial to a ground station, where an open-source interface written in C++ with the OpenGL graphics library converts each measurement into Cartesian coordinates and renders it as a cube whose color encodes angle and distance. The operator can fly through the emerging 3D point cloud while scanning is still in progress, and completed scans can be saved to file and reloaded for later analysis, producing the kind of terrain model that helps responders understand where paths, debris fields and structures lie.

The airframe itself follows a build path that any reasonably well-equipped workshop could follow. Six A2212 brushless motors rated at 1000 KV are bolted to the arms of an F550 hexacopter frame, with their bullet connectors routed through the arms before the electronic speed controllers are positioned. A power distribution board mounted on the lower frame plate feeds all six speed controllers, and every solder joint is tinned with flux to prevent the cold joints that can fail mid-flight. The top plate then encloses the electronics, with the battery carefully balanced along the centerline. An APM 2.8 flight controller with a built-in compass coordinates the six rotors, receiving commands from a Flysky FS-i6X 2.4 GHz transmitter and its ten-channel receiver. The team configured and calibrated the entire flight system using Mission Planner, an open-source ground-control application: firmware for the Hexa X frame is installed over a serial connection, the accelerometer and compass are calibrated by rotating the vehicle until calibration markers clear, the radio sticks are swept to their extremes to record their ranges, and the electronic speed controllers are programmed through a sequence of throttle-position beeps. Anti-vibration dampers sit between the frame and the controller to keep sensor readings clean.

Field validation put the whole assembly through its paces. The team performed pilot testing and initial flight calibrations before moving to full flight and field testing of the integrated platform, confirming aerial monitoring, environmental sensing and real-time wireless transmission of the thermal feed. Wireless transmission tests measured system performance characteristics such as the maximum achievable frame rate for the life-form detection stream, data that defines how quickly an operator on the ground can see a heat signature appear. The platform’s architecture also leaves room for growth: the authors note that the airframe can support a gripper mechanism for object retrieval, temperature and humidity sensors for environmental monitoring, and integration with Geographic Information System databases to sharpen situational awareness during response operations. Multi-sensor fusion of the LiDAR, thermal and environmental streams is presented as the path toward a comprehensive picture of a disaster zone, and the communication module could in principle relay temporary network connectivity into areas where infrastructure has been knocked out.

The build is also refreshingly honest about its limits, several of which will be familiar to anyone who has tried to put serious sensors on a small drone. The single biggest disappointment came during flight testing of the LiDAR: the team was unable to operate the scanner airborne because of the platform’s significant payload capacity and power limitations, a constraint the authors note is common to compact UAV-based LiDAR systems. Continuous LiDAR operation demands sustained laser emission and real-time computation, while the thermal camera needs uninterrupted power for image acquisition, so balancing sensor operation against flight stability and mission duration remains a central engineering challenge. The Garmin sensor itself is also slow: a complete scan yields roughly 92,000 data points, a process that can stretch over several hours when performed thoroughly. On the ground, that is acceptable; in the air, it would be impractical without a faster scanning approach or a lighter, more efficient sensor.

Other constraints are more exotic but no less instructive. The LiDAR’s laser ranging degrades on highly reflective surfaces and transparent materials such as glass, which scatter or pass through the beam rather than returning a clean reflection. Reliable serial communication required fixed-length data strings, so the software pads angular measurements and sensor values with leading zeros, an unglamorous detail that nonetheless determines whether the ground station parses the stream correctly. Perhaps most surprisingly for a civilian disaster tool, the FLIR Lepton 3.5 and certain LiDAR technologies fall under the International Traffic in Arms Regulations, the United States export-control regime for sensitive technologies. Complying with ITAR imposed administrative burdens that affected the project’s timelines, supplier selection and collaboration opportunities, and even sharing LiDAR-derived data with foreign entities required careful legal review. For a low-budget undergraduate project, these regulatory hurdles were as real an obstacle as any engineering problem.

What makes the publication notable is less any single component than the demonstrated end-to-end feasibility of the whole stack. A six-rotor aircraft with rotor-failure redundancy, a radiometric thermal camera that can spot living organisms through conditions that blind optical cameras, a LiDAR system that builds navigable 3D terrain models, and wireless streaming of the thermal data to an operator on the ground were all assembled for a total bill of materials of Rs. 73,689.35, with the single most expensive item, the LiDAR-Lite V3 at Rs. 23,215, costing less than many commercial camera drones on its own. All design files, from the LiDAR schematic and Arduino code to the STL enclosure prints and the tCam-Mini firmware, are archived in Zenodo repositories under Creative Commons and GNU General Public licenses. The authors frame the modular platform as a foundation for continued research in drone-assisted disaster management, with future work pointing toward machine-learning techniques for autonomous navigation that would reduce response times and keep human rescuers out of hazardous zones. For research groups and volunteer response teams in resource-constrained settings, it is a blueprint that turns a disaster-response capability once reserved for well-funded agencies into something a small team can build, fly and improve.

Subject of Research: Low-cost hexacopter platform integrating thermal imaging and LiDAR mapping for disaster management

Article Title: Thermal imaging and LiDAR mapping for disaster management with hexacopter

Article References: Vuucha, H., Jomon, J., Somasundaram, D., & M., M. S. (2026). Thermal imaging and LiDAR mapping for disaster management with hexacopter. HardwareX, Article e00841. https://doi.org/10.1016/j.ohx.2026.e00841

Image Credits: AI Generated

DOI: 10.1016/j.ohx.2026.e00841

Keywords: hexacopter, thermal imaging, LiDAR, disaster management, FLIR Lepton 3.5, Garmin LiDAR-Lite V3, open-source hardware, UAV, search and rescue, ESP32, 3D point cloud, Mission Planner

The stakes are enormous. Cardiovascular diseases claimed an estimated 19.8 million lives in 2022, roughly 32 percent of all deaths worldwide, and 85 percent of those deaths were caused by acute myocardial infarction and stroke. Heart attacks can be treated with high survival probabilities when patients receive early attention, but symptoms are not always obvious. Chest pain is the classic warning, yet subtler signs—breathing difficulty, dizziness, tachycardia, sweating—are frequently missed or dismissed. Even when patients reach a hospital, standard diagnostic protocols have limits: electrocardiography detects significant ST-segment changes in only about 57 percent of heart attacks, and some patients never return after discharge. The researchers argue that prevention and early detection depend on affordable technology that can watch the cardiovascular system continuously, in daily life, rather than in a single hospital snapshot.

The system the team built consists of two synchronized devices. The first is a wristband built around a MAX30105 photoplethysmography sensor, a reflectance-mode optical sensor that shines red, green, and infrared light into the skin to capture the cardiac pulse wave and estimate blood oxygen saturation. The second is a chest band featuring a novel optical respiratory-detection mechanism alongside an MPU6050 inertial measurement unit that tracks body movement. Both devices are driven by ESP32-C3 microcontrollers—compact 32-bit RISC-V chips with built-in Wi-Fi and Bluetooth 5.0 Low Energy—and stream data to a custom mobile application, which in turn uploads everything to a cloud database. The total component cost of both devices comes to $82.88 at single-unit distributor prices, a figure the researchers note is one to two orders of magnitude below the clinical instruments they validated against.

The chest band’s respiratory sensor is the system’s most inventive element. Instead of impedance electrodes or strain gauges, the team designed a retractable optical assembly: a 3-millimeter red LED paired with a photoresistor, mounted so that thoracic expansion during inspiration pulls the band’s sliding structures apart and allows light to pass. At rest, the structures overlap and block the light; when the chest expands by up to 12 millimeters, light leaks through and the photoresistor registers the change. The microcontroller’s 12-bit analog-to-digital converter reads values ranging from 215 in the fully contracted position to 4095 when fully stretched—a resolution of roughly 3 micrometers. Each upward crossing of a threshold set at 500 counts corresponds to one complete inspiration, so counting these transitions over a 60-second window yields the respiratory rate directly from mechanical chest-wall displacement.

That direct mechanical approach appears to pay off in accuracy. Respiratory rate measured by the chest band showed a mean absolute percentage error of 3.47 percent after physical activity and 4.69 percent at rest, with a Pearson correlation of 0.99 against the ADInstruments physiograph reference. The post-activity root-mean-square error of 1.14 breaths per minute is lower than the 1.5 to 3.5 breaths per minute typically reported for impedance-based wearable systems, which the researchers attribute to the optical method’s resistance to motion-induced degradation. Heart rate, derived offline from the raw photoplethysmographic waveform using a zero-phase Butterworth band-pass filter and dynamic peak detection, achieved errors of 1.99 percent at rest and 1.94 percent post-activity, with Bland–Altman limits of agreement spanning roughly ±5 beats per minute—performance comparable to leading commercial wrist-worn optical sensors such as the Polar OH1.

Blood oxygen saturation told a more complicated story, and the researchers are unusually candid about it. The wristband’s SpO2 error was small—1.64 percent at rest and 2.98 percent after exercise—yet its correlation with the reference monitor was weak, just 0.35 at rest and 0.21 post-activity. The team explains this apparent contradiction as range restriction: in a cohort of healthy young adults, reference saturation varied only between 94 and 100 percent, a spread so narrow that the correlation coefficient is mathematically driven toward zero regardless of how well the devices agree. Correlation measures association, not agreement, and the Bland–Altman analysis is the appropriate yardstick here. What the correlation genuinely obscured, however, was a real degradation after exercise: five of twenty post-activity acquisitions deviated by five percentage points or more, with a maximum underestimation of nine points, likely reflecting altered peripheral perfusion and a sensor configuration optimized for pulse-wave quality rather than oximetry. The researchers explicitly state the device is not yet suitable for applications requiring absolute saturation values, and validation under controlled desaturation per the ISO 80601-2-61 standard remains outstanding.

The validation study itself involved 43 healthy volunteers aged 19 to 28, who produced 50 acquisitions: 30 at rest and 20 after running approximately 100 meters. Each participant lay supine and motionless for one minute while wearing both prototype devices alongside the ADInstruments physiograph and the EDAN iM8 vital-signs monitor. The inertial unit was tested separately across six activity conditions, from sitting motionless to running, and a movement-percentage threshold of 40 percent cleanly separated rest from activity: sitting registered 3 percent, walking 64 percent, and running 87 percent, with no false positives at rest. This movement context matters, the researchers argue, because it allows cardiovascular algorithms to distinguish genuine physiological changes from motion artifacts—potentially reducing false alarms in future risk-detection systems.

What distinguishes this platform from consumer smartwatches is architectural. Devices like the Apple Watch and Garmin smartwatches calculate derived metrics but do not expose raw photoplethysmographic signals, which are essential for developing more sophisticated algorithms. The Colombian-Brazilian system deliberately transmits complete raw pulse and respiration waveforms to the cloud, leaving filtering, feature extraction, and analysis to the researcher rather than fixing them in firmware. As processing methods evolve, the same acquisitions can be re-analyzed with new techniques—statistical, spectral, or non-linear heart-rate-variability indices, for instance—without changing the hardware. The devices can even upload autonomously over Wi-Fi without an intermediary smartphone, an advantage for older adults and patients with limited mobility who are the system’s ultimate target users.

The team is equally forthright about the study’s limits. This is a technical feasibility study in healthy young volunteers, not a clinical validation. No participant had cardiovascular disease, no elderly participants were included, and skin phototype and body mass index were not recorded—a material omission given documented disparities in optical sensing accuracy across skin tones. The optical respiratory sensor also showed more variable errors in female participants, with limits of agreement 1.7 times wider than in males, which the researchers attribute plausibly to diaphragmatic breathing patterns and mechanical interference of breast tissue with band positioning, though the difference did not reach statistical significance. All acquisitions were performed with participants supine and stationary, so performance under free-living ambulatory conditions—where motion artifact dominates—remains untested. Three acquisitions even had to be repeated after wristband battery depletion produced implausible readings, motivating future battery-state monitoring and signal-quality indices.

The pathway from here is already in motion. A data-collection study with a consultant cardiologist at Clínica Valle de Pubenza in Popayán, Colombia, has assessed 50 high-risk older adults using the system, and a preliminary risk-estimation model has been trained—though the researchers deliberately report that its performance does not yet reach clinical-application range, owing to the small labeled dataset. The long-term vision is twofold: rapid, low-burden cardiovascular screening in settings where access to cardiology is limited, and home-based longitudinal monitoring of patients already at high risk. If the platform’s accuracy holds in the populations it was designed for, an $83 wearable that streams raw cardiac and respiratory waveforms to the cloud could become the foundation for machine-learning systems that flag a heart attack before it strikes—turning the wrist into a window on the heart that no hospital waiting room can match.

Subject of Research: Development and validation of a low-cost wearable system for continuous cardiorespiratory monitoring using photoplethysmography and optical respiratory detection

Article Title: Development and validation of an integrated wearable system for continuous cardiorespiratory monitoring using photoplethysmography and optical respiratory detection

Article References: Ramírez, J. D., Toro, N., Campo, J. M., Gómez-Peña, G., Villamarín-Muñoz, J. A., & Mosquera-Sánchez, J. A. (2026). Development and validation of an integrated wearable system for continuous cardiorespiratory monitoring using photoplethysmography and optical respiratory detection. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-026-03677-y

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03677-y

Keywords: wearable technology, photoplethysmography, cardiorespiratory monitoring, heart attack detection, SpO2, respiratory rate, heart rate variability, Bland-Altman analysis, ESP32, mobile health, cardiovascular risk, biomedical sensors

The timing of such a system is hardly accidental. Irrigated agriculture already accounts for approximately 70 percent of global freshwater withdrawals, while arable land continues to shrink under pressure from degradation and rapid urbanization. The Food and Agriculture Organization estimates that global agricultural production must increase by about 50 percent by 2050 relative to 2012 levels to feed a growing population. Controlled-environment agriculture, including vertical farming in urban buildings, has emerged as one strategy to raise productivity while cutting resource consumption, but its energy appetite has remained a persistent obstacle, with energy representing between 10 and 40 percent of total production costs in commercial greenhouses.

Among soilless cultivation techniques, aeroponics stands out for its extreme water efficiency. Unlike hydroponics, where roots sit submerged in nutrient solution, or aquaponics, which couples plant cultivation with fish farming, aeroponics suspends bare roots inside a closed chamber and sprays them with a fine nutrient mist. Reported water savings reach 90 to 98 percent compared with soil-based cultivation, while direct exposure to oxygen enhances root respiration and nutrient uptake. The technique also reduces exposure to soil-borne diseases, allows precise moisture control, and has even attracted attention from space agencies for food production under microgravity.

Yet aeroponics has a critical vulnerability: it depends on continuous electrical power. Because roots hang in air rather than water, a power outage can desiccate a crop within hours. Commercial vertical aeroponic towers, such as the Tower Garden FLEX, the Nutraponics Pro AeroTower, and the Agrotonomy Tower Farm system, all rely on grid electricity and proprietary designs, and none include integrated filtration to protect their misting nozzles. Prices range from $765 to $5,250 per unit, and academic prototypes published in the literature typically omit fabrication files, making them difficult or impossible to reproduce.

TOTEM was designed to close these gaps. The system stacks eight mechanically coupled modules, each housed in standard 20-centimeter-diameter PVC components: growth, reservoir, pump, filter, inlet-drainage, control, power, and solar. The growth module is a 1.5-meter PVC tower holding 24 net cups arranged in six staggered rows, with modified 45-degree elbow fittings angling each plant site to optimize light distribution. Inside the chamber, a vertical distribution pipe carries pressurized nutrient solution past 24 misting nozzles that atomize it into fine droplets directed at the suspended roots.

The hydraulic loop is deliberately simple and closed. A 12-volt DC diaphragm pump draws solution from the reservoir and pushes it through an AZUD Modular 100 mesh filter, which captures suspended particles before they can clog the nozzles, a failure mode that plagues many commercial systems. Filtered solution travels up the internal pipe, is atomized across the root zones, and the excess drains by gravity back into the reservoir for recirculation. During the basil trial, this closed loop required only about 54 liters of nutrient solution for 24 plants over the entire 55-day cycle, an estimated 0.041 liters per plant per day, far below reference values of roughly 0.6 liters per plant per day for protected agriculture and 12.6 liters for open-field cultivation.

At the system’s electronic core sits an ESP32-WROOM microcontroller mounted on a custom printed circuit board alongside an LM2596 voltage regulator, an FOD852 optocoupler, and an RAS1220M relay that switches the pump. A waterproof DS18B20 digital temperature sensor tracks conditions inside the growth chamber while a water-level sensor monitors the reservoir, enabling basic fault detection such as low-solution alerts. The microcontroller executes an intermittent irrigation schedule, in the validation trial running the pump for 0.5 minutes followed by 29.5 minutes of rest, equivalent to 48 activation cycles per day. Over Wi-Fi, the controller transmits data to a remote server that logs performance history and serves a web-based graphical user interface, allowing users to adjust irrigation parameters and monitor the system remotely, a capability none of the surveyed commercial towers offer.

Energy autonomy comes from a 50-watt polycrystalline solar panel mounted on a three-meter galvanized steel mast, angled at 24 degrees and oriented southward for maximum insolation. Power flows through a 40-amp MPPT charge controller into two 12-volt, 12-amp-hour rechargeable batteries wired in parallel, providing 288 watt-hours of nominal storage and roughly 122 watt-hours of usable energy at a conservative 50 percent depth of discharge. That corresponds to approximately 2.7 days of autonomous operation without any sunshine. The designers sized the array against the worst month of the year in La Paz, Baja California Sur, where December delivers only 3.9 peak sun hours per day; even with a photovoltaic derating factor of 0.70, the panel supplies an estimated 136.5 watt-hours daily, comfortably exceeding the calculated demand of 44.9 watt-hours after accounting for the pump’s 19.2 watt-hours, the continuously active microcontroller’s 19.0 watt-hours, and 85 percent system efficiency.

The validation experiment grew basil, a fast-growing crop prized in controlled-environment agriculture for its sensitivity to water availability, under uncontrolled shade-greenhouse conditions. After transplanting, formative pruning at 30 days promoted lateral branching, followed by a first harvest at 45 days and a final harvest at day 55. Sampled plants averaged 517.97 square centimeters of leaf area, 31.33 leaves, an 18.8-centimeter shoot length, and a mean relative water content of 82.18 percent, indicating that intermittent misting maintained adequate hydration throughout. Chamber temperatures fluctuated naturally between 14 and 38 degrees Celsius, and the system completed every programmed irrigation cycle without interruption across the full cultivation period.

The authors are candid about limitations. The final nutrient solution pH drifted to 7.39, above the mildly acidic range typically targeted in soilless production, and the team recommends adding calibrated pH and electrical conductivity sensors with closed-loop control in future iterations. The single-cycle trial demonstrated feasibility rather than long-term endurance, and the growth measurements from three plants provide descriptive data rather than statistically rigorous agronomic comparison. Still, the study’s principal contribution lies in its radical transparency: complete STEP CAD files for every module, circuit schematics, PCB layouts, ESP32 firmware in C++, and web interface code are released under CERN Open Hardware License v2-S and GNU General Public License v3, accompanied by an itemized bill of materials totaling $719.51. With a footprint of just 0.078 cubic meters, less than a ninth of comparable commercial towers, TOTEM makes off-grid aeroponic farming something a workshop, a school, or a resource-limited community can actually build, repair, and adapt for itself.

Subject of Research: An open-source, low-cost, solar-powered aeroponic vertical farming system for sustainable urban agriculture

Article Title: TOTEM: A low-cost solar-powered aeroponic system for vertical agriculture

Article References: Von Borstel, F. D., Villa-Medina, J. F., Nieto-Garibay, A., & Gutiérrez, J. (2026). TOTEM: A low-cost solar-powered aeroponic system for vertical agriculture. HardwareX, 28, Article e00838. https://doi.org/10.1016/j.ohx.2026.e00838

Image Credits: AI Generated

DOI: 10.1016/j.ohx.2026.e00838

Keywords: aeroponics, vertical farming, open-source hardware, solar power, urban agriculture, ESP32, water efficiency, controlled environment agriculture, basil cultivation, food security, IoT, sustainable agriculture

The platform, called AgroYachay—’yachay’ meaning knowledge in Quechua—is described in the journal SoftwareX and released under the MIT licence with an archived DOI, version 1.0.1. It is the cloud counterpart of a companion tool, AgroCommish, which handles the upstream work of commissioning ESP32 sensor nodes: flashing firmware, discovering sensor pins, provisioning WiFi credentials and registering each verified device with the cloud. Together, the two tools form a complete, reproducible device-to-decision pipeline that a small research group or agricultural programme can deploy entirely on open-source software and commodity hardware, with no per-seat licensing or proprietary cloud lock-in.

Architecturally, AgroYachay follows a three-tier edge-cloud design. At the edge, ESP32 microcontrollers wired to a DHT11 sensor for air temperature and humidity and an FC-28 probe for soil moisture post JSON readings over WiFi every ten seconds. The backend is a Flask 3 application of roughly 7,300 lines of Python, organised into controllers, feature modules and a service layer, with state persisted across fifteen normalised tables in PostgreSQL 17. Authentication relies on JSON Web Tokens with bcrypt password hashing and optional Google OAuth. The frontend is a React 18 single-page application of about 13,000 lines styled with Tailwind CSS, charting live sensor series with Recharts. More than ninety REST API routes tie the system together.

The intelligence layer is deliberately hybrid. Image-based pest diagnosis runs on a self-hosted open multimodal model—qwen2.5-VL-3B, served through Ollama—which keeps farmers’ photographs on local hardware rather than shipping them to a third-party image service. A grower photographs an affected leaf; the vision model, prompted as a phytopathologist specialising in Andean crops, returns a structured verdict covering health status, disease name, confidence, severity, causes, treatment and prevention. A follow-up text-only call to a cloud large language model hosted by Groq converts that verdict into an actionable management plan: urgency, step-by-step actions, chemical products with doses available in Peru, an organic alternative, and guidance on when to consult a human agronomist. The prompt targets potato, quinoa, maize, faba bean, oca and cañihua, the staple crops of the region.

The team validated the vision module with unusual candour, evaluating it as a plant-health triage aid rather than a fine-grained diagnostic tool. Using 509 images—300 laboratory photographs from PlantVillage and 209 field images from PlantDoc—they found perfect sensitivity: every one of the 409 diseased leaves was flagged as diseased, the property that matters most for an early-warning system. Specificity was lower at 62 percent, meaning the model over-flags healthy leaves as diseased, a conservative bias favouring missed nothing over missed disease. Species-level discrimination was weaker, with late blight frequently confused with early blight and a macro-F1 of 0.55. The authors state plainly that the module supports binary field triage, not reliable differential diagnosis, and that performance on real Andean field imagery remains untested. Repeated queries at the deployed temperature of zero produced identical classifications in ten out of ten trials, confirming reproducibility.

Beyond diagnosis, a context-aware conversational assistant answers free-form questions in Spanish, conditioned on the farmer’s registered crops, region and latest sensor values. Separate services translate the five-day OpenWeather forecast into a risk level, a weekly activity plan and optimal-day recommendations for spraying, sowing and irrigation. A crop management module records species, variety, parcel area, GPS location, sowing date and phenological stage, ensuring that every reading, alert, diagnosis and estimate is tied to a specific crop and device. An inputs module calculates fertiliser and agrochemical requirements from parcel area and target yield, while an advisory module queues requests for human agronomists, keeping an expert in the loop when the automated tools should not act alone.

The platform’s most distinctive feature may be its economic layer. Rather than a black-box regressor, the yield-and-revenue estimator is a transparent multiplicative factor model requiring no training data, which suits the sparse-data reality of Andean deployment. A crop-specific base yield—15.5 tonnes per hectare for potato, 1.8 for quinoa, 8.5 for oca—is adjusted by climate factors computed from mean temperature, humidity and accumulated rainfall within phenological windows, by a phenological factor that grows with crop progress, and by a piecewise area factor. The projected tonnage is multiplied by a regional reference price to give expected gross revenue. Because the model is purely multiplicative, its sensitivity is fully transparent: a 10 percent error in any climate factor moves the estimate by exactly 10 percent, and every contributing factor is displayed to the user. Results include pessimistic, probable and optimistic scenarios with a confidence score, and can be exported as styled PDF reports suitable for credit or crop-insurance applications.

The multilingual interface is treated as a first-class feature rather than an afterthought. Of 1,488 UI strings, 94 percent are localised into both Quechua and Aymara, with a complete English and Spanish locale; only technical terms such as ‘IoT’ and ‘ESP32’ are deliberately left untranslated. This matters because a substantial share of Andean smallholders operates primarily in indigenous languages, and generic farm-management platforms assume Spanish or English literacy. The authors note an honest limitation here: while the interface is multilingual, the LLM-generated advice, pest verdicts and exported reports are currently produced in Spanish only, so a Quechua-speaking user still encounters Spanish text at the moment of advice. Extending the prompts and report templates to indigenous languages is the next inclusion step.

Performance measurements on the production server show that core operations are genuinely interactive: lightweight REST calls complete in 2.6 milliseconds at the median, dashboard reads in under 10 milliseconds, and sensor ingestion in 19 milliseconds. A 22-minute bench test of a real ESP32 node delivered 132 of 134 expected sampling cycles, a 98.5 percent success rate with a maximum reconnection gap of 20 seconds. The exception is the vision module: on the CPU-only demonstration server a single diagnosis takes minutes, whereas on a consumer GPU it completes in seconds, so GPU-backed deployment is recommended for interactive use. The conversational assistant, dominated by external inference, averages around two seconds per response.

The developers, based at the Universidad Nacional del Altiplano in Puno, drew requirements from direct field experience in a smallholder economy centred on potato, quinoa, cañihua and Andean livestock, where commercial precision-agriculture suites are cost-prohibitive. They self-tested the platform on their own plantings and received positive informal feedback from local farmer demonstrations, though a formal quantitative user-acceptance study remains future work. So do calibration of the yield estimator against actual harvest records, an expert-labelled Andean field image benchmark, an on-device offline fallback for pest screening, and security hardening measures such as per-device authentication, database encryption at rest and automated secret rotation. Even without the LLM layer, sensor monitoring, alerts, the factor-model estimator and reporting already function. As an early but complete contribution, AgroYachay demonstrates that the entire chain—from flashing a sensor in a highland workshop to issuing a financed-harvest report—can now run on open-source software, potentially reshaping who gets to participate in precision agriculture across Latin America and other developing regions.

Subject of Research: An open-source IoT and large-language-model platform supporting agronomic and economic decision-making for Andean smallholder farmers.

Article Title: AgroYachay: An open-source IoT and large-language-model platform supporting agronomic and economic decision-making for Andean smallholders

Article References: Torres-Cruz, F., Vilca-Solorzano, R. A., Yana-Yucra, D. M., Ibañez-Quispe, V., & Fuentes-Navarro, E. L. (2026). AgroYachay: An open-source IoT and large-language-model platform supporting agronomic and economic decision-making for Andean smallholders. SoftwareX, 36, Article 103037. https://doi.org/10.1016/j.softx.2026.103037

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103037

Keywords: AgroYachay, precision agriculture, IoT, ESP32, large language models, smallholder farming, Andean agriculture, plant disease diagnosis, open-source software, yield prediction, Quechua, Peru

The system, called SA-IDS, was developed by Mahdi Ajdani and Maziar Asmani of the Department of Computer Engineering at Islamic Azad University, Qeshm branch, in Iran, and described in a study published in the journal Cluster Computing. Its central insight is that the traditional recipe for intrusion detection, training a supervised machine learning model on a large, carefully labeled dataset of attacks, simply does not work in real industrial settings. Labeled attack data is scarce, expensive, and quickly outdated, because adversaries constantly invent new techniques. Centralized retraining pipelines, in which data from thousands of devices is shipped back to a server for periodic model updates, add latency, bandwidth costs, and privacy risks that many industrial operators cannot accept.

SA-IDS takes a fundamentally different approach built on self-supervised learning, a technique that has transformed fields from computer vision to speech recognition. Instead of requiring humans to label data, the model learns by solving an artificial task on unlabeled data, in this case through contrastive learning. The idea is elegant: the system takes streams of benign telemetry from industrial sensors and controllers, creates altered versions of that data, and trains an embedding network to recognize which samples are original views of the same underlying behavior and which are different. By pulling representations of similar behavior together and pushing dissimilar ones apart, the network develops a rich internal map of what normal operation looks like, all without ever being told what an attack is.

Once that embedding space exists, anomaly detection becomes a geometric problem rather than a classification problem. SA-IDS measures the Mahalanobis distance between each new observation and the distribution of normal behavior learned during training. Unlike a simple Euclidean distance, the Mahalanobis metric accounts for the shape and correlations of the data, so it can distinguish a genuinely unusual reading from one that is merely rare but consistent with normal variation. If a new telemetry sample lands far from the cluster of benign behavior in the embedding space, the system flags it as a potential intrusion. This design means the detector can catch attacks it has never seen before, including zero-day exploits and novel variants, because it does not depend on memorizing signatures of known threats.

But detecting anomalies is only half the battle. Industrial environments are not static. Machines age, production lines are reconfigured, seasonal demand shifts operating schedules, and firmware updates change how devices communicate. In machine learning terms, the data distribution drifts over time, a phenomenon known as concept drift, and a model frozen at deployment gradually loses accuracy, either missing real attacks or drowning operators in false alarms. Conventional solutions involve periodic retraining with fresh labeled data, which is precisely the resource industrial edge deployments lack. SA-IDS addresses this with a lightweight, unsupervised adaptation mechanism that continuously updates the model using only high-confidence benign samples, allowing the learned representation of normal behavior to evolve as the environment does, without any human annotation or cloud round-trip.

The practicality of the approach was tested on a real-world IIoT testbed built from the kind of hardware actually deployed in industrial settings: Raspberry Pi single-board computers and ESP32 microcontrollers. These are the workhorses of low-cost industrial sensing, and they are severely constrained in memory, processing power, and energy budget. Running a modern deep learning system on such hardware is a genuine engineering challenge, and many state-of-the-art detectors simply cannot operate there. The researchers demonstrated that SA-IDS maintains real-time performance on these devices, meaning detection happens locally at the edge, where attacks can be caught and responded to before they spread, rather than after data has made a round trip to a distant server.

The numbers reported in the study are striking for a system operating under such constraints. SA-IDS achieved a detection accuracy of 96.3 percent, an F1-score of 0.92, and a false positive rate of just 3.1 percent, outperforming several state-of-the-art baseline methods. The false positive figure deserves particular attention in an industrial context. In a factory, every alarm triggers human investigation, and if a detector cries wolf too often, security teams begin ignoring it, a failure mode that can be as dangerous as missing an attack. A low false positive rate combined with high detection accuracy means the system can be trusted to run continuously without exhausting its operators.

The broader significance of the work lies in what it removes from the deployment equation. No labeled attack data. No centralized retraining. No dependence on cloud connectivity for security decisions. Each of these has been a persistent obstacle to deploying machine learning-based security in the industrial sector, where operational technology networks are often air-gapped or semi-isolated by design, and where downtime is measured in thousands of dollars per minute. A detector that learns from the data flowing past it and adapts on its own fits the operational reality of industrial environments far better than models that must be periodically fed curated datasets by security specialists.

The researchers have also made the system reproducible and extensible. The source code, trained models, and experimental scripts for SA-IDS are publicly available on GitHub, including the self-supervised training pipeline, the Mahalanobis distance implementation, the online adaptation module, and the testbed scripts needed to reproduce the experiments on Raspberry Pi and ESP32 hardware. This openness matters, because industrial cybersecurity is a field where independent verification and community scrutiny can accelerate adoption far faster than proprietary black boxes. Security tools that researchers and operators can inspect, test, and improve are inherently more trustworthy than those that cannot be examined.

Challenges remain, of course. Self-supervised anomaly detectors are ultimately bounded by the quality of their representation of normal behavior, and a sufficiently patient adversary who behaves in ways that mimic benign patterns may still evade detection. The adaptation mechanism, while designed to update only on high-confidence benign samples, must also be robust against poisoning attempts in which an attacker tries to redefine normality itself. Still, SA-IDS represents a meaningful step toward security systems that match the constraints of the devices they protect: small, self-sufficient, always learning, and always watching. As industrial networks continue to expand into critical infrastructure around the world, approaches like this one, which bring sophisticated machine learning to the very edge of the network without demanding resources those edges do not have, may well define the next generation of industrial cyber defense.

Subject of Research: A self-supervised, adaptive intrusion detection system for resource-constrained IIoT edge devices that detects anomalies via Mahalanobis distance and adapts to concept drift without labeled data.

Article Title: SA-IDS: a self-supervised and adaptive intrusion detection system for edge-based IIoT security with label-free drift resilience

Article References: Ajdani, M., & Asmani, M. (2026). SA-IDS: a self-supervised and adaptive intrusion detection system for edge-based IIoT security with label-free drift resilience. Cluster Computing, 29(13), Article 776. https://doi.org/10.1007/s10586-026-06562-3

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06562-3

Keywords: Industrial Internet of Things, intrusion detection system, self-supervised learning, contrastive learning, Mahalanobis distance, concept drift, anomaly detection, edge computing, adaptive security, Raspberry Pi, ESP32, cybersecurity

According to the researchers, existing voice-controlled home service robots typically cost between 150,000 and 250,000 rupees, placing them far beyond the reach of most households in developing economies. Those that are affordable tend to be narrow in capability, addressing a single subsystem such as autonomous movement or internet-of-things appliance control in isolation. And almost without exception, commercial platforms restrict voice interaction to English, excluding the hundreds of millions of users whose first languages are Hindi, Gujarati, or any of the other tongues spoken across linguistically diverse regions. Harley’s designers argue that no single commercial product currently combines mobile navigation, standard-protocol IoT control, and multilingual conversational AI at any comparable price point.

Technically, Harley rests on a deliberate division of labour between two compute tiers. A Raspberry Pi serves as the high-computing device, handling speech recognition, natural language understanding, navigation decision-making, and the coordination of subsystems. An ESP32 microcontroller sits at the other end of the architecture, tasked with driving the relays and circuits that switch high-power household appliances safely. The bridge between them is MQTT, the lightweight publish-subscribe messaging protocol that has become a de facto standard in industrial and consumer IoT deployments. By structuring communication around MQTT topics, the team decoupled the intelligence layer from the actuation layer, allowing the Raspberry Pi to issue commands to the ESP32 without direct electrical coupling and making the appliance-control subsystem extensible to any MQTT-compliant device.

The conversational layer is powered by the Google Gemini API, which the robot uses to interpret spoken commands and generate context-aware responses across multiple languages. Rather than relying on rigid, keyword-matched command grammars, Harley pipes captured speech to the large language model, which can handle paraphrase, mixed-language utterances, and open-ended questions in addition to discrete instructions such as switching off a fan or navigating to the kitchen. The researchers report that this integration enabled natural, context-aware communication, and that in experimental trials the system earned an average conversational quality rating of 4.1 out of 5.0 across multiple languages — a figure that, while drawn from their own residential evaluation, suggests the interaction model was judged substantially better than the brittle command-and-response behaviour typical of low-cost voice devices.

Navigation is the third pillar of the design. Harley supports voice-enabled navigation commands that direct the robot to predefined locations inside the home, using stored coordinates recorded within the environment. In practice, a user can ask the robot to travel to a named waypoint — a room, a charging dock, or a designated position — and the robot plans and executes the movement autonomously. The waypoint approach trades the complexity of full simultaneous localisation and mapping for a simpler, more robust scheme suited to static home layouts, and it keeps the computational burden low enough to run on modest hardware. The team notes that experimental validation in a residential setting measuring 75 square metres across four rooms demonstrated a high voice recognition accuracy and an improved navigation success rate alongside reliable appliance control with low latency.

The economics of the build are arguably its most striking feature. The researchers itemised the total component cost of Harley at 9,247 rupees, which they benchmarked against a mid-range alternative configuration assembled from commercial off-the-shelf parts at 47,850 rupees — an 80 percent component-level saving. Compared with the 150,000 to 250,000 rupee range quoted for existing voice-controlled home service robots, Harley comes in at roughly 19 percent of the cost of an equivalent system, and a small fraction of the premium end of the market. The authors frame this not merely as an engineering achievement but as a question of access: high-performance home automation robotics, they argue, is achievable at a fraction of current market cost, with particular relevance to multilingual and cost-sensitive environments, including those serving elderly users in developing economies.

That last point carries social weight. India’s ageing population, like much of Asia’s, is growing faster than the infrastructure available to support it, and voice-first interfaces are among the most accessible technologies for users with limited mobility, declining eyesight, or low digital literacy. A robot that understands commands in a user’s native language and can physically move through the home — fetching itself to a bedside, switching appliances, answering questions — addresses a genuinely different user group than the wall-mounted smart speakers and app-controlled plug ecosystems that dominate the consumer market. The researchers position Harley within a broader body of work on service robots in aged care, noting that prior deployments have shown promise but frequently stumble on cost, language support, and the integration gap between mobility and control functions.

The study is candid about its experimental scope. Validation was conducted in a single residential setting rather than across a fleet of homes, and the waypoint-based navigation scheme presumes a relatively stable floor plan, meaning a significant rearrangement of furniture would require re-recording stored coordinates. The conversational quality scores, while encouraging, originate from the development team’s own evaluation, and larger, independent user studies with elderly and non-technical participants would be needed to establish how the system performs under real-world acoustic conditions, accents, and background noise. The cloud dependence of the Gemini API also raises questions about latency, privacy, and behaviour during internet outages — a consideration the authors’ own cited literature on offline voice interaction for elderly-care robots highlights as an open challenge in the field.

Even so, the work signals a shift in what low-cost robotics can credibly deliver. By pairing commodity single-board computers with lightweight IoT protocols and cloud-scale language models, small academic teams can now assemble systems that until recently demanded enterprise budgets. If the cost trajectory holds, the practical consequences could reach well beyond gadget enthusiasts: assistive robots for elderly users who speak minority languages, affordable automation for small clinics and care homes, and educational platforms that let students experiment with integrated robotics without institutional funding. The Nirma University team, which received no specific grant from any funding agency and developed the system with laboratory facilities and laser-cutting equipment at the institute, has demonstrated that the barrier to a genuinely useful, multilingual, mobile home robot is no longer primarily technical or financial — it is a matter of integration, and Harley shows one concrete way to close it.

Subject of Research: A low-cost multilingual voice-interactive home service robot integrating navigation, IoT control, and conversational AI

Article Title: Harley: a voice interactive intelligent robot for home services and automation

Article References: Chokshi, T., Shah, D., & Pandya, V. (2026). Harley: a voice interactive intelligent robot for home services and automation. International Journal of Intelligent Robotics and Applications. https://doi.org/10.1007/s41315-026-00587-y

Image Credits: AI Generated

DOI: 10.1007/s41315-026-00587-y

Keywords: robotics, home automation, voice interaction, MQTT, Raspberry Pi, ESP32, conversational AI, Google Gemini, indoor navigation, IoT, multilingual technology, assistive robots