The core of their argument is a concept they call data endogeneity. In a classical machine learning pipeline, data is collected, cleaned, labeled, and then fed to a model; the model’s predictions do not change how the data was generated. In a learning-enabled 6G network, the opposite happens. The observations a learning system sees—channel measurements, traffic statistics, interference patterns, latency readings—are produced by the very communication, control, and architectural decisions the network itself makes. When a scheduler reallocates spectrum, the traffic statistics shift. When a beam-training algorithm changes its probing strategy, the channel measurements it later receives are different from what they would have been otherwise. The data is not a static resource sitting in a warehouse; it is a native byproduct of network dynamics, continuously regenerated by the system’s own behavior.

That feedback loop may sound like a philosophical nicety, but the survey shows it has hard technical consequences. Standard machine learning theory leans on assumptions such as independent and identically distributed samples, stationarity of the underlying process, and a clean separation between the training phase and the deployment phase. Wireless networks violate all three. The wireless channel is famously non-stationary: users move, obstacles appear, interference sources come and go, and the statistical properties of the signal environment drift on timescales from milliseconds to hours. Interference, in particular, exhibits temporal correlation under common fading models, meaning that successive observations are not independent draws from a fixed distribution but tightly coupled samples of an evolving process. A model trained on yesterday’s channel statistics may be stale by lunchtime.

Partial observability compounds the problem. No network element sees the whole system. A base station observes its own cells in detail and neighboring cells only dimly; a user device sees its local link but nothing of the broader topology. Learning algorithms deployed in such conditions must make decisions from incomplete, biased, and correlated views of the state, and the survey emphasizes that the placement of learning within the network architecture—what the authors call edge–cloud data locality—becomes a first-order design question rather than an implementation detail. Where data is aggregated, and where models are trained and updated, determines the statistical coherence of the learning problem and the timescales over which the network can actually adapt. A model refreshed every second at the network edge faces a fundamentally different learning problem from one retrained nightly in a central cloud, even if the underlying algorithm is identical.

Latency and resource constraints add a further layer of realism that much of the algorithmic literature ignores. 6G envisions services such as ultra-reliable low-latency communications, where decisions must be made within millisecond budgets and where failures carry safety implications for applications like industrial automation and vehicular control. Learning systems cannot simply pause the network while they retrain. Inference must run on hardware with finite compute and energy, often on edge devices, and the communication cost of coordinating distributed learning—federated updates, model synchronization, data shipment—competes directly with the network’s primary job of moving user traffic. The survey synthesizes work on edge intelligence, federated learning with non-IID data, and communication-efficient distributed training to map how these constraints reshape what is learnable in practice.

One of the most striking threads in the analysis draws on the emerging theory of performative prediction. In conventional supervised learning, the model is a passive observer of a fixed world. In performative settings, the model’s own predictions change the distribution it is predicting—a phenomenon first formalized in the machine learning literature and now, the authors argue, unavoidable in closed communication and control loops. A scheduling policy trained on observed traffic will alter that traffic once deployed; a reinforcement learning agent controlling radio resources changes the interference environment that its next observations will reflect. The survey reviews performative reinforcement learning in gradually shifting environments and state-dependent performative prediction, arguing that 6G networks are a natural habitat for these effects and that ignoring them leads to systematic, reproducible failures rather than random noise.

The closed-loop nature of learning-enabled networking also raises the stakes on stability and safety. When a learning controller sits inside a feedback loop with the physical channel and the network’s control plane, the question is no longer just whether the model’s accuracy is high but whether the coupled system remains stable. The authors survey work on learning-based model predictive control, predictive safety filters, control barrier functions, and safe reinforcement learning under partial observability—techniques developed largely in robotics and control theory that they argue must migrate into network design. Open RAN platforms, which expose programmable closed-loop control interfaces for machine-learning applications, are highlighted as both an opportunity and a testbed: they make it possible to study these dynamics experimentally, on real radio hardware, rather than only in simulation.

Perhaps the most uncomfortable part of the survey is its critique of evaluation practice. The authors examine benchmarking protocols commonly used in the field and identify systematic mismatches between those protocols and operational 6G reality. Static datasets stand in for live networks; independent test sets stand in for temporally correlated streams; offline metrics stand in for closed-loop performance. Time series forecasting research has long shown that naive performance estimation can badly mislead when data is non-stationary, and the survey argues that networking research inherits the same trap. The result is a recurring pattern of failure modes induced by abstractions: models that look excellent in the lab and degrade unpredictably in deployment, adaptation mechanisms that oscillate or diverge when coupled with the network, and benchmarks that reward exactly the assumptions real systems violate. The authors also invoke the broader machine learning literature on evaluation gaps and on data cascades in high-stakes AI, where the unglamorous work of data quality is skipped in favor of model work, with predictable consequences.

What the survey deliberately does not do is catalog algorithms. There are already numerous surveys enumerating machine learning techniques for resource allocation, network slicing, beam management, and physical-layer design. Trigka and Dritsas instead ask a prior question: under what system-level conditions does data-driven learning remain valid, stable, and interpretable inside an operational network? Their answer reframes Big Data and machine learning as endogenous system functions of 6G—components woven into the network’s own operation, subject to its physics, its latencies, and its feedback loops—rather than add-on intelligence layered on top. That reframing carries practical implications for how the field should evaluate claims, design experiments, and architect the AI-native networks that standards bodies and vendors are already building.

The timing is not incidental. Research programs on AI-native and task-oriented 6G architectures, foundation-model-based cloud–edge–end collaboration, and integrated sensing and communication are moving from vision papers to engineering specifications, and the decisions made in the next few years will harden into infrastructure that lasts decades. If the survey’s central claim is right—that the classical machine learning playbook quietly breaks when the learner and the learned system are the same machine—then the most important innovations in 6G intelligence may not be new architectures or exotic algorithms at all, but a more honest account of what happens when a network learns from itself. The authors, who contributed equally to the work, frame their contribution as a foundation for understanding, evaluating, and designing learning-enabled 6G systems beyond the algorithm-centric paradigms that have dominated the conversation so far. Whether the field listens may determine how much of the 6G hype survives contact with reality.

Subject of Research: System-level integration of Big Data and machine learning in 6G communication networks

Article Title: Big Data–driven machine learning for 6G communications: from algorithmic promises to system-level realities

Article References: Trigka, M., & Dritsas, E. (2026). Big Data–driven machine learning for 6G communications: from algorithmic promises to system-level realities. International Journal of Data Science and Analytics, 22(1), Article 292. https://doi.org/10.1007/s41060-026-01274-8

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01274-8

Keywords: 6G, machine learning, Big Data, wireless networks, edge intelligence, federated learning, non-stationarity, closed-loop control, performative prediction, O-RAN, network slicing, evaluation benchmarks

The core insight behind CLATS is that alert optimization is not purely a statistical problem; it is an operational one. Most deployed deterioration-warning systems rely on static or ad hoc thresholds set when the software is installed. But hospitals are not static environments. Patient case-mix shifts between winter and summer, staffing levels rise and fall across shifts, and the workload a unit can absorb on a Tuesday afternoon differs from what it can absorb at three in the morning. A threshold that keeps missed events acceptably rare during calm periods may flood the same unit with alerts during a surge, or vice versa. The researchers argue that any sensible alert policy must therefore balance two competing quantities that most systems treat separately: the sensitivity of the detector, meaning its ability to catch true deterioration events, and the workload burden it imposes on the clinical staff who must respond.

CLATS addresses this trade-off with a deliberately transparent architecture that the authors describe as auditable and deterministic. The system pairs a calibrated predictive model with a two-threshold, three-tier triage policy. Patients whose deterioration risk score exceeds the upper threshold L2 land in the highest-alert tier and trigger immediate attention. Patients scoring between the lower threshold L1 and L2 fall into a middle tier, and those below L1 require no action. This tiered structure matters because it lets the optimizer treat false positives with nuance: an alert that merely asks a nurse to glance at a chart costs far less than one that demands an emergency response team. Rather than counting every false alarm equally, CLATS assigns tiered false-positive costs, alongside explicit penalties for missed events and for system overload situations in which the volume of alerts exceeds what the ward can realistically process.

The machinery that makes this work is a penalized, operations-aware cost function. In essence, the optimizer searches for the pair of thresholds that minimizes a weighted sum of harm: the penalty for each missed deterioration event, the accumulated costs of false alarms in each tier, and the burden of overload. Crucially, the search operates under a hard constraint, a fixed sensitivity target that the system must never fall below, no matter how attractive the cost savings. This design decision reflects a clinical reality that pure cost-benefit mathematics can obscure: in medicine, missing a true deterioration event is generally far more dangerous than responding to a few extra false ones, so the safety floor is enforced structurally rather than traded away when convenient.

Technically, the optimization proceeds in two phases. Initial thresholds are determined on validation data using isotonic calibration, a non-parametric technique that maps raw model scores into well-calibrated probabilities by fitting a monotonic, non-decreasing function. Isotonic calibration is prized in clinical prediction because it makes no assumptions about the shape of the relationship between scores and outcomes and produces probabilities that can be trusted as genuine risk estimates. On top of those calibrated scores, a discrete grid search enumerates candidate threshold pairs and identifies the combination that minimizes the operations-aware cost. Grid search is a blunt but dependable instrument: because the space of threshold pairs is finite and enumerable, every decision the system makes can be traced, reproduced, and explained to a hospital governance committee, a property many black-box alternatives lack.

What distinguishes CLATS from a one-time tuning exercise is its closed loop. Once deployed, the system does not let its thresholds fossilize. Instead, it operates in review cycles, re-estimating the optimal thresholds as new data accumulates and as staffing and patient-population conditions shift. The refinement process uses exponential smoothing, which blends the newly computed thresholds with the previous values while weighting recent observations more heavily, and guardrail constraints, which bound how far thresholds may move in a single cycle. The guardrails are a safeguard against instability: a sudden data anomaly or transient shift in the ward could otherwise cause a wild threshold swing that disrupts clinical workflows. With smoothing and bounded updates, the system adapts gradually and predictably, in the spirit of stochastic learning and adaptive control rather than abrupt reconfiguration.

The evaluation reported in the paper is synthetic, and the authors are unusually explicit about what that means for interpretation. Working with simulated data and an isotonic-calibrated CatBoost model, a gradient-boosted decision tree method known for strong performance on tabular clinical data, CLATS maintained a sensitivity of at least 0.989 across the tested conditions. Under the controlled data-generating assumptions of the simulation, the system produced no false positive alerts, a result that must be read carefully: it validates the optimization mechanism itself, not clinical performance, because real hospital data contain noise, label ambiguity, and distribution shift that simulations rarely capture. The authors state plainly that these findings support the feasibility of the CLATS mechanism rather than constituting clinical evidence, a caveat that responsible readers should carry forward.

The more operationally interesting result concerns the closed-loop dynamics. During simulated re-optimization cycles, the adaptive threshold updates reduced the operational cost objective by an average of 4.6 percent per review cycle, without reducing recall in this setting. Compounded across repeated cycles, savings of that magnitude suggest that static thresholds leave meaningful efficiency on the table, particularly in environments where patient acuity and staffing fluctuate. Because the cost function explicitly encodes workload and overload penalties, the improvements translate conceptually into fewer disruptive alarms and less wasted clinical attention, precisely the resources that alarm fatigue squanders. Prior literature has linked clinician workload to patient safety outcomes, and operations-management research has long documented how overburdened staff experience degraded performance, giving the workload-aware formulation an empirical footing beyond simple intuition.

Several design choices position CLATS for eventual real-world adoption, even though that step awaits validation on genuine longitudinal clinical data. The framework is model-agnostic: the CatBoost predictor in the paper could be swapped for any sufficiently accurate deterioration model, since the optimization layer interacts only with calibrated risk scores. The system is also governance-compatible, offering a transparent, deterministic basis for decisions that hospital safety committees and regulators can inspect line by line. That transparency addresses a persistent obstacle in clinical artificial intelligence deployment. Studies of real-world sepsis prediction systems have shown that technical accuracy alone does not guarantee clinical uptake; integration into routine workflows, trust, and auditability often decide whether a predictive tool helps or gathers dust. A framework whose every threshold decision can be explained and reproduced speaks directly to those implementation barriers.

The research lands at a moment when healthcare institutions are actively searching for ways to tame alert burden, from remote patient monitoring programs to decision-support systems in intensive care, and the bibliography of the paper reflects that urgency, spanning early warning score evaluations, alert reduction frameworks, and reinforcement learning approaches to healthcare operations. CLATS contributes a specific and disciplined piece to that puzzle: not a new predictor, but a principled mechanism for turning any calibrated predictor into an operationally sustainable alert policy. The caveats remain clear. The results are synthetic, the data-generating assumptions are idealized, and the authors themselves, along with their funders at the Gust Engineering and Applied Innovation Research Center, frame the work as a foundation for future validation. If that future validation on real hospital data holds, closed-loop, operations-aware threshold optimization could become a standard layer in the clinical monitoring stack, quietly retuning the alarm systems that clinicians have learned, at their peril, to tune out.

Subject of Research: A closed-loop adaptive optimization framework for clinical deterioration alert thresholds

Article Title: CLATS: a closed-loop adaptive triage system for operations-aware clinical alert optimization

Article References: Alzhouri, F., Daraghmeh, M., Agarwal, A., & Ebrahimi, D. (2026). CLATS: a closed-loop adaptive triage system for operations-aware clinical alert optimization. Cluster Computing, 29(14), Article 793. https://doi.org/10.1007/s10586-026-06611-x

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06611-x

Keywords: clinical alerts, alert fatigue, adaptive triage, machine learning, CatBoost, isotonic calibration, threshold optimization, hospital operations, sensitivity, closed-loop control, clinical decision support, predictive modeling

The authors, Yongqi Zhang, Eric M. Yeatman and Ranan Dasgupta, frame the problem in terms of the two fundamentally different failure modes of the lower urinary tract. In detrusor underactivity, the bladder cannot contract strongly enough to void, so mechanical assistance is needed during emptying. In detrusor overactivity, by contrast, the bladder contracts aberrantly during the storage phase, producing urgency and incontinence, so the therapeutic goal is inhibition rather than assistance. A single soft-robotic construct, they argue, could in principle address both phenotypes — but only if its control logic is matched to the underlying physiology, and only if a series of formidable biomechanical and regulatory challenges can be overcome.

The core of the proposal is a soft actuator sleeve that conforms to the exterior of the bladder. Unlike traditional rigid implants, which create stress concentrations and can damage delicate tissue, soft actuators made from elastomers, pneumatics or magnetic materials distribute forces gently across the organ wall. The conceptual architecture involves sensing bladder volume and pressure in real time, deciding when assistance is appropriate, and then applying controlled compression to raise intravesical pressure and drive urine through the urethra. The authors describe an idealized pressure–flow relationship for voiding assistance, drawing on Laplace-law insights from ultrasound urodynamics: because wall tension depends on both pressure and radius, a compliant sleeve can amplify the effectiveness of modest actuation forces as the bladder empties and shrinks.

Crucially, the authors insist that voiding assistance must be strictly patient initiated — a human-in-the-loop control philosophy. Micturition is not merely a mechanical reflex; it is gated by supraspinal brain circuits that integrate social context, and functional brain imaging has shown that urgency and continence involve forebrain influences on the pontine micturition switch. An implant that squeezed the bladder autonomously whenever it detected fullness would override this behavioural gating and could cause socially catastrophic emptying. Preserving social continence therefore requires that the machine act only when the patient commands it, with the algorithm serving as an amplifier of intent rather than a replacement for it. Emerging brain–computer interface work decoding urination motor attempts in spinal cord injury patients suggests that even severely injured patients may retain the neural signals needed to trigger such systems.

The opposite phenotype demands the mirror-image strategy. For detrusor overactivity, the authors propose autonomous, closed-loop neuromodulation that detects and inhibits aberrant bladder micromotions during storage without any conscious patient intervention. Unregulated autonomous micromotions of the bladder wall have been implicated in both overactive bladder and detrusor underactivity, and animal studies have demonstrated that closed-loop stimulation triggered by the frequency spectrum of non-voiding bladder activity can suppress unwanted contractions. Recent advances in precise tibial nerve stimulation, guided by evoked compound action potential feedback, point toward implantable systems that could continuously monitor bladder electrical or mechanical signals and deliver inhibitory neuromodulation the moment pathological activity begins — a genuinely artificial continence reflex.

Three families of actuators are emerging as candidates for the mechanical side of the problem, each at a different level of technology readiness. Pneumatic artificial muscles, including PneuNet-type bending actuators, offer high forces and simple fabrication but require pneumatic lines or pumps that complicate implantation. Dielectric elastomer actuators, which squeeze a soft elastomer film between compliant electrodes at high voltage, deliver large strains and fast response, and recent multilayer designs have achieved impressive performance — yet they face dielectric breakdown risks and the challenge of generating kilovolt-level fields safely inside the body. Magnetic soft actuators, in which embedded magnetic particles allow an implant to be deformed by external fields, have already been used to build a magnetically controlled robotic bladder that enhanced urine flow in experimental work, and they eliminate the need for on-board power electronics at the cost of requiring an external field source.

Whatever the actuator technology, the authors identify a set of biomechanical constraints that any clinical system must solve. During the filling phase, a snugly fitted sleeve can create a suction effect that resists bladder expansion, so the design must incorporate fail-safe open mechanical architectures that relax passively as the organ fills. Long-term implantation inevitably provokes fibrotic encapsulation, the foreign-body response that thickens tissue interfaces and degrades both sensing fidelity and mechanical coupling. Power delivery is equally thorny: implantable batteries add volume and eventually require replacement, driving interest in wireless approaches ranging from ultra-low-frequency magnetic energy focusing to ultrasound and magnetoelectric transduction. Encapsulation films built on atomic-layer-deposited nanolaminates must keep body fluids out for years, and any magnetic components must satisfy MRI safety standards such as ISO/TS 10974 and the relevant ASTM test methods for heating, torque and displacement.

Sensing and computation, meanwhile, are advancing rapidly on the soft-electronics front. Fully implantable, sensorized artificial bladders have been demonstrated that monitor volume and fullness continuously, and wireless bioelectronic harnesses with soft strain sensors can track bladder function through surgical recovery. Stretchable sensors based on liquid metals, graphene, conductive hydrogels and high-linearity capacitive designs provide the raw signals, while in-sensor and near-sensor computing — the emerging discipline of tiny machine learning — allows classification of bladder states on milliwatt-scale edge processors rather than in the cloud. The authors point to benchmark suites such as MLPerf Tiny as evidence that the computational hardware needed for on-board, adaptive control is arriving just as the actuator hardware matures.

The final hurdles are ethical and regulatory rather than purely technical. Algorithmic continence control raises questions about autonomy, consent and failure modes: what happens when an adaptive machine learning system drifts, or is compromised? The authors note that regulatory pathways for adaptive artificial intelligence in bioelectronics are still being created — the US Food and Drug Administration has only recently finalized guidance on predetermined change control plans for AI-enabled device software and on cybersecurity in medical devices — and that a definitive roadmap must outline how continuously learning implants will be validated, updated and monitored over a lifetime of use. Sterilization standards, biocompatibility evaluation under ISO 10993 and radio-spectrum rules for medical implants further shape the engineering envelope.

None of these obstacles, the authors conclude, is fatal; each is the kind of problem that interdisciplinary collaboration between engineers, urologists and neuroscientists has solved before in adjacent fields, most visibly in soft robotic cardiac sleeves that restored pumping function in experimental hearts. If the field can integrate fail-safe mechanics, robust sensing, patient-centred control and trustworthy adaptive algorithms, soft-robotic bladder orthoses could transform the management of lower urinary tract dysfunction — replacing the catheter bag with an invisible, compliant machine that restores not just voiding, but the quiet, unremarkable social confidence that continence makes possible. For a condition that has seen so little therapeutic progress over the past few decades, that would be nothing short of revolutionary.

Subject of Research: Soft robotic bladder implants for restoring urinary voiding and continence control

Article Title: The potential of soft robotics for the restoration of urinary voiding

Article References: Zhang, Y., Yeatman, E. M., & Dasgupta, R. (2026). The potential of soft robotics for the restoration of urinary voiding. Nature Reviews Urology. https://doi.org/10.1038/s41585-026-01186-z

Image Credits: AI Generated

DOI: 10.1038/s41585-026-01186-z

Keywords: soft robotics, bladder, urinary retention, detrusor underactivity, detrusor overactivity, neuromodulation, bioelectronics, implantable devices, continence, dielectric elastomer actuators, closed-loop control, urology

The key to the new platform is the integration of three functions on a scale measured in tens of micrometres. Each robot carries a temperature sensor, a programmable CMOS control circuit and arrays of electrochemical actuators that beat like biological cilia. The actuators are built from a titanium–palladium stack that bends when voltage is applied, driving ions into and out of the palladium layer and causing controlled bending at engineered hinges. Because the cilia are hinged, with two rigid panels connected by rotational joints, their beat cycle can be programmed to break the time-reversal symmetry that governs fluid motion at low Reynolds number. At microscopic scales, where viscosity dominates over inertia, fluid flows are reversible unless the stroke and the recovery stroke differ in shape, a constraint famously articulated in Lighthill’s analysis of flagellar hydrodynamics. The hinged design allows the robots to sweep the fluid with a fast, extended stroke and a slow, folded recovery, producing net pumping in a chosen direction.

Powering and controlling such tiny machines is a formidable engineering challenge, and the team solved it with microscale photovoltaic regions that convert incident laser light at 635 nanometres into electrical current for both the logic and the actuators. The binary sensing circuit switches its output phase configuration at threshold temperatures of 26 and 30 degrees Celsius, and can deliver actuation frequencies ranging from 0.4 to 12.8 hertz, with 1.6 hertz used for cilium beating in the reported demonstrations. The photovoltaic supply generates currents of roughly 0.78 microamperes at one sun of illumination, rising to about 4.8 microamperes at ten suns, enough to run the circuit and drive the cilia without any tether. This architecture builds on earlier work from the same collaboration, including electronically integrated mass-manufactured microscopic robots, cilia metasurfaces for programmable microfluidic manipulation, and microscopic robots with onboard digital control, but it adds something those systems lacked: a genuine sensory loop in which the robot’s action depends on what it measures.

The researchers demonstrated three distinct temperature-responsive modalities, each producing a different coupling between robot behaviour and environmental cues. The first is binary sensing, in which the onboard sensor acts as a threshold detector. Below the threshold the cilia pump in one direction; above it, the circuit reconfigures the wiring to the actuators and the pumping reverses. The team showed that this simple switch can reverse unidirectional flow, reverse the rotation of a single vortex, or reverse both vortices in a symmetric counter-rotating pair, all in response to nothing more than the ambient temperature crossing a set point. Because the flow patterns are generated by arrays of individually wired cilia, the same sensing principle can be routed into dramatically different hydrodynamic outcomes simply by changing how the circuit output is distributed across the array.

The second modality replaces the sharp threshold with continuous sensing, implemented using pulse-coupled oscillator circuits built from dynamic-leakage-suppression logic gates. In this scheme, the oscillation frequency of the onboard circuit depends on temperature through the exponential dependence of subthreshold leakage currents, so the beat frequency of the cilia rises smoothly as the fluid warms. The pumping speed therefore tracks temperature continuously rather than switching between two states. The oscillator architecture also supports synchronisation: coupling pulses from a designated leader oscillator advance the follower until the two lock with a fixed phase offset, a mechanism related to the pulse-coupled designs previously used for coordinating autonomous microscopic machines through local electronic pulses. This phase-locking provides a route to coordinated actuation across a cilia array without any central controller.

The third modality exploits spatial temperature gradients rather than absolute temperature. Using a scalable ultra-low-power temperature gradient sensor based on pulse-coupled oscillators, the robot determines which side of its body is warmer and aligns its pumping accordingly. When the imposed gradient is reversed, the leader and follower roles in the oscillator network exchange, and the pumping direction flips. The result is a microscopic machine that orients its fluid-mechanical output along the local thermal landscape, in loose analogy to the way ciliated protists orient their swimming relative to environmental cues. Together, the three modalities, threshold switching, continuous modulation and gradient alignment, form a toolkit for programming how a microrobot’s behaviour responds to its surroundings.

The most striking demonstration is the closing of the loop. Because the cilia-driven flows move fluid around, they transport heat, and the team showed that a collective array of cilia can actively reshape the local thermal field. In their experiments, unidirectional pumping advected fluid from a cooler region towards a hotter region, flattening the temperature gradient that the sensors were measuring. As the temperature field changed, the oscillator frequencies shifted, which in turn altered the pumping, which further modified the thermal field. The researchers modelled this feedback with a reduced two-dimensional description of thermally advected flow and quantified it through a thermal stretching length that grows cycle by cycle, fitting the dynamics with an exponential law. The system thus exhibits a genuine closed-loop interaction among sensing, actuation and environment, the microscopic analogue of an organism modifying its own habitat.

The experimental platform itself is a tour de force of integration. The devices were fabricated at the Cornell NanoScale Facility, with the CMOS circuits exposed by etching the top dielectric, then interconnected with titanium–platinum leads, re-encapsulated in silicon dioxide, shielded with a grounded layer, and finally released with an aluminium nitride sacrificial layer before the titanium–palladium actuator stack and rigid panels were added. Fluid temperatures were controlled with a hotplate and characterised by infrared thermography on dry samples, which showed a uniform region with a spatial standard deviation of about 0.11 degrees Celsius and a gradient region of roughly 0.5 degrees Celsius per millimetre; in liquid, where phosphate-buffered saline strongly absorbs mid-infrared radiation, the team used a micro-thermocouple instead. Flow fields were measured with particle image velocimetry and compared against three-dimensional hydrodynamic simulations of the beating cilia, validating the theoretical model of the two-hinged kinematics.

The implications reach well beyond the laboratory demonstration. Artificial cilia that sense and respond to their environment could regulate temperature and chemical gradients in lab-on-chip systems, create fluid microhabitats for recruiting cells or controlling microbiomes in biomedical contexts, and serve as building blocks for emergent collective behaviours in swarms of autonomous micromachines. The theoretical framing, developed with physicists studying non-reciprocal phase transitions and adaptive active solids, suggests that populations of such robots could self-organise into patterns no single robot is programmed to produce, much as bacterial colonies generate large-scale spiral waves from local interactions. Because the platform is built with standard CMOS processes and mass-manufacturable techniques, scaling to larger and more capable microrobotic collectives appears feasible. The work was supported primarily by the National Science Foundation and the Army Research Office, with additional support from the Kavli Institute at Cornell and Westlake University, and the team has filed patent applications covering the actuators and control electronics.

What makes this advance conceptually important is the shift from open-loop microrobots, which execute fixed motions, to machines that participate in a dynamic dialogue with their world. In nature, the coupling between organism behaviour and environmental modification underlies nest construction, swarm thermoregulation and microbiome recruitment; until now, engineered microrobots could sense or act, but rarely both in a feedback loop. By embedding thermal sensing, programmable logic and ciliary actuation on a single chiplet smaller than a grain of salt, the Cornell-led team has created the first microscopic robotic platform in which the environment is not merely a medium the robot moves through but a variable the robot actively regulates. As the researchers and their collaborators refine the sensing modalities and extend them to chemical and mechanical cues, the boundary between living, environment-shaping matter and engineered machines at the microscale looks set to blur further.

Subject of Research: Microscopic robots with onboard sensing, logic and cilia actuators that couple their behaviour to thermal environmental cues in a closed loop

Article Title: Microscopic robots that sense and reshape their environment

Article References: Wang, W., Zhang, J., Chaudhari, P., Shim, K., Severn, J., Zheng, C., Liang, Z., Ji, Y., Pelster, J., Griniasty, I., Seara, D., Vitelli, V., Lauga, E., Apsel, A., & Cohen, I. (2026). Microscopic robots that sense and reshape their environment. Nature Electronics. https://doi.org/10.1038/s41928-026-01709-x

Image Credits: AI Generated

DOI: 10.1038/s41928-026-01709-x

Keywords: microrobotics, artificial cilia, CMOS, electrochemical actuators, temperature sensing, microfluidics, closed-loop control, NEMS, collective behaviour, thermal regulation, Cornell University, Nature Electronics

Arc-DED is one of the most economical additive manufacturing routes for large metallic structures. Instead of a laser, it uses an electric arc to melt wire feedstock, achieving high deposition rates at low cost. Gas tungsten arc-DED (GTA-DED), the variant studied here, employs a non-consumable tungsten electrode and produces a smoother transition of molten material, making it attractive for high-quality fabrication in alloys ranging from aluminum to nickel and titanium. Yet the process is notoriously sensitive to disturbance. Heat accumulates layer upon layer, substrate conditions vary, and the surface state of each previously deposited bead changes the geometry the torch encounters next. When the deposited width drifts from the design value, defects such as porosity, poor fusion, and bead collapse appear between adjacent beads. When the actual height drifts from the programmed lifting height of the torch, the working distance between electrode and part degrades—too short and the tungsten electrode can collide with the surface, too long and shielding gas fails to protect the pool or the arc simply extinguishes, halting production.

Previous strategies to tame these deviations—offline heat-input models, path compensation, interlayer active cooling—share a common weakness: they do not respond to the dynamic geometry of the molten pool as it forms. Earlier sensing approaches each captured only half the picture. Infrared cameras can map pool isotherms but are expensive at the required frame rates and resolutions. Passive vision, using a CCD camera and image processing, reliably measures deposition width but struggles to track height without a side-mounted camera that collides with complex parts. Electrical sensing of arc voltage, by contrast, cheaply and almost instantaneously reflects the distance from torch to deposition surface—effectively the height—but says nothing about width. The innovation of the new study is to fuse the two: arc voltage for height, vision for width, each sampled at the moment it performs best.

Raw arc voltage signals, however, are noisy. Fluctuations in the conductive channel of the arc plasma corrupt the measurement, and classical threshold-based wavelet filters, derived under idealized Gaussian white-noise assumptions, either over-filter and erase faint signal features or leave too much residual noise. The team’s answer was to let an ant colony optimization algorithm search for the optimal wavelet threshold. Inspired by the pheromone trails ants lay to find shortest paths, the algorithm iteratively updates pheromone concentrations across candidate thresholds, guided by the mean squared error of the filtered signal. Because the search is global rather than fixed by a closed-form expression, the filter adapts to the actual character of each signal, preserving transient peaks and abrupt changes that carry real physical information about deposition height stability.

Calibration posed a subtler challenge. In a thick-walled, multi-layer, multi-bead part, the arc does not look the same everywhere. On the first layer, heat dissipates efficiently into the substrate and no closed-loop control is needed. On the second layer, the arc straddles the first layer’s surface and the substrate. By the third layer, the arc touches the previous layer’s surface and side wall, and from the second bead onward it spans two deposited layers simultaneously. Each morphology produces distinct electrical signatures, so a single arc-length model would be hopelessly inaccurate. The researchers therefore constructed three separate plane-fitting models linking peak arc voltage, arc length, and peak current—one for the first bead of the second layer, one for the first bead of the third layer, and one for subsequent beads—and deployed the appropriate model as deposition progressed.

Vision sensing required equally careful choreography. The system uses pulsed current, and at peak current the arc blazes so brightly that it saturates the camera’s dynamic range, drowning the molten pool in glare and reflections from neighboring beads. At base current, arc-light interference drops dramatically and image quality improves. The team therefore timed every image capture to the eightieth millisecond after the current transitioned from base to peak, using a camera fitted with a 25 mm lens, a 2 percent neutral density filter, and a narrow-band 685 nm filter. Image processing proceeded through Gaussian filtering to suppress noise, a Laplacian operator to detect pool edges, and a Hough transform to extract them, with two small 50-by-380-pixel windows positioned at the tail of the molten pool—far enough from the arc glare at the pool head, yet not so far back that solidification introduced lag. A chessboard calibration, cross-referenced against the arc-voltage-derived arc length, converted pixel coordinates into millimeters, yielding a width monitoring error below 0.04 mm.

With sensing in place, the control architecture split the problem in two. A model reference adaptive controller (MRAC) regulated wire feeding speed to stabilize deposition height: a PID controller computed wire-speed corrections from the deviation between detected and reference arc voltage, while an adaptation mechanism continuously retuned the PID gains based on sensitivity functions, allowing the controller to track a time-varying nonlinear process. In parallel, a self-tuning fuzzy controller (FSTC) adjusted peak current to hold the pool width at target. Rather than fixing the fuzzy controller’s quantization and scale factors in advance, the FSTC recalculated them on the fly from the width error and its rate of change, following rules distilled from expert experience and requiring no precise quantitative model of the process. Simulation of the width controller showed that when the target pool width jumped from 5 to 5.5 mm, the system tracked the change within five seconds without significant overshoot.

The experimental contrast was striking. With constant process parameters, a ten-layer, ten-bead thick-walled part of ER70S-6 steel wire on Q235B substrate grew increasingly erratic: peak arc voltage climbed layer by layer as height deviations accumulated, marginal beads deposited taller than interior beads, and the pool width drifted from 4.5 to about 5.3 mm as heat dissipation conditions evolved. By the eleventh layer, the tungsten-to-substrate distance exceeded the maximum at which an arc could be initiated, and the process simply stopped. Dents pocked the termination region, fusion lines between beads were irregular, and the finished part showed misalignment and bulging, with a height deviation reaching 5 mm across the part.

Under closed-loop control, the same system printed a fifty-layer, ten-bead part continuously and without manual intervention. The MRAC held arc voltage deviations to a maximum of roughly 0.27 to 0.31 volts during stable burning, while the FSTC pinned the molten pool width near 4.5 mm regardless of layer number, with a maximum absolute width error no greater than 0.45 mm and a mean squared error below 0.25 mm². Three-dimensional scanning of the finished components told the definitive story: height deviations of only 0.3 mm after fifty layers and width deviations of just 0.1 mm, compared with 5 mm and 0.7 mm respectively for the open-loop part. The top surface was smooth, bead-to-bead consistency within each layer was excellent, and the dents and defects that plagued constant-parameter printing vanished.

The authors are candid about limits. The filtering algorithm currently applies only to direct-current arc voltage signals in steel; highly reflective metals such as aluminum, typically welded with alternating current, will demand new denoising and vision algorithms to suppress noise and glare. The demonstration parts were simple cubes, and extending closed-loop control to curved geometries will strain the arc-length models, since curved trajectories continuously shift local heat accumulation, wire position, and arc shape. Still, the strategy offers a practical path forward: as a data-driven software solution, it can be integrated into existing industrial control systems through modest secondary development, without expensive hardware replacement or production downtime. For an industry seeking to print large, accurate metal parts at low cost, two humble sensors and a pair of cleverly tuned controllers may prove to be the difference between a promising laboratory process and a factory-floor reality.

Subject of Research: Collaborative closed-loop control of deposition width and height in gas tungsten arc-directed energy deposition using integrated arc voltage and vision sensing

Article Title: Collaborative control of deposition width and height for thick-walled parts in arc-directed energy deposition via integrating arc voltage and vision sensing

Article References: Cai, Y., Chen, D., Chen, H., Zhang, G., Pan, Z., & Xiong, J. (2026). Collaborative control of deposition width and height for thick-walled parts in arc-directed energy deposition via integrating arc voltage and vision sensing. Advanced Materials Joining, 1(1), Article 15. https://doi.org/10.1007/s44500-026-00017-w

Image Credits: AI Generated

DOI: 10.1007/s44500-026-00017-w

Keywords: arc-directed energy deposition, GTA-DED, additive manufacturing, arc voltage sensing, vision sensing, closed-loop control, model reference adaptive control, fuzzy control, molten pool monitoring, thick-walled parts, wire arc additive manufacturing, forming accuracy

The central challenge in cardiac electrophysiology is that heart cells communicate through rapidly propagating electrical waves. In a healthy heart, a precisely timed wave of depolarization sweeps across the muscle, triggering coordinated contraction. In diseased tissue, these waves can fragment, circle back on themselves, or originate from ectopic sites, producing arrhythmias that range from benign to lethal. Conventional interventions, from antiarrhythmic drugs to implanted pacemakers and ablation catheters, act on slow timescales or with coarse spatial resolution. What has been missing is a tool that can both observe and modulate cardiac electrical activity at the scale of individual cells, on millisecond timescales, within a continuous feedback loop.

The new work addresses this gap by exploiting optogenetics, a technique in which light-sensitive proteins borrowed from microbes are expressed in target cells. When blue light strikes channelrhodopsin, a light-gated ion channel embedded in the cell membrane, the channel opens and positive ions flow inward, depolarizing the cell and triggering an action potential. By genetically engineering human induced pluripotent stem cell-derived cardiomyocytes to express such opsins, researchers gain a remote, genetically specified actuator: any region of the cellular network can be electrically stimulated simply by illuminating it, with no physical contact required.

Stimulation alone, however, is only half of the control problem. The other half is sensing. The system pairs optogenetic actuation with optical voltage imaging, using fluorescent indicators whose emission changes with membrane potential. High-speed cameras capture the fluorescence of the cardiomyocyte network frame by frame, allowing the researchers to reconstruct the electrical state of the tissue in real time: which cells are resting, which are firing, and how excitation waves are propagating across the culture. This optical readout replaces the electrode arrays traditionally used to map cardiac activity, eliminating the invasiveness, wiring complexity, and spatial limitations of contact-based sensing.

The truly novel element is the holographic light engine that ties sensing and actuation together. Rather than illuminating the culture with a uniform beam or scanning a single laser spot, the researchers use a spatial light modulator to shape light into arbitrary two-dimensional patterns, projected onto the cell layer through holographic principles. A computer-generated hologram determines, pixel by pixel, where light intensity is delivered. This means the system can stimulate a single cell, a stripe of tissue, a curved wavefront mimicking the sinus node, or multiple disconnected regions simultaneously, all with subcellular spatial resolution and microsecond-scale temporal precision. The hologram can be updated faster than the dynamics of a cardiac action potential, which is essential for genuine real-time control.

Closing the loop requires software that can translate what the cameras see into what the light projector should do next. The control algorithm continuously monitors the optical voltage signals, compares the observed electrical behavior against a desired target state, and computes the illumination pattern needed to drive the network toward that state. If an excitation wave propagates too slowly, the system can deliver light pulses ahead of the wavefront to accelerate it. If an unwanted wave appears in the wrong location, the system can suppress it or redirect it. If the goal is a specific pacing frequency, the controller adjusts the timing and geometry of optical stimuli on every beat, compensating for the natural variability of biological tissue. This is the defining feature of closed-loop control: the intervention is not preprogrammed but continuously recalculated from live measurements.

The researchers demonstrated that this architecture can reliably entrain human cardiomyocyte networks to desired pacing patterns, guiding the rhythm of electrically active tissue that would otherwise beat at its own intrinsic rate. Beyond simple pacing, the holographic system’s spatial freedom enables more sophisticated interventions, such as shaping the direction and curvature of propagating waves or confining activity to defined regions of the network. Such capabilities are directly relevant to the study of arrhythmia mechanisms, where reentrant waves, spiral waves, and conduction blocks are the underlying culprits. A tool that can create, steer, and terminate such waves on demand in human-derived tissue provides an unprecedented experimental platform for arrhythmia research.

The significance for drug development and precision medicine is considerable. Human induced pluripotent stem cell-derived cardiomyocytes already allow pharmaceutical researchers to test compounds on human heart cells rather than animal tissue, but standard assays capture only bulk behavior, such as average beat rate or field potential duration. A closed-loop optical system adds an active dimension: it can probe how a tissue responds to perturbation, measure its vulnerability to arrhythmia induction, and quantify the effects of drugs on conduction velocity, refractory periods, and wave dynamics under precisely controlled stimulation conditions. In principle, patient-specific cell lines could be engineered with opsins and screened not just for passive responses but for behavior under stress, revealing proarrhythmic risks that conventional tests miss.

Looking further ahead, the all-optical nature of the approach suggests possibilities beyond the laboratory dish. Because neither sensing nor actuation requires physical contact, the conceptual framework is compatible with future cardiac therapies in which light delivered through optical fibers or implanted micro-LEDs could pace or resynchronize heart tissue in a feedback-controlled manner, guided by optical or electrical sensors. Such light-based pacemakers could adapt their stimulation pattern beat by beat, something conventional devices, which deliver fixed electrical pulses on fixed schedules, cannot do. Significant hurdles remain before any clinical translation, including delivering opsins safely to adult human myocardium, achieving sufficient light penetration in thick tissue, and ensuring long-term stability of both the genetic and optical components. The current study is confined to engineered cell networks in vitro, and the authors’ achievement should be understood as a foundational demonstration of control methodology rather than a therapy.

Even within that scope, the work marks a conceptual milestone. It shows that a living, electrically excitable human tissue can be observed, modeled, and steered in real time by a machine that touches nothing, intervening only through shaped light. The convergence of optogenetics, holographic projection, fast fluorescence imaging, and feedback control points toward a broader paradigm in synthetic biology and bioelectronic medicine: organs and organoids treated not as passive specimens but as dynamic systems that can be regulated the way engineers regulate any other process. For cardiac science, where rhythm is everything, the ability to write rhythm into human heart tissue with light, and to correct it when it goes wrong, may reshape how arrhythmias are studied, how drugs are validated, and, eventually, how failing electrical systems in the heart are repaired.

Subject of Research: All-optical closed-loop control of human cardiomyocyte networks using holographic optogenetics

Article Title: All-optical closed-loop control of human cardiomyocyte networks exploiting holographic optogenetics

Article References: Wendland, R., Schmieder, F., Sikandar, M. A., Knüppel, F. P., Zimmermann, W.-H., Bergmann, O., Büttner, L., & Czarske, J. W. (2026). All-optical closed-loop control of human cardiomyocyte networks exploiting holographic optogenetics. Communications Engineering, 5(1), Article 159. https://doi.org/10.1038/s44172-026-00779-1

Image Credits: AI Generated

DOI: 10.1038/s44172-026-00779-1

Keywords: holographic optogenetics, cardiomyocytes, closed-loop control, cardiac electrophysiology, optical voltage imaging, arrhythmia, induced pluripotent stem cells, channelrhodopsin, spatial light modulator, cardiac tissue engineering, bioelectronic medicine, optogenetic pacing