In the basement of a university engineering building, a small drone hovers in a vertical wind tunnel. The air rushes upward from beneath it — not uniformly, but in a carefully sculpted pattern that mirrors what the drone would encounter soaring on the edge of a thermal updraft in the real world. The drone's wings catch this artificial breeze with surprising grace, staying aloft far longer than it could on its own. Outside, no wind blows. No thermals rise. And yet this machine has learned to ride invisible waves of air, because researchers have learned how to paint those waves into existence.
This is the quiet promise of a new approach to wind-tunnel control, developed by Ghadeer Elmkaiel and Michael Muehlemann at the University of Stuttgart — a method for sculpting airflow with the precision of a chef seasoning a dish, using feedback and learning to shape the invisible currents that determine whether a flying robot soars or stalls. The implications stretch far beyond this basement laboratory. Wind tunnels are the backbone of aerospace engineering, the controlled arenas where everything from paper airplanes to supersonic jets prove their worth. But traditional wind tunnels offer a blunt instrument: uniform, homogeneous airflow, like a flat skillet when you need a gradient. The researchers' work suggests we can do better. Much better.
The Invisible Problem With Wind Tunnels
To understand why this matters, you need to understand what wind tunnels actually do — and why their limitations have constrained aerial robotics research for decades.
A wind tunnel is, at its core, a controlled environment for studying how objects interact with moving air. Air is drawn or pushed through a large tube, and a test object sits within that flow. Engineers can measure forces, observe behavior, and identify failures before a design ever takes flight in the real world. The physics of flight — drag, lift, turbulence, stall — all reveal themselves in these controlled conditions.
But here's the problem: real-world airflow is almost never uniform. Birds soaring on thermals encounter columns of rising air surrounded by still air or even sinking currents. A drone flying near a building encounters chaotic wakes, turbulence shed from edges, pressure differentials that change moment by moment. A micro air vehicle navigating a forest canopy experiences wildly different conditions as it passes from sunlit clearings into shaded patches where temperature gradients create invisible eddies.
Wind tunnels, for all their utility, have historically offered something closer to a laboratory approximation than a faithful reproduction of real-world conditions. You can change the speed. You can change the angle. But creating complex, spatially varying airflow patterns — the kind of structured complexity that defines natural flight environments — has been prohibitively difficult.
Multi-fan wind tunnels offer a partial solution. Rather than a single large fan pushing air through the entire tunnel, these systems use an array of smaller fans distributed across the tunnel's cross-section. Each fan can be controlled independently, allowing different regions of the tunnel to experience different airflow velocities. In principle, this enables the creation of complex flow profiles — parabolic gradients, Gaussian peaks, complex vector fields that match the structure of real-world atmospheric conditions.
In principle. In practice, controlling these systems to produce a desired flow pattern has proven remarkably difficult. The challenge is that the relationship between fan speeds and the resulting airflow is nonlinear, complex, and highly dependent on the specific geometry of the tunnel and the properties of the air itself. Small changes in one region can propagate through the entire system in unpredictable ways. Traditional control approaches, built on precise mathematical models and well-defined dynamics, struggle with this kind of coupled, interdependent system.
"The development and testing of advanced aerial robots require experiments in controlled environments with tailored airflow profiles," the researchers note, framing the problem with quiet understatement. What they're really saying is that the next generation of aerial robotics — drones that can perch like birds, glide like albatrosses, soar like hawks — cannot be developed and refined in the blunt instruments that most laboratories have access to. The field needs a new kind of wind tunnel. One that can speak the complex language of natural airflow.
Learning to Shape the Wind
The core innovation in Elmkaiel and Muehlemann's work is an algorithm that learns to control a multi-fan wind tunnel by combining two sources of knowledge: a simplified physical model of how airflow behaves, and iterative measurements of what actually happens when the fans are set to particular speeds.
The approach is called online learning model predictive control, though the name matters less than the philosophy. The researchers start with a basic understanding of fluid dynamics — the mathematical framework that describes how air moves and how forces propagate through it. This provides a foundation, a set of constraints that any realistic solution must respect. But they don't try to build a complete, perfectly accurate model from first principles, because such models are notoriously difficult to construct and computationally expensive to solve.
Instead, they accept that their simplified model will be imperfect. And they compensate for that imperfection through iterative learning: run an experiment, measure the actual airflow that results, compare it to what was desired, update the model, adjust the fan speeds, and try again. Each iteration brings the system closer to the target, with the learning algorithm automatically determining how to adjust controls based on the observed errors.
This iterative, measurement-based approach has several advantages. First, it doesn't require exhaustive first-principles modeling of every aspect of the system. The algorithm learns the quirks and idiosyncrasies of the specific wind tunnel it's installed in — the particular way fans interact with each other, the subtle asymmetries introduced by structural elements, the small nonlinearities that make real systems diverge from idealized mathematical descriptions. Second, it's adaptive. If something changes — a fan ages and produces slightly different flow characteristics, a new test object is introduced that alters the flow field, the room temperature shifts and changes air density — the algorithm can adapt without requiring manual recalibration. Third, it's sample-efficient. The researchers emphasize this point: their method achieves convergence to desired airflow distributions with relatively few measurements, making it practical for real laboratory use rather than requiring thousands of experiments to train a model.
The mathematics underlying this approach draws from the field of adaptive control and Gaussian process regression. The simplified physical model provides a baseline prediction of how fan inputs map to flow outputs. Gaussian processes — a flexible, nonparametric approach to statistical modeling — are used to represent the uncertainty in that model and to update it based on new measurements. Model predictive control provides the framework for determining optimal fan settings at each iteration, accounting for both the current best estimate of the system dynamics and the remaining uncertainty about those dynamics.
It's a principled approach that balances prior knowledge (from the simplified physics) with empirical data (from iterative measurements), allowing the algorithm to exploit what it knows while remaining open to learning what it doesn't.
From Uniform Flow to Complex Patterns
The versatility of the method is demonstrated through its ability to generate several different types of airflow profiles — not just uniform flow that any decent wind tunnel can produce, but complex patterns that capture the structure of real atmospheric conditions.
A uniform airflow profile is the simplest case: wind moving at the same velocity everywhere in the tunnel cross-section. This is what traditional wind tunnels offer, and it's useful for many baseline experiments. But the researchers go beyond this to generate Gaussian profiles — airflow that peaks in the center and falls off smoothly toward the edges, mimicking the structure of a thermal column rising from a sun-warmed patch of ground. They also demonstrate parabolic profiles, with velocity varying quadratically across the tunnel — the kind of boundary layer flow that develops near surfaces due to viscous effects.
Each of these profiles represents a qualitatively different flow condition, and the ability to generate them with a single hardware platform — simply by changing the target that the learning algorithm is asked to match — dramatically increases the experimental flexibility available to researchers. A single wind tunnel becomes a multi-environment testing facility, capable of simulating conditions ranging from open atmosphere to edge-of-thermal to near-ground effects.
The learning algorithm's job is to determine the specific fan speed configuration that produces each target profile. For a uniform profile, this might mean setting all fans to similar speeds. For a Gaussian profile centered in the tunnel, it means identifying the combination of fan speeds that creates a velocity peak in the center with appropriate falloff. For a parabolic profile, it means solving a different optimization problem entirely.
The researchers show that their algorithm converges to these different target profiles from scratch — without requiring a-priori knowledge of the exact fan-speed-to-flow mapping for each profile type. Start from an initial guess based on the simplified physical model, measure the actual flow that results, update, adjust, and repeat. Within a finite number of iterations, the system settles into a configuration that produces the desired flow pattern to within specified tolerances.
Soaring Without Effort
But perhaps the most compelling demonstration is saved for the end: the algorithm's ability to produce an airflow profile specifically designed for passive soaring.
Passive soaring is the art of staying aloft without consuming energy — using the energy already present in the atmosphere to maintain or gain altitude. Birds have mastered this skill over millions of years of evolution. Albatrosses ride ocean waves of wind, gaining energy from vertical wind gradients near the ocean surface. Hawks circle in thermals, converting the upward motion of warm air into altitude they can later use to glide elsewhere. Humans have attempted to replicate this with gliders and sailplanes, but these require initial launch and still experience drag that gradually depletes their energy reserves.
For small unmanned aerial vehicles (UAVs), passive soaring represents a tantalizing possibility: a robot that can stay airborne for hours or even days without battery consumption, powered entirely by the energy inherent in atmospheric flows. This has obvious applications for long-endurance surveillance, environmental monitoring, and communication relay. But achieving reliable passive soaring has proven challenging, in part because it requires maintaining precise trim and orientation as the aircraft rides spatially varying wind fields.
Elmkaiel and Muehlemann demonstrate that their algorithm can generate a wind profile specifically tailored to enable passive soaring behavior in a small test aircraft. They design a target flow field that provides the optimal combination of lift and energy input for their test platform — a wind profile that mirrors the conditions a soaring bird might encounter at the edge of a thermal or along a wind shear layer. The learning algorithm then converges to a fan configuration that produces this profile.
When the test aircraft is placed in this artificially sculpted airflow, the results are dramatic. The researchers report greatly enhanced flight performance — the aircraft maintains stable flight with minimal or no control input, riding the gradient like a surfer on a wave. This is passive soaring in a wind tunnel: the aircraft extracts energy from the structured flow field without any active propulsion.
This demonstration connects the algorithm's abstract capability (producing complex flow profiles) to a concrete, application-relevant outcome (enabling energy-efficient UAV flight). It's one thing to generate a Gaussian profile that matches a theoretical curve. It's another thing entirely to generate an airflow that lets a robot soar.
The Engineering Details
To understand how this works in practice, it helps to have a sense of the experimental setup.
The researchers use a multi-fan vertical wind tunnel with an array of independently controllable fans distributed across its cross-section. The exact number of fans and their arrangement would be specified in the detailed methods section of the paper, but the general architecture is an array capable of producing spatially varying flow fields rather than uniform conditions. Each fan's speed can be individually commanded, allowing the system to create different velocity patterns across the tunnel's cross-section.
The learning algorithm operates iteratively. At each step, it receives measurements of the actual airflow at various points in the tunnel — this could come from an array of sensors, perhaps hot-wire anemometers or particle image velocimetry, depending on the specific implementation. The algorithm compares these measurements to the target profile and updates its estimate of the relationship between fan speeds and flow outputs.
The simplified physical model that anchors the approach is derived from fluid dynamics principles. In its simplest form, this might be a linear model relating fan speeds to local flow velocities, but with corrections for the most important nonlinear effects. The learning component then captures deviations from this simplified model based on empirical observations.
The update strategy uses principles from Bayesian optimization and Gaussian process regression, allowing the algorithm to balance exploration (trying configurations where it's uncertain about the outcome) with exploitation (trying configurations that are likely to improve performance based on current knowledge). This is crucial for achieving sample efficiency — the algorithm learns quickly rather than requiring exhaustive testing of every possible fan configuration.
The researchers validate their approach by demonstrating multiple profile types and showing that their algorithm can generalize across different targets. They also test robustness: what happens when the number of operating fans is reduced, either due to failure or deliberate modification? The learning algorithm adapts, finding new fan configurations that approximate the target profile despite the reduced degrees of freedom.
Why Sample Efficiency Matters
The emphasis on sample efficiency deserves special attention, because it's a crucial practical consideration that distinguishes this work from naive learning approaches.
In an ideal world, you might imagine that the solution to complex control problems is simply to collect more data: try many configurations, measure the results, and use machine learning to infer the optimal control strategy. This approach has shown success in some domains — board games, video games, robot manipulation tasks — but it typically requires enormous amounts of data. Millions of trials, billions of observations, weeks or months of computation.
Wind tunnels don't typically accommodate this approach. Experiments take time to set up and run. Measurements require physical sensors that must be positioned and calibrated. The cost per trial isn't negligible. An algorithm that requires thousands of iterations to converge would be impractical for real laboratory use.
The researchers' emphasis on sample efficiency means their algorithm can achieve good performance with a manageable number of measurements — enough to characterize the system and refine the model, but not so many that the process becomes impractical. This is achieved through careful choice of the learning algorithm (focusing measurements where they're most informative), appropriate use of the simplified physical model (reducing how much must be learned from scratch), and principled update strategies (avoiding redundant or low-value experiments).
The practical implication is that this approach could be deployed in actual laboratory settings, not just demonstrated in research prototypes. A research group could implement this algorithm, run the learning procedure to characterize their specific wind tunnel, and then use the resulting control system to generate complex flow profiles on demand.
From the Lab to the World
What does this mean more broadly? Why should people outside aerospace engineering or robotics care about a better algorithm for wind-tunnel control?
The answer lies in what wind tunnels enable. These are the facilities where the machines that will populate our skies are developed and refined. Drones that inspect infrastructure, deliver packages, monitor crops, and survey disaster zones all pass through wind tunnels during their development. So do the aircraft that will eventually fly passengers and cargo without pilots. The insights gained in these controlled environments shape designs that will operate in the uncontrolled, unpredictable real world.
If wind tunnels can only produce uniform flow, then the testing they enable is limited to uniform conditions. Drones developed and refined in uniform flow may perform well when encountering uniform conditions in the real world — and may struggle or fail when encountering the complex, structured flows that actually characterize most environments. This creates a gap between laboratory performance and field performance that engineers must close through conservative design margins, extensive field testing, or iterative redesign.
Better wind tunnel capabilities enable better development processes. If you can test a drone in conditions that more closely approximate real-world flight — if you can expose it to thermal-like gradients before it encounters a real thermal, if you can simulate the wind shear near buildings before deploying it for urban delivery — then you can identify and address performance limitations earlier, when they're cheaper to fix.
The researchers specifically highlight "advanced aerial robots" as the beneficiaries of their approach. This category encompasses a wide range of systems: drones that perch like birds, that glide between thermals, that navigate complex urban environments with precision. These systems don't just need to fly — they need to exploit the structure of the airflow they encounter, converting environmental energy into sustained flight. Developing such systems requires testing platforms that can reproduce the structured complexity of real atmospheric flows.
Passive soaring represents a particularly compelling application. If UAVs can reliably extract energy from atmospheric flows, their endurance limitations disappear. A drone that can soar doesn't need larger batteries or more fuel. It simply rides the energy that's already there, like a sailboat catching a favorable wind. This has implications for everything from environmental monitoring (imagine a drone that stays aloft over a forest for weeks, monitoring for fires) to infrastructure inspection (a drone that patrols power lines indefinitely without needing to land for recharging).
But achieving reliable passive soaring requires understanding how aircraft interact with spatially varying flows — and that understanding requires experimental platforms capable of producing those spatially varying flows. The researchers' work provides such a platform, demonstrating not just that complex flows can be generated but that they can be generated in ways that enable specific, application-relevant behaviors like passive soaring.
The Broader Context: Adaptive Systems and the Real World
Beyond the specific application to wind tunnels, this work touches on a broader theme in robotics and control: the challenge of deploying autonomous systems in the real world, where conditions are rarely exactly as models predict.
Traditional control theory excels at problems where the system dynamics are well understood and can be captured in precise mathematical models. Rocket guidance, process control in chemical plants, stabilization of rigid bodies — these are domains where the physics is well-characterized and the models are accurate. Control algorithms built on these models can achieve remarkable precision.
But many systems resist this treatment. Biological systems, complex robots with many interacting degrees of freedom, environments where conditions change over time — these are domains where our models are necessarily incomplete, where the gap between theory and reality cannot be closed through more careful mathematics alone.
The adaptive, learning-based approach demonstrated in this paper represents one response to this challenge. Rather than trying to build a perfect model from first principles, acknowledge that the model will be imperfect and design the control system to compensate. Use iterative learning to bridge the gap between model and reality. Adapt when conditions change.
This philosophy is increasingly common in robotics research. It reflects a growing recognition that pure model-based approaches have limits, and that incorporating learning — from data, from experience, from feedback — can enable robust performance in domains where purely analytical approaches struggle.
The wind tunnel application is particularly clean for exploring these ideas because the learning can happen during dedicated experiments before the system is deployed for testing. This separates the learning phase from the testing phase in a way that wouldn't be possible for, say, a drone navigating unknown terrain. But the principles — use physical insight where you have it, learn from data where you need to — transfer to many domains.
Implications for Testing Methodologies
The research also raises interesting questions about how aerial robots should be tested and validated.
Current standard practice involves extensive flight testing in real environments, often supplemented by simpler laboratory experiments that capture limited aspects of the flight envelope. Wind tunnels are used for certain tests, but the limitation to uniform flow means they can only capture a subset of the conditions that real aircraft will encounter.
If wind tunnels could more faithfully reproduce real-world flow conditions, testing methodologies could evolve. More development could happen in the controlled laboratory environment, where failures are safe and conditions are reproducible. Less testing would need to happen in the field, where failures can be expensive and where conditions are harder to control and reproduce.
This matters because field testing is expensive and time-consuming. Field failures are costly. And field testing doesn't allow the kind of systematic exploration of edge cases that laboratory experiments enable. If you can test your drone's behavior in a thermal-like gradient, a wind shear layer, a complex urban wake — all in the lab, all under controlled conditions — you can develop and validate robust performance before deploying to the field.
The researchers' work doesn't immediately enable this vision — it would need to be extended and validated further before becoming a standard testing methodology. But it points in this direction. An algorithm that can reliably generate complex flow profiles in wind tunnels, adaptable to different target conditions and robust to component variations, is a step toward more capable testing infrastructure.
Open Questions and Future Directions
The paper demonstrates a compelling proof of concept, but several questions remain for future work.
First, how well does the approach scale to larger wind tunnels? The demonstrations in the paper likely occurred in a relatively modest-sized facility. Scaling to industrial-scale wind tunnels, where the flow volumes are larger and the number of fans is greater, would require additional work on computational efficiency and algorithm scalability. The learning algorithm must run fast enough to converge in a reasonable number of iterations, and the computational cost of updating the model must not grow too quickly with system complexity.
Second, how robust is the approach to disturbances during actual testing? The current demonstrations focus on establishing target flow profiles when the tunnel is operating with a stationary test article. But what happens when a drone is actively flying within the flow field? Does its presence disturb the flow in ways that degrade performance? Can the system adapt in real time to disturbances introduced by the test article? These are practical questions that would need to be addressed for operational deployment.
Third, what types of flow profiles can actually be generated in practice? The paper demonstrates uniform, Gaussian, and parabolic profiles, as well as a custom soaring profile. But atmospheric flows exhibit many other patterns: helical vortices, turbulent cascades, multi-pole flows with complex spatial structure. Understanding the full space of achievable profiles — what can and cannot be produced by a finite array of fans — is important for setting realistic expectations and identifying the cases where the approach is most valuable.
Fourth, how does the algorithm perform across different air densities and temperatures? Air properties affect flow behavior, and wind tunnels are typically operated under controlled conditions. If the algorithm relies on specific assumptions about air properties, its performance might vary with atmospheric conditions. This could matter for outdoor testing applications, where temperature and pressure vary with weather and altitude.
These are not fatal flaws in the approach — rather, they're natural questions for future research to address. The paper establishes the core idea and provides compelling demonstrations; extending the approach to broader scenarios and validating it at scale are natural next steps.
The Bigger Picture: Environment-Aware Robotics
Stepping back from the specific technical details, this work fits into a broader narrative about the evolution of aerial robotics.
Early drones were largely indifferent to their environment. They flew using simple control strategies that treated wind as a disturbance to be rejected rather than a resource to be exploited. If a gust pushed them off course, they corrected. If they encountered turbulent conditions, they struggled. Flight was about overcoming the environment, not working with it.
This is changing. The next generation of aerial robots is increasingly designed to sense and respond to atmospheric conditions — to find and ride thermals, to exploit wind gradients, to navigate complex flows rather than simply surviving them. This requires both better sensing (knowing what the wind is doing) and better control (knowing how to use that information).
But it also requires better testing. If you want to develop a drone that soars, you need to test it in soaring conditions. If you want to develop a drone that navigates urban wakes, you need to test it in simulated wake conditions. The development and testing of environment-aware aerial robots requires experimental platforms that can produce the structured, spatially varying flows that these robots will encounter in the field.
The researchers' work is a step toward such platforms. By enabling wind tunnels to generate complex flow profiles on demand, it expands the space of conditions available for testing and development. An algorithm that can learn to produce arbitrary target profiles — adapting to the specific tunnel, accounting for its idiosyncrasies — is a practical tool for enabling this vision.
The broader goal is robots that are graceful rather than just stable, that exploit their environment rather than fighting it. This is what birds have achieved through millions of years of evolution. It's what engineers are now beginning to achieve through careful design and, increasingly, through learning.
A Different Kind of Wind
The next time you see a drone hovering in a demonstration, consider the invisible constraints that shaped its design. Somewhere, in a wind tunnel or a simulation, it was tested against conditions that its engineers hoped would approximate the real world. If those tests involved uniform flow — flat, homogeneous, unlike anything in nature — then the drone was developed to perform in conditions it may never actually encounter.
The approach developed by Elmkaiel and Muehlemann suggests a different path: wind tunnels that can paint complex flows into existence, that can reproduce the structure of thermals and wind shears and urban wakes, that can give engineers the testing environments their robots actually need. An algorithm that learns to shape the wind.
This isn't just a technical achievement. It's a shift in what's possible — in what questions can be asked, what designs can be tested, what robots can be built. The flying machines of the future may be designed in wind tunnels that can speak the complex language of real atmospheric flows. They may be tested in conditions that prepare them for the messy, structured, unpredictable world they'll actually encounter.
And they may soar on winds that engineers have sculpted to teach them how.
The Path Forward
The work demonstrates that complex wind profiles can be reliably generated in multi-fan wind tunnels using iterative learning algorithms that combine physical insight with empirical measurement. The key achievements — generating diverse profile types, enabling passive soaring behavior, adapting to component variations — point toward practical utility for aerial robot development.
What remains is the work of translation: turning this proof-of-concept into a deployable system, scaling it to larger facilities, integrating it into standard testing workflows, and extending it to cover the full range of flow conditions that aerial robots may encounter. This is the agenda for future research, and it's a tractable one.
The researchers have shown the core idea works. They have demonstrated it in compelling application contexts. They have highlighted its practical advantages — sample efficiency, adaptability, robustness — that make it suitable for real laboratory deployment. What's needed now is continued development and validation.
For the broader community interested in the development of aerial robotics, this work offers a glimpse of testing capabilities that don't yet exist at scale but could. For researchers working on adaptive control and learning-based robotics, it offers an application context where these techniques provide clear practical value. For engineers designing the next generation of flying machines, it suggests that the limitations of current testing infrastructure may not be fundamental — that they can be overcome with clever algorithms and principled integration of models and data.
The wind can be shaped. The tunnel can learn. And the robots that fly in these engineered winds may one day soar on patterns that their designers learned to paint.
The work stands at the intersection of multiple trends: the increasing sophistication of aerial robots, the growing interest in environment-aware autonomous systems, the continued development of learning-based approaches to control, and the practical need for better testing and validation infrastructure. These trends will continue, and the work they enable will build on foundations like the one laid here.
For now, the demonstration speaks for itself: in a wind tunnel in Stuttgart, a small drone has learned to soar on artificially sculpted wind. Outside the tunnel, the air is still. But inside, the wind has been taught a new trick. And the drone has learned to ride it.