FALCON-S: Fixed-wing ground-effect Aerodynamics Simulator and Flight Control Learning Suite

The sky is not a uniform void for an aircraft; it is a landscape that changes dramatically the closer a vehicle gets to the earth. When a fixed-wing plane flies high above the ground, the air flows around its wings in a predictable, steady manner. However, as that same plane descends to within a few feet of the surface, the air trapped between the wing and the ground becomes compressed. This phenomenon, known as ground effect, acts like a cushion, increasing the lift that keeps the plane aloft while simultaneously reducing the drag that tries to slow it down. For pilots and engineers, this is a powerful force that can make a heavy aircraft feel lighter and more responsive, but it also introduces a layer of complexity that standard flight models often ignore. Designing autonomous systems that can safely navigate this low-altitude zone requires a deep understanding of these shifting aerodynamic forces, yet creating a virtual environment that accurately captures them has long been a significant hurdle for researchers.

To bridge this gap, a team of researchers has developed a new, highly detailed simulation framework specifically designed for fixed-wing aircraft flying near the ground. Unlike previous tools that either simplified the physics to save computing power or focused on hovering drones, this system models the full, complex motion of a rigid airplane as it interacts with the earth below. The researchers built a modular engine that simulates not just the flight itself, but also the imperfections found in the real world, such as the slight delays in a motor's response, the static in a sensor's signal, and the unpredictable gusts of wind. By combining this physical realism with the ability to run thousands of simulations simultaneously on powerful graphics processors, they have created a testing ground where artificial intelligence can learn to fly with a level of precision that was previously impossible to achieve in a virtual setting.

The core of this work is a simulator that treats the aircraft as a solid object moving through six degrees of freedom, meaning it can move forward, backward, left, right, up, and down, while also rotating in all directions. The software calculates the forces acting on the plane by using a semi-empirical model, which blends theoretical physics with real-world data derived from the aircraft's shape. Crucially, the system applies a specific correction for ground effect, adjusting the lift and drag calculations based on how close the wings are to the surface. This allows the simulation to replicate the exact moment a plane feels the "cushion" of air, altering its behavior in ways that a standard flight model would miss. The framework is flexible enough to test different types of aircraft, from small remote-controlled models to larger prototypes, and it can be run on standard computer processors or accelerated massively using graphics cards to speed up the learning process.

In their experiments, the researchers put this simulator to the test by pitting two different types of flight controllers against each other: traditional mathematical controllers that rely on pre-calculated rules, and modern learning-based controllers that improve through trial and error. They tasked these systems with two primary challenges: maintaining a steady altitude after climbing to a specific height, and following a series of complex maneuvers like circles and figure-eights. The results revealed a fascinating trade-off between stability and adaptability. The traditional controllers were incredibly precise when the conditions were exactly as predicted, holding an altitude with near-perfect accuracy. However, they struggled when the aircraft had to climb from a distance or when the conditions changed slightly, often failing to reach the target height at all.

In contrast, the learning-based controllers, specifically one trained using a method called Proximal Policy Optimization, proved to be far more robust. These systems were able to successfully reach and hold the target altitude in every single test run, regardless of the starting position. They also demonstrated a remarkable ability to track the complex maneuvers, keeping the aircraft on course with a high degree of consistency. While the traditional controllers sometimes failed to even complete the task, the learning agents managed to fly through the entire sequence of turns and climbs without crashing. The study showed that while the traditional methods could be more accurate in a perfect, static environment, the learning-based approaches were better at handling the dynamic and unpredictable nature of real flight, especially when ground effect was involved.

The researchers also investigated how these systems performed when faced with the imperfections of the real world, such as noisy sensors, delayed commands, and turbulent winds. They found that the learning-based controllers were surprisingly resilient to sensor noise, maintaining their performance even when the data they received was slightly corrupted. However, they were more sensitive to delays in the system, where a lag between the command and the aircraft's movement caused the plane to wobble or lose control. Turbulence proved to be the most challenging factor, with the aircraft sometimes being pushed out of its safe flight envelope by strong gusts. Interestingly, the size and design of the aircraft mattered significantly; a smaller, lighter model was able to absorb some of the delays better than a larger one, yet it was the smaller model that was more easily knocked off course by the wind.

One of the most significant findings of the study was the quantification of how much thrust an aircraft needs to save when flying in ground effect. The simulations showed that as a plane descends to a height equal to one-quarter of its wingspan, it requires roughly 25 percent less power to maintain level flight compared to flying higher up. This reduction in power demand is a direct result of the increased lift and reduced drag provided by the ground effect. The researchers confirmed that this saving is real and measurable, noting that the aircraft can fly the same path with significantly less energy expenditure when it is close to the ground. This insight is vital for designing efficient autonomous systems that need to conserve battery life or fuel while operating in low-altitude environments.

The team validated their simulator by comparing its results with two other established flight simulation tools, ensuring that the underlying physics were sound. When ground effect was turned off, their system produced flight paths that matched the other tools almost perfectly, confirming that the basic aerodynamics were correct. When ground effect was enabled, the system showed a distinct and systematic change in behavior that the other tools did not replicate, proving that the new model was successfully capturing the unique aerodynamic phenomena of low-altitude flight. This cross-validation gives confidence that the simulator is not just a theoretical exercise, but a reliable tool for engineering and research.

Ultimately, this work provides a unified platform for testing and comparing different flight control strategies under realistic conditions. It demonstrates that while traditional control methods have their place, learning-based approaches offer a superior ability to adapt to the complex and changing environment of low-altitude flight. The simulator allows researchers to run millions of flight scenarios in a matter of minutes, accelerating the development of autonomous aircraft that can safely navigate the ground effect zone. By providing a common, high-fidelity benchmark, this framework helps to standardize how flight controllers are evaluated, moving the field closer to the day when autonomous fixed-wing aircraft can operate reliably in the challenging airspace just above the earth.