⚡ electrical engineering

GES: Ground Effect Stabilizer — Onboard-Sensor-Based Feedforward Estimation of Ground Effect on Irregular Terrain via Machine Learning versus Analytical Models

This paper demonstrates that while machine learning models can accurately predict ground effect on irregular terrain offline, their deployment in closed-loop control often fails to improve performance over analytical baselines due to unaccounted actuator-level dynamics and asymmetric sensitivity to overprediction biases, necessitating specific correction strategies like asymmetric clamping rather than relying solely on predictive accuracy.

Original authors: Muwon Lee

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muwon Lee

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When a drone hovers just above the ground, the air pushed down by its spinning blades hits the surface and bounces back up. This returning air pushes against the underside of the rotors, giving the drone a sudden, uncommanded boost in lift. Engineers call this the ground effect. For a drone flying over a perfectly flat floor, this boost is predictable; for a drone flying over a rocky hillside or a bumpy field, the effect changes constantly as the distance to the ground shifts and the shape of the terrain alters how the air flows. If the drone's computer does not anticipate this extra lift, the vehicle can start to bob up and down uncontrollably or crash into the ground. The challenge for modern robotics is to teach these machines to sense this invisible push and adjust their motors instantly, keeping them steady even when the ground beneath them is uneven and unpredictable.

A researcher set out to solve this problem using a new approach: instead of relying on fixed mathematical formulas to guess how the air would behave, they trained a computer to learn the pattern directly from the drone's own sensors. They built a highly detailed virtual world filled with procedurally generated, irregular terrain—landscapes that were rough, bumpy, and unique, rather than smooth and flat. Inside this simulation, they flew a virtual drone equipped with standard sensors that measure how fast it is spinning, how much power the motors are using, and how far it is from the ground. They taught several different types of machine-learning models to look at these sensor readings and predict exactly how much extra lift the drone would experience in the next tenth of a second. In a test where the computer simply looked at the data and made a prediction without actually flying the drone, the machine-learning models were spectacularly successful. They predicted the ground effect with near-perfect accuracy, far outperforming the traditional mathematical formulas that have been used for decades. The models learned to see the subtle signs in the sensor data that indicated the drone was about to hit a bump or dip, allowing them to forecast the lift boost before it happened.

However, the story took a surprising turn when the researcher put these high-performing models to work in a live flight simulation. They connected the best machine-learning model to the drone's flight controller, allowing it to make real-time adjustments as the drone hovered over the same rough terrain. The result was not a victory for the artificial intelligence. Despite the model's incredible ability to predict the ground effect in a static test, the drone flown with the machine-learning controller actually performed slightly worse than the drone flown with the simple, old-fashioned mathematical formula. The machine-learning drone hovered less smoothly and had slightly larger altitude errors. This was a baffling outcome: the computer was better at guessing the future, yet the drone flew worse. To understand why, the researcher ran a series of diagnostic tests, including one where they gave the flight controller the perfect, exact answer for how much lift to expect, removing any possibility of a prediction error. Even with perfect knowledge, the controller did not fly better than the simple formula. This proved that the problem was not that the machine learning model was bad at guessing; the problem lay in how the drone's motors responded to the information.

The researcher discovered that the issue was a mismatch between the prediction and the physical action required to cancel it. The drone's motors generate lift based on the square of their speed, meaning a small change in speed creates a large change in lift. When the machine-learning model predicted a strong ground effect, the flight controller tried to compensate by reducing the motor speed. Because the relationship between speed and lift is non-linear, the controller was applying the wrong amount of correction. It was like trying to turn off a faucet that reacts to pressure in a way that is not a straight line; if you turn the handle by a certain amount expecting a specific drop in water, you might actually get too much or too little. The researcher found that when the model slightly overestimated the ground effect, the drone would dip dangerously low, reducing its safety margin. This was a critical safety risk, even though the overall average height of the flight looked fine. The system was stable and would not spiral out of control, but the specific type of error caused by overestimating the lift made the drone fly closer to the ground than necessary.

To fix this, the researcher did not throw away the machine-learning model. Instead, they added a simple safety rule: if the model predicted a ground effect that was too strong, the system would ignore that extreme prediction and stick to the safer, more conservative estimate. This "asymmetric clamp" allowed the drone to use the machine learning's insights while preventing it from making dangerous overcorrections. With this adjustment, the performance gap between the machine-learning drone and the simple formula disappeared, and the drone flew just as well as the traditional method. The study concludes that for safety-critical tasks like flying a drone, being able to predict the future with high accuracy is not enough. A model must also be tested in the real loop of action and reaction, because a perfect prediction can still lead to a crash if the physical system does not respond in the way the computer expects. The researcher showed that even the smartest artificial intelligence needs a safety net when it interacts with the messy, non-linear reality of the physical world.

Technical Summary: GES — Ground Effect Stabilizer

Problem Statement
Multirotor unmanned aerial vehicles (UAVs) experience unmodeled thrust augmentation, known as ground effect (GE), when operating near the ground. While classical analytical models (e.g., Cheeseman–Bennett) effectively describe GE over flat, infinite ground planes, they fail to capture the complex aerodynamic interactions occurring over irregular terrain. On non-flat surfaces, local slope and curvature alter the recirculation of rotor downwash, leading to unmodeled thrust variations that cause altitude oscillations and control degradation. The central problem addressed is whether a data-driven model, trained solely on onboard-observable sensors (range-finder altitude, gyroscope rates, motor RPM/current, and commanded thrust), can outperform classical analytical approximations in predicting GE factors on irregular terrain, and crucially, whether such offline predictive accuracy translates to improved closed-loop control performance.

Methodology
The study utilizes a physics-based multirotor simulator featuring a rigid-body dynamics engine, a per-rotor motor model, and a procedurally generated irregular terrain generator. The terrain is created using a seeded pseudorandom number generator to ensure geometric diversity across training and evaluation sets.

Ground Truth Model: The simulator employs a heuristic "ground truth" GE model based on the Cheeseman–Bennett flat-ground formula, augmented with slope and curvature corrections. These corrections are heuristic rather than fitted to external validation data.
Offline Evaluation: Six machine learning (ML) regression models (Linear Regression, Decision Tree, Random Forest, Gradient Boosting, MLP, LSTM, and Transformer) were trained to predict the GE factor 100 ms into the future. A strict "seed-wise" data split was employed (240 training, 30 validation, 30 test seeds) to ensure no terrain data leakage between sets, guaranteeing generalization to unseen terrain geometries.
Closed-Loop Deployment: The best offline model (an MLP) was integrated into a PID-based flight controller as a feedforward correction term. The controller scales the throttle command by $1/\sqrt{\hat{f}}$ (where $\hat{f}$ is the predicted GE factor) to correctly account for the quadratic relationship between rotor speed and thrust.
Diagnostic Experiments:
- Oracle Controller: A controller using the simulator's exact ground-truth GE factor (zero prediction error) was tested to isolate prediction accuracy from actuator-level effects.
- Causal Bias Intervention: Fixed positive and negative biases were injected into the GE factor estimate to determine the causal impact of over- vs. under-prediction on safety margins.
- Stability Analysis: A time-limited bias pulse and linearized closed-loop pole analysis were conducted to determine if GE prediction errors induce self-reinforcing feedback loops.

Key Results

Offline Superiority: Under the seed-wise split, all ML models substantially outperformed the analytical baseline on held-out terrain seeds. The best model (Random Forest) achieved an $R^2$ of 0.9987 compared to 0.851 for the analytical baseline, confirming that ML can accurately learn the nonlinear mapping from sensors to GE factors on irregular terrain.
Closed-Loop Shortfall: Despite superior offline accuracy, the deployed hybrid ML controller performed consistently worse than the pure analytical baseline in closed-loop simulations across 30 unseen terrain seeds. The ML hybrid exhibited higher AGL tracking RMS error and gyroscope RMS.
The Oracle Finding: Critically, the "oracle" controller, which had access to perfect ground-truth GE factors (zero prediction error), performed no better than the simple analytical baseline. This indicates that the closed-loop performance gap is not primarily caused by prediction inaccuracy but by how GE estimates interact with the actuator mapping and control loop dynamics.
Causal Mechanism: Controlled bias experiments revealed that overprediction of the GE factor (positive bias) significantly degrades safety-relevant metrics, specifically reducing the 5th-percentile clearance and increasing low-clearance duration, even though aggregate AGL tracking RMS was minimally affected.
Stability: Contrary to initial hypotheses of a self-reinforcing feedback loop, the system was found to be stable. A transient overprediction pulse caused a temporary clearance loss, but the system returned to baseline without divergence. Closed-loop pole analysis confirmed negative real parts across the operating range, indicating an underdamped but stable return to equilibrium.
Mitigation: An asymmetric clamp, which specifically limits GE-factor overprediction while leaving underprediction unconstrained, successfully closed most of the performance gap between the ML hybrid and the analytical baseline.

Significance and Claims
The paper presents a dissociation between offline predictive accuracy and closed-loop control performance. Its primary contributions are:

Validation of Offline Metrics: The study demonstrates that high offline $R^2$ or RMSE is not a reliable proxy for closed-loop control performance. A model can perfectly predict the ground truth (as shown by the oracle) yet fail to improve control performance if the actuator mapping or control structure is not separately accounted for.
Methodological Rigor: The author emphasizes that plausible mechanistic stories based on correlation signs are insufficient. Direct causal intervention (bias injection) and dynamical analysis (pole analysis) are required to verify mechanisms, as initial assumptions about "self-reinforcing feedback loops" and causal directions were corrected through these rigorous tests.
Practical Implication: For learned control corrections, an "oracle ablation" (substituting the learned predictor with ground truth in simulation) is a valuable diagnostic tool to distinguish between prediction errors and structural/actuator-level issues.
Targeted Mitigation: The study shows that closed-loop shortcomings do not necessarily require abandoning learned models; rather, targeted interventions, such as asymmetric clamping of specific error modes (overprediction), can recover performance while retaining offline advantages.

The author concludes that while ML models can accurately predict GE on irregular terrain, their deployment requires careful consideration of actuator dynamics and safety margins, and offline metrics alone are insufficient for validating control-relevant learned models.

Technical Summary: GES — Ground Effect Stabilizer

More like this