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

Abstract

Ground effect (GE) causes multirotor unmanned aerial vehicles (UAVs) to experience unmodeled thrust augmentation when flying close to irregular terrain, producing altitude oscillation and control degradation that classical flat-ground analytical models (e.g., Cheeseman–Bennett) cannot capture. We built a physics-based simulator with procedurally generated irregular terrain and trained machine learning (ML) models (Random Forest, Gradient Boosting, LSTM, Transformer, and MLP) to predict GE factor from onboard-observable sensors alone. Offline, all ML models substantially outperform the analytical baseline under a seed-wise train/validation/test split that guarantees no terrain seed leaks across splits (best model R2 = 0.998 vs. R2 = 0.851 for analytical on 30 held-out seeds), confirming that the offline advantage is not an artifact of seed leakage. In closed-loop evaluation across 30 unseen terrain seeds, however, a deployed hybrid MLP controller remained consistently, if modestly, worse than the pure analytical baseline, and — critically — an oracle controller given the simulator’s exact ground-truth GE factor performed no better than the analytical baseline either. This indicates the shortfall is not primarily a prediction-accuracy problem. Using the corrected actuator mapping (1/√f̂, since thrust is quadratic in throttle-proportional rotor speed), a controlled bias-injection experiment confirms a genuine causal effect: injecting a positive (overprediction) GE-factor bias significantly degrades safety-relevant clearance margins relative to an equivalent negative bias (p < 10−5 on 5th-percentile clearance and low-clearance duration), even though its effect on aggregate AGL tracking RMS is small and in the opposite direction. A time-limited bias pulse and closed-loop pole analysis further show that this effect does not compound into a self-reinforcing feedback loop: the system’s poles have negative real parts across the operating range of GE factors, and the observed response is an underdamped but stable return to baseline rather than divergence. An asymmetric clamp on GE-factor overprediction, tuned only on validation seeds, closes most of the gap to the analytical baseline. Across nine slope/curvature correction configurations (1,620 total runs), the hybrid-versus-analytical gap is stable at approximately 0.0010 m. These results show that offline predictive accuracy is not a reliable proxy for closed-loop performance, that a plausible correlation-based mechanistic story requires direct causal and dynamical verification before being reported as established, and that even a perfectly accurate predictor does not guarantee closed-loop improvement when actuator-level effects are not separately accounted for.