Accurate modeling of quadrotor flight dynamics is a prerequisite for high-performance model-based control. Classical first-principles models derived from the Newton-Euler rigid-body equations efficiently capture the dominant dynamics but systematically neglect aerodynamic effects such as rotor drag and blade flapping, accumulating prediction errors during open-loop rollouts. Purely data-driven neural networks can learn these effects from flight data, but are prone to overfitting in small-dataset regimes and provide no guarantees of physical consistency. This thesis proposes a Physics-Informed Machine Learning (PIML) framework for quadrotor dynamics identification. A Multilayer Perceptron is trained with a composite loss combining a data loss — penalizing next-state prediction errors after Euler integration — and a physics loss — penalizing deviations from the Newton-Euler equations. An adaptive weighting strategy balances the two terms throughout training. The framework is validated in simulation (PX4 SITL + Gazebo, x500 quadrotor) on two datasets with different excitation ranges, comparing three configurations: a data-driven baseline, a physics-informed model with adaptive weighting, and one with the ConFIG gradient method. The physics-informed model with adaptive weighting consistently outperforms the baseline on multi-step open-loop rollouts, within the training distribution and on an unseen circular trajectory. The physics loss prevents overfitting even with a larger network and anchors predictions to physically consistent solutions. The ConFIG method fails to maintain stable joint convergence. Finally, dataset diversity is the primary driver of out-of-distribution generalization: the restricted-range model is more accurate in-distribution, while the extended-range model generalizes better to unseen conditions.
La modellazione accurata della dinamica di volo di un quadricottero è un prerequisito per il controllo ad alte prestazioni. I modelli classici derivati dai principi fisici, basati sulle equazioni di Newton-Eulero, descrivono efficientemente le dinamiche dominanti ma trascurano sistematicamente effetti aerodinamici quali il drag dei rotori e il flapping delle pale, accumulando errori di predizione durante la simulazione di traiettorie. I modelli puramente data-driven possono apprendere questi effetti dai dati di volo, ma sono soggetti a overfitting con dataset limitati e non offrono garanzie di consistenza fisica. Questa tesi propone un approccio per il Physics-Informed Machine Learning (PIML) per l'identificazione della dinamica di un quadricottero. Un Multilayer Perceptron è addestrato con una funzione di costo composita che combina un costo sui dati — che penalizza gli errori di predizione dello stato successivo dopo integrazione di Eulero — e un costo fisico — che penalizza le deviazioni dalle equazioni di Newton-Eulero. Una strategia di pesatura adattativa bilancia i due termini durante il training. L'approccio è validato in simulazione (PX4 SITL + Gazebo, quadricottero x500) su due dataset con diversi range di eccitazione, confrontando tre configurazioni: un baseline data-driven, un modello physics-informed con pesatura adattiva, e uno con il metodo ConFIG. Il modello physics-informed con pesatura adattiva supera costantemente le prestazioni offerte dagli altri metodi. La cifra di costo fisica previene il fenomeno dell'overfitting anche con una rete più ampia e porta le predizioni ad essere fisicamente consistenti. Il metodo ConFIG non mantiene una convergenza congiunta stabile. Infine, la varietà del dataset è particolarmente rilevante per la generalizzabilità del modello in condizioni outof-distribution: il modello a range ristretto è più accurato in-distribution, mentre quello a range esteso generalizza meglio a condizioni non visitate.
Physics-informed machine learning for quadcopter dynamics identification
VILLA, LORENZO
2025/2026
Abstract
Accurate modeling of quadrotor flight dynamics is a prerequisite for high-performance model-based control. Classical first-principles models derived from the Newton-Euler rigid-body equations efficiently capture the dominant dynamics but systematically neglect aerodynamic effects such as rotor drag and blade flapping, accumulating prediction errors during open-loop rollouts. Purely data-driven neural networks can learn these effects from flight data, but are prone to overfitting in small-dataset regimes and provide no guarantees of physical consistency. This thesis proposes a Physics-Informed Machine Learning (PIML) framework for quadrotor dynamics identification. A Multilayer Perceptron is trained with a composite loss combining a data loss — penalizing next-state prediction errors after Euler integration — and a physics loss — penalizing deviations from the Newton-Euler equations. An adaptive weighting strategy balances the two terms throughout training. The framework is validated in simulation (PX4 SITL + Gazebo, x500 quadrotor) on two datasets with different excitation ranges, comparing three configurations: a data-driven baseline, a physics-informed model with adaptive weighting, and one with the ConFIG gradient method. The physics-informed model with adaptive weighting consistently outperforms the baseline on multi-step open-loop rollouts, within the training distribution and on an unseen circular trajectory. The physics loss prevents overfitting even with a larger network and anchors predictions to physically consistent solutions. The ConFIG method fails to maintain stable joint convergence. Finally, dataset diversity is the primary driver of out-of-distribution generalization: the restricted-range model is more accurate in-distribution, while the extended-range model generalizes better to unseen conditions.| File | Dimensione | Formato | |
|---|---|---|---|
|
executive_summary.pdf
non accessibile
Dimensione
417.14 kB
Formato
Adobe PDF
|
417.14 kB | Adobe PDF | Visualizza/Apri |
|
Thesis.pdf
accessibile in internet per tutti
Dimensione
3.45 MB
Formato
Adobe PDF
|
3.45 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/261417