Unmodeled and often complex aerodynamics are among the most notable challenges to precise flight control. Especially aerodynamic forces pose a challenge, as they depend on hidden state variables like airflow, which cannot be easily measured. Furthermore, the individual downwash induced by the rotors interacts with both the frame and the blades depending on the current state of the platform [1]. These unsteady and nonlinear aerodynamic effects substantially degrade the performance of conventional UAV control methods
that neglect to account for them in the control design. This aspect is particularly crucial for emerging generic UAV designs, characterized by mutual interference among propellers that can be positioned and oriented in different, even time-varying ways. Prior approaches partially capture these effects with simple linear or quadratic air drag models, which limit
the tracking performance [2]. In such contexts, a data-driven model can be used for prediction. While pure DNN models have been shown to provide competitive predictions, they are often found to be poor at generalizing and extrapolating, especially when trained with small or sparse datasets [3]. Moreover, they do not implicitly guarantee adherence to even basic physics laws and exhibit challenging-to-interpret black-box behaviour. In this thesis project, we propose a multi-rotor dynamics model that can capture complex aerodynamic
effects by integrating a simple partial physics multi-rotor model with a Neural Network that learns the residual dynamics.
Thus the main research question is whether an AI-powered hybrid dynamics model for generic multi-rotor systems demonstrates more accurate and robust predictive capabilities compared to existing approaches, especially regarding mutual aerodynamic interference. To optimize this novel model, the following key areas will be investigated:
• Designing a Neural Network architecture tailored to effectively capture and predict the residual aerodynamics.
• Investigating a range of loss functions and back-propagation techniques that are compatible with physical models or drawing inspiration from physics principles to improve the accuracy and efficiency of the hybrid model.
• Validation of the proposed model through real-world experiments on multi-rotor platforms such as the Omnimorph in order to provide empirical evidence of its performance.
• If feasible within the project timeframe, implementing an offline-online learning scheme, aiming to adapt the model in real-time during physical experiments to enhance it’s adaptability and robustness.
References
Bibliography
[1] L. Bauersfeld, E. Kaufmann, P. Foehn, S. Sun, and D. Scaramuzza, “NeuroBEM: Hybrid Aerodynamic Quadrotor Model,” in Robotics: Science and Systems XVII, Jul. 2021, arXiv:2106.08015 [cs].
[2] A. Saviolo and G. Loianno, “Learning quadrotor dynamics for precise, safe, and agile flight control,” Annual Reviews in Control, vol. 55, pp. 45–60, 2023.
[3] A. Behjat, C. Zeng, R. Rai, I. Matei, D. Doermann, and S. Chowdhury, “A physics-aware learning architecture with input transfer networks for predictive modelling,” Applied Soft Computing, vol. 96, p. 106665, Nov. 2020.