Paolino, Antonello (2025) Deep Learning for Aerodynamics Modelling and Control of Flying Multi-body Robots. [Tesi di dottorato]

Documento PDF paolino_antonello_38.pdf Visibile a [TBR] Amministratori dell'archivio Download (90MB) | Richiedi una copia

Abstract

In nature, many organisms developed multimodal locomotion as an evolutionary strategy to adapt to diverse environments. This strategy has inspired scientists to develop multimodal robotic systems aspiring to expand their operational capabilities. The strive for innovation in the design of such systems led to the emergence of aerial humanoid robotics, combining bipedal and aerial locomotions in a single platform. Flying humanoid robots present unique challenges in terms of design, modeling, and control, since they are designed to operate in complex non-controlled environments, requiring advanced capabilities to navigate and interact with their surroundings. Aerial humanoid robots are exposed to aerodynamic forces during flight, and the accurate modeling and control of these forces is crucial to ensure stability and performance. This poses significant challenges, as the aerodynamic flow field is highly nonlinear due to the complex shape of the robot, and the strong dependency on its configuration and state. This thesis aims to investigate the challenges and solutions related to the modeling and control of the aerodynamic forces acting on flying humanoid robots during flight. We begin our investigation by measuring the aerodynamic forces acting on the aerial humanoid robot iRonCub-Mk1 with wind tunnel experiments, and we proceed to the validation of reliable RANS CFD simulations for the collection of the distributed aerodynamic forces dataset. We then leverage the aerodynamic forces dataset to train a deep neural network and a linear regression model. These models are designed to predict the aerodynamic forces acting on the robot during flight, enabling real-time control. The trained models are then integrated into a whole-body simulator and into an aerodynamic-aware controller, which is validated through flight simulations and balancing experiments on the iRonCub-Mk1 robot. Finally, we propose a new deep learning approach to predict the surface aerodynamic quantities on flying multi-body robots. We start from the CFD simulations data collected on the iRonCub-Mk3 robot and introduce diverse data manipulation techniques to enable the use of convolutional autoencoders on 3D simulation data. Then, we compare the results of three different learning architectures, namely a convolutional autoencoder, a fully-connected multi-layer perceptron, and a graph convolutional network. The studies presented in this thesis demonstrate that data-driven techniques are an effective solution to model and control the aerodynamic forces acting on flying humanoid robots. The improvements achieved in this research pave the way for future development of aerodynamics-based design and control strategies for the next generation of aerial humanoid robots.

Downloads

Actions (login required)

Modifica documento