Documenti full-text disponibili:
![[img]](http://amsdottorato.unibo.it/style/images/fileicons/application_pdf.png) |
Documento PDF (English)
- Accesso riservato fino a 31 Dicembre 2024
- Richiede un lettore di PDF come Xpdf o Adobe Acrobat Reader
Disponibile con Licenza: Salvo eventuali più ampie autorizzazioni dell'autore, la tesi può essere liberamente consultata e può essere effettuato il salvataggio e la stampa di una copia per fini strettamente personali di studio, di ricerca e di insegnamento, con espresso divieto di qualunque utilizzo direttamente o indirettamente commerciale. Ogni altro diritto sul materiale è riservato.
Download (30MB)
| Contatta l'autore
|
Abstract
Today, artificial intelligence (AI) is rapidly advancing, enabling increasingly complex capabilities in tiny flying robots.
Tiny AI-driven unmanned aerial vehicles (UAVs) are envisioned to achieve intelligence akin to biological systems like insects. Bees, for example, can pursue multiple goals concurrently with full autonomy. Developing sophisticated skills in miniaturized UAVs holds the potential for wide-ranging applications, significantly impacting many aspects of our lives.
However, the miniaturization of UAVs presents several challenges. Nano-drones, approximately 10cm in diameter, are at the forefront but cannot yet execute multiple intelligence tasks concurrently due to limited size and payload, which restrict them to ultra-low-power (ULP) processors with stringent computational and memory constraints.
This thesis aims to narrow the intelligence gap between tiny flying robots and insects by enabling concurrent execution of multiple real-time AI-based perception tasks on autonomous nano-UAVs.
First, we present methodologies and software tools for automating and optimizing convolutional neural networks (CNNs) deployment on nano-UAVs, adhering to their ULP processor constraints, and we apply our methodology to a CNN for visual autonomous navigation.
Second, we minimize the CNN workload on nano-drones. We identify inactive neurons within the CNN and introduce architecture modifications to shrink the network.
Applying this methodology to a state-of-the-art visual-based autonomous navigation CNN, we achieve a network that is 50x smaller and 8.5x faster than the baseline without compromising performance.
Third, leveraging the freed computational resources, we enable nano-UAVs to perform multiple AI tasks in real-time by deploying a CNN for object detection alongside the visual-based navigation CNN.
Combining techniques for CNN optimization and automated deployment and integrating two CNNs on a ULP processor, we demonstrate the ability to overcome computational and memory constraints, allowing simultaneous execution of multiple AI-based perception tasks on nano-UAVs. This milestone brings tiny flying robots closer to the high-level intelligence and multi-tasking capabilities of biological systems.
Abstract
Today, artificial intelligence (AI) is rapidly advancing, enabling increasingly complex capabilities in tiny flying robots.
Tiny AI-driven unmanned aerial vehicles (UAVs) are envisioned to achieve intelligence akin to biological systems like insects. Bees, for example, can pursue multiple goals concurrently with full autonomy. Developing sophisticated skills in miniaturized UAVs holds the potential for wide-ranging applications, significantly impacting many aspects of our lives.
However, the miniaturization of UAVs presents several challenges. Nano-drones, approximately 10cm in diameter, are at the forefront but cannot yet execute multiple intelligence tasks concurrently due to limited size and payload, which restrict them to ultra-low-power (ULP) processors with stringent computational and memory constraints.
This thesis aims to narrow the intelligence gap between tiny flying robots and insects by enabling concurrent execution of multiple real-time AI-based perception tasks on autonomous nano-UAVs.
First, we present methodologies and software tools for automating and optimizing convolutional neural networks (CNNs) deployment on nano-UAVs, adhering to their ULP processor constraints, and we apply our methodology to a CNN for visual autonomous navigation.
Second, we minimize the CNN workload on nano-drones. We identify inactive neurons within the CNN and introduce architecture modifications to shrink the network.
Applying this methodology to a state-of-the-art visual-based autonomous navigation CNN, we achieve a network that is 50x smaller and 8.5x faster than the baseline without compromising performance.
Third, leveraging the freed computational resources, we enable nano-UAVs to perform multiple AI tasks in real-time by deploying a CNN for object detection alongside the visual-based navigation CNN.
Combining techniques for CNN optimization and automated deployment and integrating two CNNs on a ULP processor, we demonstrate the ability to overcome computational and memory constraints, allowing simultaneous execution of multiple AI-based perception tasks on nano-UAVs. This milestone brings tiny flying robots closer to the high-level intelligence and multi-tasking capabilities of biological systems.
Tipologia del documento
Tesi di dottorato
Autore
Lamberti, Lorenzo
Supervisore
Co-supervisore
Dottorato di ricerca
Ciclo
36
Coordinatore
Settore disciplinare
Settore concorsuale
Parole chiave
Autonomous Nano-UAV, Artificial Intelligence, Embedded systems, Ultra-low-power, Energy-efficiency.
URN:NBN
Data di discussione
9 Luglio 2024
URI
Altri metadati
Tipologia del documento
Tesi di dottorato
Autore
Lamberti, Lorenzo
Supervisore
Co-supervisore
Dottorato di ricerca
Ciclo
36
Coordinatore
Settore disciplinare
Settore concorsuale
Parole chiave
Autonomous Nano-UAV, Artificial Intelligence, Embedded systems, Ultra-low-power, Energy-efficiency.
URN:NBN
Data di discussione
9 Luglio 2024
URI
Gestione del documento: