Deep Learning on Ultra-low-power Autonomous Flying Nano-drones
keywords ARTIFICIAL INTELLIGENCE, CONVOLUTIONAL NEURAL NETWORKS, DEEP LEARNING, DEEP NEURAL NETWORKS, DRONES, EMBEDDED SYSTEMS, ENERGY EFFICIENCY, FIRMWARE DEVELOPMENT, LOW POWER, MICROCONTROLLERS, ROBOTICS, SOFTWARE, UAV
Reference persons DANIELE JAHIER PAGLIARI
External reference persons Dr. Alessio Burrello (UniversitÓ di Bologna), Dr. Daniele Palossi (SUPSI-USI, UniversitÓ della Svizzera Italiana)
Thesis type EMBEDDED SOFTWARE DEVELOPMENT, EXPERIMENTAL, SOFTWARE DEVELOPMENT
Description Nano-sized unmanned aerial vehicles (UAVs) are ultra-low-power devices, often organized in swarms, that require on-board "intelligence" to carry out various recognition and control tasks based on data from optical cameras. The state-of-the-art algorithms for such type of task are deep neural networks. However, those models are typically very computationally intensive and memory-hungry, which prevents their deployment on nano-drones, for which the energy-efficiency of the processing is the number one requirement.
To tackle this gap, the thesis candidate will explore different optimization techniques for deep learning models, oriented at balancing the trade-off between accuracy and efficiency, i.e., at obtaining a highly accurate model while minimizing the number of operations and bytes of memory required for its execution on the drone's hardware. The explored techniques will include: Efficient Neural Architecture Search (NAS), Quantization and Mixed-Precision Search (MiPS), and run-time adaptive/dynamic inference. The efficacy of these techniques will be then evaluated on a real working system (in a lab in Lugano), targeting tasks such as drone-to-human and drone-to-drone pose estimation. The target hardware platform will be the ultra-low-power multi-core IoT end-node GAP8.
The thesis can be carried out fully from Torino (leaving in-field tests to our external collaborators), or partly in Lugano (allowing the candidate to run in-field tests and directly touch the results of the work).
Required skills Required skills include C and Python programming. Further, a basic knowledge of computer architectures and embedded systems is necessary. Required skills also include some familiarity with basic machine/deep learning concepts and corresponding models.
Notes Thesis in collaboration with Dalle Molle Institute for Artificial Intelligence (IDSIA) SUPSI-USI, Lugano, Switzerland. The thesis can be carried out fully from Torino (leaving in-field tests to our external collaborators), or partly in Lugano (allowing the candidate to run in-field tests and directly touch the results of the work).
Deadline 13/06/2023 PROPONI LA TUA CANDIDATURA