Open access
Author
Date
2019Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
Whether we like it or not, flying robots are rapidly evolving, getting ever smaller. Pioneering research groups have achieved fully-functional insect-scale unmanned aerial vehicles – i.e., pico-size UAVs. This new wave of innovation has huge economic and social game-changing potentialities, enabling a vast number of visionary applications. However, there are still several challenges preventing us from entering into the insect-size intelligent robot era.
In this thesis, we aim at getting beyond, what we call, the “wall of onboard intelligence”. Nowadays, relatively big-sized drones (i.e., standard- micro-size UAVs) have been proven capable of impressive
autonomous sense-and-act capabilities, but this intelligence comes at the price of onboard bulky and power-hungry computational units. These high-end devices, capable of a tremendous peak throughput of tens of TOps/s within tens of Watts, are unaffordable for small-size UAVs, as their power budget and available payload are severely limited. To date, this “wall of onboard intelligence” stands at the nano-size class of vehicle, characterized by a mass of few tens of grams, and a computational power budget in the range of a few hundred mW. Therefore, this thesis tries to answer the fundamental question of how to bring state-of-the-art autonomous navigation capabilities aboard nano- (and potentially pico-) size UAVs.
The thesis starts investigating on the hardware/software key enablers for high-level energy efficiency and cutting down the computational requirements of navigation workloads, without compromising the overall quality of results. As a domain-specific example application, we address a global path planner. Our findings demonstrate that i) heterogeneous hardware architectures, ii) the parallel computing paradigm, and iii) approximate computing techniques can enable more
than an order-of-magnitude improvement on energy efficiency. Additionally, we show that the introduction of approximate computing in a cyber-physical system (CPS), under strict real-time constraints, implies a “semantic shift” of the same paradigm. Traditionally, approximate computing trade-off numerical accuracy in favor of higher energy efficiency; in this context, the gain, due to the “simplified” computation, can also turn in a higher quality of the mission (e.g., reduced response time and increased reactivity).
The second part of the thesis branches our energy investigation to focus on the system level, aiming at lifetime maximization of nano-size UAVs. Selecting a nano-blimp platform, consisting of a helium
balloon and a rotorcraft, we show how the lifetime of the CPS can be significantly extended by i) dynamically adapting the motor control, ii) duty-cycling high power actuators, and iii) adding solar harvesting. According to our study, even in modest lighting conditions, the ultimate limiting factor for a self-sustainable nano-blimp is the balloon’s deflation rate.
In the third part of the thesis, we explore the applicability of the heterogeneous model in the ultra-low power (ULP) domain – i.e., representative of nano- pico-size UAVs power budget. We demonstrate the feasibility of coupling a low power Cortex-M microcontroller with a programmable parallel ULP accelerator (PULP) for speeding-up computation-intensive algorithms at a milliwatt-scale budget. To further validate the proposed heterogeneous model, we extend our analysis to two concrete example applications representative of the autonomous navigation domain: a visual servoing and a visual odometry pipeline. Our evaluation shows that a speedup of more than an order-of-magnitude is achievable vs. single-core MCUs in a ULP setting, without compromising the platform’s programmability, guaranteeing real-time performance, and leaving enough computational bandwidth to execute additional onboard tasks.
The last part of the thesis is devoted to combining all the previous basic building blocks with a novel bio-inspired class of autonomous navigation algorithms. Leveraging the DroNet convolutional neural
network (CNN), we developed a complete deployment methodology targeted at enabling the execution of complex CNNs directly aboard resource-constrained milliwatt-scale nodes. We achieve energy-efficient real-time processing of DroNet on board a nano-UAV, with a peak throughput of ∼1.5 GOps/s, and energy efficiency up to ∼7.7 GOps/s/W. Our field-proven, closed-loop, autonomous nano-UAV, can follow a street/corridor, being at the same time capable of preventing collisions with dynamic obstacles. We answer our initial question creating – to the best of our knowledge – the first example of a new intelligent nano-size species of UAVs. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000400028Publication status
publishedExternal links
Search print copy at ETH Library
Publisher
ETH ZurichSubject
bio-inspired; artificial intelligence; Unmanned aerial vehicles (UAVs); Convolutional neural network (CNN); deep learning; Nano-UAVs; Autonomous UAVs; energy efficiency; path planning; Approximate computing; Heterogeneous hardware architectures; parallel computing; cyber-physical systems (CPS); nano-blimp; self-sustainability; ultra-low power architectures; visual servoing; visual odometryOrganisational unit
03996 - Benini, Luca / Benini, Luca
More
Show all metadata
ETH Bibliography
yes
Altmetrics