Any problem becomes tractable with enough power. Since the inception of wireless sensornetworks, researchers have searched for ways to do more with less. Integrated circuits and sensors have continued to shrink in size, cost, and active and quiescent power. This has resulted in sensors with increasing computational power and longer lifetimes. By comparison, however, the options for and quantity of power available for a wireless sensor has stagnated. The energy density of non-rechargeable batteries as well as photovoltaic efficiency have approached a plateau. As a result, wireless sensors are constrained in either lifetime or power.

Technology improvements like integrated circuits that perform efficient maximum powertracking and voltage boosting, near-threshold computing, the advent of non-volatile memory technologies, and the rapid improvement of supercapacitor technology has enabled the development of sensors that can operate entirely on harvested power without batteries, and without a finite lifetime. But the lack of a reliable power source necessarily results in a system that is fundamentally unreliable. Despite this significant flaw, researchers have pursued this design archetype tirelessly, producing an impressive corpus of methods, systems, and solutions that attempt to improve batteryless design. Proponents of batteryless systems are convinced that batteries are a threat to the future of wireless sensing, and that batteryless sensing is the only way forward. Despite this, batteryless sensing has not seen widespread adoption by industry. There is a rift of design understanding between those who value reliability, and those who do not.

The core argument of this dissertation is that there is not a single design dogma, be itbatteryless or battery-powered, that can provide a solution for all applications. Instead, the correct design process must involve a balancing act of the inclusion and sizing of energy harvesting, rechargeable, and non-rechargeable energy storage to meet the goals of the application. This design space is large and difficult to navigate, resulting in many system designers defaulting to following a predetermined design template archetype instead of fully reasoning about their application and its requirements. In this dissertation, we develop a design framework for energy harvesting systems that provides reasoned guidance for the inclusion and sizing of various power supply elements. In particular, we develop analytical and simulation tools to size rechargeable energy capacity in a more reasoned way than current heuristics and arbitrary methods.

To develop this design framework, this dissertation explores previous wireless sensor applications, identifying the appropriateness of different approaches qualitatively and quantitatively.We explore the system-level effects of harvester size and rechargeable and non-rechargeable energy capacity on wireless sensor application performance. To determine rechargeable energy capacity selection and sizing, we develop a novel heuristic for determining the minimum sufficient capacity for a sensor workload and expected energy income. We verify this heuristic through the use of a custom wireless sensor energy state simulator to estimate energy utilization and system performance. To identify technology options for energy capacity, we quantitatively compare energy buffer types and reevaluate the many qualitative claims made against rechargeable batteries by batteryless proponents, concluding that many of them are without merit. Finally, we utilize the design framework developed within this dissertation, including the heuristics and simulation tool, to design and implement wireless sensor systems to address two real indoor sensing applications that achieve long-lived operation with consistent and reliable sensing

The Internet of Things (IoT) encompasses a broad array of technologies that connect the physical world with large-scale data processing and storage. Today, we can build ultra-low power devices that last for decades on ambient energy and we can deploy highly scalable internet services to process streams of IoT data. The dominant challenge of IoT lies in connecting these two domains. To meet this challenge, many new networks have been developed with low-energy, low-throughput, and predominantly uplink use cases in mind.

In this work, we explore the capabilities and limitations of several recent IoT-focused networks. While these novel networks are not yet deployed at scale, modeling aspects of them enables us to predict the success or failure of particular applications and explore the ramifications of potential protocol improvements.

In the local-area domain, Bluetooth Low Energy (BLE) has arisen as an ubiquitous method of connecting deployed devices directly to users' smartphones. One communication mechanism that BLE provides is the advertisement a simple, periodic, broadcast message intended for device discovery. With advertisements, we can create a single-hop, star-topology network in full compliance with the BLE specification that allows any number of devices to send data to any number of gateways. By modeling advertisement transmissions, we can accurately predict congestion in advertisement networks, which enables an a priori understanding of performance and an in-situ adaptation mechanism to meet reliability expectations.

In the wide-area domain, multiple low-power, wide-area networks (LPWANs) have arisen to enable city-scale deployments. Their ability to transmit at ranges over a kilometer while drawing only a few hundred milliwatts could enable exciting new applications. To understand LPWAN capabilities, we propose a new metric, bit flux, which describes communication in terms of throughput over a coverage area. By using bit flux to model the performance of networks and the needs of applications, we demonstrate problems with the existing designs of several LPWANs that make them unsuitable for many real-world deployments.

In both domains, we explore the strengths and weaknesses of communication technologies and their potential to serve real-world applications. The potential for these networks to promote a new generation of easy-to-deploy sensors is high. However, through the application of communication models, we demonstrate both potential use cases and also very real concerns that may limit them. By uncovering concerns that future deployments will face, we hope to guide improvements to these protocols that will improve their use and support the growth of the Internet of Things.

Contemporary tools for the design of embedded systems, task-specific electronic devices, arebuilt on a paradigm that has not fundamentally changed since the era of pen-and-paper drafting despite new computational tools that enable new, better workflows for designers and engineers. Embedded systems are the glue we use to connect the digital and physical worlds, letting us leverage the connectivity and automation of computers to solve problems from home automation to battlefield awareness. More user friendly tooling would reduce skill requirements, speed up design cycles, and allow more people to solve their problems with embedded systems. Contemporary Electronic Design Automation (EDA) tools are designed around a single step in the design process, board layout, where an electrical schematic is turned into a design for a printed circuit board, the copper and fiberglass base that connects all the other components in a system. However, engineers go through a series of phases before they reach board layout: exploring the problem they are trying to solve, sketching out a high-level system architecture, and refining that into a well-defined electrical circuit. Better EDA tools would fit more naturally into this workflow, existing to support users in earlier phases, presenting them with information as it becomes relevant, and automating routine or repetitive work. This dissertation describes two such tools, both built by formulating the design process in mathematical terms and using algorithms to reason about our formalisms, all while wrapped in user-friendly interfaces. The first tool, Embedded Design Generation (EDG) is a proof-of-concept system meant to push the limits of automation in the embedded design process. Given a high-level specification for a device it uses satisfiability solvers to synthesize, from whole cloth, a design meeting that spec. The second, Polymorphic Blocks, uses block diagrams to represent designs in arbitrary stages of construction and propagator semantics for error checking, predictive suggestions, and other features.

Detailed contact tracing that not only captures the social interaction graph, but also precise interaction distance and duration could prove useful in a wide variety of applications. Most notably, we have seen this play out in the global COVID-19 pandemic, where social distancing and contact tracing have proven critical in efforts to combat disease spread. Traditionally, contact tracing has relied on manual reporting, which provides only coarse grained data and relies on the subjectivity of human memory. These factors have led to a drive for wearable sensor based solutions which can provide objective face-to-face interaction data. Ideally, these sensors would provide precise interaction distances and durations, and would only report these metrics when users are actually facing each other and are not separated by a barrier. Current contact tracing sensors can generally be divided into two camps. First are sensors that can provide precise interaction distances, but require infrastructure to run, making them difficult to deploy in practice. Second are sensors that do not require infrastructure, but only provide a rough sense of proximity, making it difficult to analyze which interactions are significant. The majority of these systems also cannot determine if there is a barrier separating users, or if the users are facing each other.

To address these issues, we present Opo, a wearable sensor which requires no infrastructure to run, provides interaction distance accurate to 5~cm, and only records interaction distances when users are facing each other with no barriers between them. The key problem we identify is that systems that provide precise interaction distances require RF based neighbor discovery protocols to synchronize nodes before performing ranging operations to get interaction distance. Instead, Opo utilizes ultrasonic passive vigilance, to perform neighbor discovery and ranging at the same time, lowering system complexity and power usage.

In addition, while current wearables for contact tracing have largely focused on detecting interactions, in practice this information is greatly enriched by knowledge of health behaviors and symptoms. For example, researchers are often interested in detecting hand-washing behavior due to its importance in combating a wide variety of infectious diseases. Current hand washing sensor systems generally use wrist mounted accelerometers or utilize smart badges and soap-dispenser mounted sensors. Of these two system types, only smart badge plus soap dispenser sensor systems are able to capture if a person uses soap, a key consideration when measuring hand washing behavior. However, smart badge systems only detect when a person washes their hands with soap, and do not sense hand washing duration. The key problem is that current smart badge systems use low-resolution ranging technologies, making it difficult for them to determine when a user approaches leaves a sink. To address this problem, we create a smart badge plus dispenser mounted sensor system by extending Opo with passive vigilance in the accelerometric domain. This extension allows soap dispenser mounted Opos to passively detect when a dispenser is used and provide precise times when a user approaches and leaves a sink. To the best of our knowledge, our hand washing system is the first that can detect and categorize both soaped and un-soaped hand washing events and measure hand washing duration.

Researchers are also often interested in when people first experience symptom onsets. In particular, researchers are often interested in when people begin coughing, due to its prominence as an early symptom in many infectious diseases. Current cough sensing systems focus on counting the number of times a person coughs over a given period of time. These sensors require a user to wear a voice recorder and record all of their audio over the period of time. These systems then identify and count coughs in post-processing. This technique has shown very promising results, but requires a massive invasion of user privacy, making them difficult to deploy in many situations. In addition, our review of prior work on coughing shows that in many applications, simply knowing when a person starts coughing or general trends in a person's cough counts provides significant value. To fill this niche, we create CoughNote, a wearable privacy preserving cough sensor. Instead of constantly recording audio, CoughNote utilizes passive vigilance in the audio domain to capture 1~s snippets of potential coughs, while avoiding recording sensitive vocalized audio such as speech. These potential coughs can then be analyzed in post processing without violating user privacy. Although CoughNote does not capture every cough, it can show general cough trends while preserving usability and being smaller, lighter, and almost three times as long lived as a typical voice recorder.

Overall, our work creates a wearable sensing kit that researchers can use to study face-to-face interactions and important contextual health information. We have conducted two pilot studies using Opos and CoughNote with epidemiologists, and hope that our sensors enable future gains in better understanding and forecasting disease spread. Furthermore, our work shows the power of using passive vigilance to create complex, high-resolution wearables, and we hope that future wearable sensor designers draw inspiration from our designs.

At its inception, the Internet of Things (IoT) was predicted to connect trillions of inexpensive smart devices to the Internet in mere decades, embedding paradigm-shifting utility, communication, and convenience into everyday things. Since then, a number of connected speakers, thermostats, doorbells, lights, and wearables have entered the market and our homes, but devices and sensors further constrained by processing power, memory, energy, and networking capabilities struggle to find their way to market. This is due, in part, to the prevailing model in industry of requiring a different application-layer gateway — often in the form of a device-specific router box or smartphone app — to link each unique IoT device to the Internet and allow users to interact with it. Unfortunately, this siloed approach cannot sustain the true scale, density, and capability of the IoT vision.

To break free from such siloed gateway standards which restrict scalability, affordability, and accessibility for constrained IoT devices, we propose a set of application-agnostic approaches to connectivity and interaction. For connectivity, we design a general-purpose network architecture consisting of ubiquitous short-range gateways that openly facilitate data transport for constrained devices to the Internet. By simplifying the gateway to its most essential parts, communications and processing, we can implement the architecture on inexpensive off-the-shelf system-on-chips and on existing smartphone infrastructure. These static and mobile approaches provide coverage for both reliable data throughput scenarios and ephemeral connectivity needs. For interaction, we design a browsing architecture that provides a seamless, scalable approach to discovering and interacting with nearby Things. The system takes advantage of multiple network patterns and modern web technologies to automatically detect devices and supply users with rich device interfaces that can enable interaction directly over local networking protocols. We implement and deploy a mobile-based browsing platform to facilitate this and provide tools for device developers. We further explore techniques for more physically-tied interaction, including extension to mixed reality.

The proposed approach to IoT connectivity and interaction opens up exciting possibilities for the future of the Internet of Things. By breaking down the barriers imposed by siloed gateway standards and introducing a set of approaches that form a general-purpose alternative, this work paves the way for the creation of more innovative and versatile IoT devices and interfaces. The design principles presented provide a simple, yet dynamic framework for enabling Internet connectivity for IoT devices, and for users to interact with them in a more intuitive and seamless way. These architectures represent significant steps forward in realizing the full potential of the Internet of Things, making it more accessible to everyone, and further advancing the development of the smart and connected world of the future.

Advancements across the computing stack now allow the realization of self-contained, sensing, computing, and communication devices in less than a cubic millimeter of volume—Smart Dust. Today’s millimeter-scale systems are often one-offs, however. They are built to prove they can be built or for one particular application, but they are not yet so numerous and diverse to satisfy the vision of ubiquitous, ambient intelligence. This work aims to identify and address challenges in the move from tens to tens of millions of Smart Dust computing systems.

The core contribution of this dissertation is bringing modularity to the millimeter-scale computing class. Much of the rich panoply of conventional systems is driven not by the diversity of individual chips alone but by the number of different ways in which they can be synthesized into novel designs. Composition-oriented design of systems is enabled in part by the interconnect technologies that lie between components. This dissertation identifies the limitations of extant interconnect technologies for the millimeter-scale computing class and then designs, implements, and evaluates a new interconnect bus and interconnect protocol that enables composable Smart Dust systems. As a final contribution, this dissertation shows how this same interconnect can be leveraged to support key operations for bringup and deployment of Smart Dust systems, in particular programming and debugging.

The system and network architecture for stationary sensornets is

largely solved today with many commercial solutions now available and

standardization efforts underway at the IEEE, IETF, ISA, and within

many industry groups. However, the existing techniques for reliable,

low-power communications in stationary sensornets fail on both counts

when confronted with mobility. In this dissertation, we argue that

awareness of real or potential mobility enables a solution that

handles the mobile case well, and supports stationary networks as a

special case. This dissertation addresses micropower mobiscopes, a

nascent class of mobile sensornets -- small, embedded, and

battery-powered systems -- that experience unpredictable but

structured mobility and are severely energy-constrained. We show how

awareness of mobility can simplify their communication challenges,

enable low-power operation, and enhance the reliability of data

delivery.

We introduce the MOV metric, a measure of mobility, and present

techniques to gather it on a near nano-power budget. We also present

iCount, a regulator-integrated energy meter design that allows nodes

to introspect their own energy usage, and adapt their behavior to the

actual energy availability and consumption. Integrating the pieces,

we present three concrete hardware platforms that support our mobile

sensing architecture. We develop a novel asynchronous neighbor

discovery algorithm called Disco that allows nodes to operate their

radios at very low duty cycles and yet still discover neighbors

without any external synchronization information. Recognizing the

necessity of beaconing in mobile networks, and the need for

mobile-stationary node interactions, we design a link layer

synchronization primitive, Backcast, and a receiver-initiated link

layer, HotMac, that are suitable for mobile sensing, but also work

for stationary networks across a range of conventional data collection

workloads and a broad range of duty cycles.

We evaluate our thesis with three mobile sensing applications that

embody our proposed architecture. The three applications --

AutoWitness, SleepTrack, and CommonSense -- are representative of

asset tracking, health and fitness, and participatory urban sensing,

and they each stress different aspects of the architecture, including

motion detection, neighbor discovery, communications, interaction

patterns, energy management, and data transport. These design points

illustrate that our architecture is general enough to enable a range

of applications but specific enough to support them well.

\begin{abstract}Electricity is increasingly considered to be a human right. However, in much of the world, basic data about power outages and voltage quality is not available, making it difficult to ensure reliable service. Smart meters are difficult to deploy and slow to be adopted in lower-resourced areas, the same areas where frequent outages and voltage sags most impede economic growth.

In this work, I present evidence that by aggregating simple measurements from networked sensors installed at outlets in households and businesses, we can detect both large and small power outages and power-quality issues, enabling a utility-independent, agile, high-resolution, and low-cost system well suited for deployment in under-instrumented areas.

I design, deploy, and operate a large network of sensors---called PowerWatch---at outlets in households and businesses across Accra, Ghana. I show that this deployment, when coupled with cloud-based analytics, matches utility-reported rates of high- and medium-voltage outages at a fraction of the cost. Further this methodology provides a good estimate of low-voltage outages, potentially filling a critical data gap present in most countries.

The deployment methodology developed allows for longitudinal data to be gathered independently from the utility. Utility engineers are not required for sensor installation, and permission is not required as utility property is not impacted. I describe a number of novel measurement and analysis opportunities enabled by this independent deployment methodology, many of which are being piloted by \textit{n}Line, the company I co-founded to continue improving this work.

Data from PowerWatch is being used by multiple governments and research institutions, including as a primary source for the Monitoring and Evaluation of the \$315 Million USD MCC led Ghana Power Compact. \end{abstract}

Cloud computing revolutionized the ease with which we can build, deploy, and scale distributed computing services. These advances, however, have not extended to the physically distributed and resource-constrained computers deployed throughout the world to collect data, and their resource constraints have thus far confined them to function as inefficient, fixed-purpose data forwarders. Breaking these distributed sensors free of their resource-constraints by including them in a dynamic, programmable, distributed system will not only enable easier deployment and scaling of applications relying on their data, but it will also give us the ability to collect and process never-before-seen data and discover new ways sensing the world around us.

We enable this vision in two parts. First we present a signpost-based platform which eases the building and deployment of sensors by providing the core services and hardware necessary for them to function. Next we explore the benefits of, and build a resource manager to form these resource-constrained sensors into a compute cluster akin to those found in the cloud. This enables multiple users to simultaneous program a cluster of sensors and quickly iterate on their programs through an application framework which abstracts away the details of scheduling and task distribution. By forming these sensors into a multiprogrammable cluster, we enable them to be accessed as a shared sensing utility rather than as a collection of individual nodes.