150813 research outputs found
Sort by
Underwater Semantic Simultaneous Localization and Mapping
No full textBuilding semantically meaningful object level maps of underwater environments is crucial for enabling higher-level autonomy, fostering human-robot collaboration, and providing compressed map representations for bandwidth-constrained underwater communications, while localizing against such maps can improve the positioning accuracy of underwater vehicles by correcting for odometric drift. However, underwater semantic simultaneous localization and mapping (SLAM) has lagged behind analogous terrestrial and aerial semantic SLAM techniques largely due to the lack of large labeled underwater datasets and the challenging sensor modalities specific to underwater environments. To address these shortcomings, this thesis develops a range of methodologies to advance underwater semantic SLAM capabilities.
First, self-supervised learning and visual foundation models are leveraged to detect and segment underwater objects in an open-set manner, i.e., objects need not be present in the training dataset to be detected. The machinery of the open-set object detection technique breaks several assumptions made by existing closed-set semantic SLAM methods. Thus, new methods for object representation and data association are proposed and demonstrated. A method to localize underwater objects is then developed through an analysis of the geometry of underwater monocular cameras and multibeam sonars.
Finally, a formulation of open-set object-level place recognition as a graph matching problem is introduced. The formulation includes a method for calculating and tracking semantic uncertainty for open-set object detections. Experimental results on both underwater and terrestrial datasets demonstrate that the proposed formulation can be used for real-time accurate open-set object-based place recognition.
In summary, techniques for underwater object detection, localization, and data association are introduced and integrated with probabilistic graphical models for open-set semantic SLAM. The proposed techniques are tested across a wide variety of scenarios, and are shown to generalize to terrestrial settings as well.Ph.D
OPTASAT: An Open-Source, Flexible Software Framework for Small Satellite Operations
No full textThe unprecedented growth in access to space has created a corresponding growth in the number of spacecraft and the number of people operating spacecraft. This has meant that many of these operators are operating spacecraft for the first time. Gone are the days when the only operators of spacecraft were national governments, militaries, and massive corporations. The operators of small spacecraft today include many early-career individuals who need the tools to enable them to make strong decisions in the behavior of their spacecraft. The tools for operating spacecraft are often overlooked by teams focusing on the spacecraft themselves, but these operating tools are critical for mission success. Spacecraft operations tools have not developed in a similarly low-cost, widespread fashion as the spacecraft have. The best tools for modeling and understanding the situation of a satellite in space remain locked behind high barriers to entry including high cost, long training, and complex interfaces. In the same way that satellites have gone from the size of automobiles to the size of toasters, the software for operating them needs to go from expensive, complicated, high-performing suites to simple, flexible, approachable options that are accessible to the democratized space operators. New spacecraft operations staff need straightforward, direct interfaces which give them the knowledge of where their spacecraft is, where it will be, and what it will be able to do, and they need to know when all the options at their disposal are viable. Operators also need to be given the capability to adjust their software in whatever ways are necessary to tailor it to the particular parameters of their missions, to reflect the incredible variety of spacecraft and missions that exist today. A gap exists in spaceflight software. Users need software that can perform their mission planning tasks in the short term and to inform them of the upcoming parameters of their spacecraft which concern them, whether this is the spacecraft’s location, solar illumination, orientation, or any other property which is relevant to their particular mission. This software must also allow the users to be aware of the expected output of their sensors, especially imaging sensors, such that they may have an understanding of what they are imaging and what it ought to look like. Finally, this software must be open-source, enabling the user to audit the software and make changes to the software to customize it to their preferences, which may differ from anything the original software developer could have imagined. Such spaceflight software does not yet exist. This dissertation develops and presents OPTASAT, the Open-source Python Tool for Awareness of Spacecraft and Analysis of Telemetry, which provides an extensible, modular interface for incorporation of multiple tools which contextualize spacecraft data in a manner which maximizes usefulness for the operators. A priority is visualization of data to facilitate rapid understanding and distillation of the complexity of a spaceflight operation. This software has been released as a fully-featured, open-source software toolkit which performs the mission analysis components deemed most crucial to those who stand to benefit from it. This software is intended to fulfill the needs of small spacecraft missions. Several particular application cases are studied, including that of an Earth Sensing mission, and Astronomy mission, and modeling communications for a real laser crosslink mission. These case studies are evaluated for their ability to present the relevant information to the operator. For Earth Sensing, this involves displaying information regarding the spacecraft’s location with respect to the Earth, and enabling the selection of ground targets for imaging. For astronomy, the relevant information concerns the stars visible in the sky, and the spacecraft’s relationship to sources of interference like the Sun and Moon. For the laser crosslink example, we study the operator’s understanding of the spacecraft as they pass over a ground station and determine the operational configurations available for this communication. OPTASAT fills gaps in the field. OPTASAT presents users with a tool which is flexible and intuitive to use for understanding data from spacecraft in a way that is not currently available in the offerings on the market. Additionally, it takes functionality that is currently available in proprietary paid software and makes it available for free, in an open source offering that is accessible to everyone. OPTASAT will allow spacecraft operators (especially those operating spacecraft for the first time) to confidently know the state of their spacecraft, enabling them to make the best decisions for their satellites. This will reduce barriers to entry and smooth the learning curve, reducing the amount of overhead to new spacecraft operators. OPTASAT will be yet another step in the ongoing process of making space more accessible to a larger pool of users.Ph.D
Optimizing Solar PV Deployment in Manufacturing: A Morphological Matrix and Fuzzy TOPSIS Approach
No full textThe growing energy demand of the industrial sector and the need for sustainable solutions highlight the importance of efficient decision making in solar photovoltaic (PV) implementation. Selecting optimal PV configuration is complex due to the interdependent technical, economic, environmental, and social factors involved. This study introduces an integrated decision-making method combining a morphological matrix and fuzzy TOPSIS to systematically select and rank optimal PV system configurations for manufacturing firms. While the morphological matrix exhaustively examines possible design solutions based on sensing, smart, sustainable, and social (S4) attributes, the fuzzy TOPSIS method ranks the alternatives by handling uncertainty in decision making. A case study conducted in a Mexican manufacturing company validates the methodology’s effectiveness. The optimal PV configuration identified comprehensively addresses operational and sustainability criteria, covering all lifecycle stages. This approach demonstrates quantitative superiority and greater robustness compared to existing fuzzy TOPSIS-based methods for solar PV applications. The findings highlight the practical value of data-driven, multi-criteria decision making for industrial solar energy adoption, enhancing project feasibility, cost efficiency, and environmental compliance. Future research will incorporate discrete event simulation (DES) to further refine energy consumption strategies in manufacturing
Organizational Forms and Practices: Essays on Implications for Frontline Workers and Performance
No full textIn three essays, this dissertation explores how organizational forms and workforce practices shape frontline work experiences and organizational performance. Using both quantitative and qualitative methods, I explore how frontline workers experience work and what factors shape their performance. In the first essay, I examine how workforce practices in nursing homes relate to organizational performance. Specifically, I evaluate performance on resident health outcomes for both pre-pandemic and COVID-19 conditions. Combining Federal and state administrative data sets with non-public data on early COVID-19 spread and mortality, I investigate the degree to which the organization of work for frontline workers predicted resident health as a measure of organizational performance for nursing homes. In a period of global stress on health and care systems, I seek to understand to what extent prepandemic predictors of performance remained important. When nurses spent more time with residents, residents experienced better care both before and during the pandemic. Yet contrary to expectation, the role of clinical outsourcing became more relevant during the pandemic, potentially reflecting greater workforce flexibility or targeted COVID-19 workforce support to facilities that outsourced nursing activities before the pandemic. These results depict how environmental changes and alternative performance measures call into question established relationships in the high-performance work systems literature. In the second essay, I use in-depth interviews and field observations to uncover the process of constructing ownership culture in an employee-owned firm. I demonstrate how workers co-create their own control system, supported by a high financial value of ownership, strategic managerial communication, peer pressure, and performance management. This critical case challenges the dominant view in the employee-ownership literature that success requires formal worker participation in decision-making. Further, it investigates the “black box” of culture-building in an employee-owned firm. The third essay builds on this understanding by evaluating the stated motives of individual worker-owners in a home care cooperative. The cooperative developed as a pilot initiative with non-profit partners to develop a home care organization that would provide quality jobs and quality care, while integrating immigrant workers. I traced the workers’ justifications for joining and participating in these cooperatives. Rather than aligning with expected motives from previous studies or with Worker Center motives, I find that these workers adapted motives to reflect their realities, such as multiple jobs and a lack of labor rights in practice. This analysis emphasizes the decoupling of workers’ experiences from stated organizational goals, emphasizing the importance of collecting workers’ perspectives. Taken together, these three essays contribute insights into how frontline workers shape organizational performance by interpreting organizational context, culture, and structure. Results indicate that organizational performance is not merely a function of workplace practices, but rather, directly influenced by frontline workers based on their individual motives and roles in workplace culture. These findings imply that by directly engaging with frontline workers’ motives, organizational leaders and policymakers can design organizations that improve work and performance.Ph.D
Interpretable and Automated Bias Detection for AI in Healthcare
No full textBiases in artificial intelligence systems and the data they operate over are a major hurdle to their application in clinical and biomedical settings. Such systems have frequently been shown to fail to generalize from their training data to the real world environment and often display differing levels of accuracy over different population subgroups, which has detrimental effects on patients' quality of care and on healthcare equality. Here, we introduce an automated framework for identifying and understanding nontrivial sources of bias in healthcare datasets and AI models. Our framework is data and model agnostic and does not rely on human-developed heuristics or assumptions to uncover bias. We demonstrate its effectiveness by uncovering serious and nontrivial sources of bias in three widely used clinical datasets and one biomedical dataset, over the diverse tasks of diabetes risk prediction, lung cancer risk prediction, and biomolecular toxicity prediction. Our framework is used to uncover biases caused by patient BMI and computed tomography (CT) scanner type in the data used by a cutting-edge lung cancer risk prediction AI model, causing AUC drops on the order of ten percent.S.M
Guiding Deep Probabilistic Models
No full textDeep probabilistic models utilize deep neural networks to learn probability distributions in high-dimensional data spaces. Learning and inference in these models are complicated due to the difficulty of direct evaluation of the differences between the model distribution and the target. This thesis addresses this challenge and develops novel algorithms for learning and inference based on the guidance of complex parameterized distributions towards desired configurations via signals from auxiliary discriminative models.
In the first part of the thesis, we develop novel stable training objectives for Generative Adversarial Networks (GANs). We show that under standard unary-discriminator objectives, most of the valid solutions, where the learned distribution is aligned with the target, are unstable. We propose training objectives based on pairwise discriminators that provably preserve distribution alignment and demonstrate improved training stability in image generation tasks.
In the second part of the thesis, we introduce distribution support alignment as an alternative to the distribution alignment objective and develop a learning algorithm that guides distributions towards support alignment. We demonstrate the effectiveness of our approach in unsupervised domain adaptation under label distribution shift. Recent works have shown that under cross-domain label distribution shift, optimizing for distribution alignment is excessively restrictive and causes performance degradation. Our algorithm, which is based on support alignment, alleviates this issue.
In the third part of the thesis, we develop a novel approach to compositional generation in iterative generative processes: diffusion models and Generative Flow Networks (GFlowNets). Motivated by the growing prominence of generative models pre-trained at scale and the high training costs, we propose composition operations and guidance-based sampling algorithms that enable the combination of multiple pre-trained iterative generative processes. We offer empirical results on image and molecular generation tasks.Ph.D
Two's More Fun than One: How the Presence of Multiple Nutrients Changes Microbial Competition and Foraging in Unexpected Ways
No full textMicrobes exist in incredibly diverse environments with many possible resources (i.e. nutrients) to compete and forage for. To make this complex system tractable, ecologists often study microbes in the presence of a single resource in order to predict and explain what happens with multiple resources. But what gets lost when we do this? Are there phenomena that only emerge in the presence of multiple resources? Here, I explore the ecological implications of three phenomena that each require the presence of at least two resources. First, I show that the diauxic lags that occur when a microbe needs to switch between resources after one is depleted can allow ‘fast-switcher’ microbes to coexist with competitors that exclude them in single-resource environments. Then, I derive a rich temporal niche structure that arises from variations in the order in which resources are depleted in ecosystems with a pulsed resource supply and show that these temporal niches reshape community structure, vastly increasing the expected diversity of microbial ecosystems. Finally, I present a novel differential strategy in which a microbe attempting to intercept a moving source of multiple resources can treat one resource as an attractant and the other as a repellent to significantly increase its chances of successfully intercepting the source as compared to just being attracted to the resources released by the source. Each of these phenomena fundamentally requires the presence of at least two resources and reshapes microbial behavior and ecology. Thus, they collectively highlight the need to carefully consider how characterizations from single-resource environments actually combine to determine what happens in multi-resource environments and what new dynamics must be accounting for in such a bottom-up approach. I conclude with an argument that the case of two resources may be particularly relevant to study due to how much complexity can emerge at just the first step up from one resource to two.Ph.D
Reducing Transformer Key-Value Cache Size with Cross-Layer Attention
No full textKey-value (KV) caching plays an essential role in accelerating decoding for transformer-based autoregressive large language models (LLMs). However, the amount of memory required to store the KV cache can become prohibitive at long sequence lengths and large batch sizes. Since the invention of the transformer, two of the most effective interventions discovered for reducing the size of the KV cache have been Multi-Query Attention (MQA) and its generalization, Grouped-Query Attention (GQA). MQA and GQA both modify the design of the attention block so that multiple query heads can share a single key/value head, reducing the number of distinct key/value heads by a large factor while only minimally degrading accuracy. In this work, we show that it is possible to take Multi-Query Attention a step further by also sharing key and value heads between adjacent layers, yielding a new attention design we call Cross-Layer Attention (CLA). With CLA, we find that it is possible to reduce the size of the KV cache by another while maintaining nearly the same accuracy as unmodified MQA. In experiments training 1B- and 3B-parameter models from scratch, we demonstrate that CLA provides a Pareto improvement over the memory/accuracy tradeoffs which are possible with traditional MQA, potentially enabling future models to operate at longer sequence lengths and larger batch sizes than would otherwise be possible.S.M
Remanufacturing and Energy Savings
No full textRemanufactured products that can substitute for new products are generally claimed to save energy. These claims are made from studies that look mainly at the differences in materials production and manufacturing. However, when the use phase is included, the situation can change radically. In this Article, 25 case studies for eight different product categories were studied, including: (1) furniture, (2) clothing, (3) computers, (4) electric motors, (5) tires, (6) appliances, (7) engines, and (8) toner cartridges. For most of these products, the use phase energy dominates that for materials production and manufacturing combined. As a result, small changes in use phase efficiency can overwhelm the claimed savings from materials production and manufacturing. These use phase energy changes are primarily due to efficiency improvements in new products, and efficiency degradation in remanufactured products. For those products with no, or an unchanging, use phase energy requirement, remanufacturing can save energy. For the 25 cases, we found that 8 cases clearly saved energy, 6 did not, and 11 were too close to call. In some cases, we could examine how the energy savings potential of remanufacturing has changed over time. Specifically, during times of significant improvements in energy efficiency, remanufacturing would often not save energy. A general design trend seems to be to add power to a previously unpowered product, and then to improve on the energy efficiency of the product over time. These trends tend to undermine the energy savings potential of remanufacturing
Scalable and Sustainable Microwave Power Beaming to Mobile Lunar Surface Assets
No full textLunar missions are hindered by the challenges of maintaining continuous operation, especially during the 14-day lunar night, when solar power sources may be unavailable, causing significant mission delays and limiting efficiency. Frequent returns to charging stations supplied by fixed lunar surface power plants further disrupt workflows and restrict the operational range of lunar vehicles. To address these issues and enhance lunar mission performance, a continuous, secure, and shareable power source is essential. While nuclear power and larger battery systems are viable options for continuous lunar energy supply, they pose challenges such as safety risks, complex deployment, and limited scalability. This thesis focuses on exploring microwave-beamed power systems as a flexible and scalable solution for sustained lunar operations. Ideally, the power source would enable 24/7 operations without requiring vehicles to return to base stations, allowing for unrestricted navigation across the lunar surface, including in permanently shadowed regions (PSR). In addition, it would support the construction of critical infrastructure, accelerating the development of the lunar economy. This thesis aims to support sustained lunar exploration and infrastructure development by exploring the design space for microwave-beamed power systems under three different demand use cases of increasing scale, loosely corresponding to the three phases of the Artemis program: Local (Shackleton Crater), Regional (navigation between equatorial regions and South Pole), and Global (entire lunar surface). A case study focused on the YUTU-2 lunar rover investigates alternative architectures for each use case, comparing power beaming from tall towers vs. satellites. Evaluation reveals that the most effective solution for the Local use case is a tower-based approach featuring a single 100m tower, >10,000 solar modules, and using 1 GHz operating frequency, at a cost of 1.7M/W - 0.8M/W. The trade studies showed that larger receiver antenna areas and lower frequencies improve performance and cost-effectiveness. Furthermore, larger microwave-beamed power systems leverage economies of scale, lowering the cost per watt by an average of $1M/W when scaling from the Regional to the Global power system, with potential for further reductions through future expansions.S.M