545298 research outputs found
Sort by
Multisensory Dexterity for Robotics
Get PDFHuman hands are fundamental to how we sense and act upon the physical world. Their unique ability to coordinate precise movements and adapt to changing conditions gives us the dexterity to grasp, manipulate, and interact with objects of different shapes and physical properties with remarkable ease. Replicating this level of dexterity in robots is critical for building general-purpose robots that can operate reliably in unstructured environments. Although modern artificial intelligence has achieved significant success in vision and language, dexterous manipulation remains a major unsolved challenge. This difficulty arises from the high dimensionality of motor control, the scarcity of real-world data, and the need to integrate diverse sensory inputs into robust behaviors. This thesis aims to close this gap by developing learning-based systems that equip robots with multisensory intelligence for dexterous manipulation. It shows how to combine vision, touch, and proprioception with scalable learning methods to enable robots to perform complex, contact-rich tasks that require coordination and adaptation.
The approach follows a structured learning paradigm. First, we focus on acquiring individual manipulation skills through large-scale training in simulation. Each skill is developed independently and grounded in principles of generalization across objects and physical properties. Second, we investigate how rich multisensory feedback, including tactile sensing, visual inputs, and proprioceptive signals, can be fused to improve both perception and control. We show that these modalities are not redundant but complementary, and their integration allows robots to perform tasks that remain infeasible with simple sensors alone. Third, we propose a compositional framework that builds on previously learned skills to acquire new behaviors efficiently. The thesis concludes by identifying key challenges ahead and outlining directions for future research in developing robots that can interact with the world as fluently and flexibly as humans do
People are (Shockingly) Bad at Hedging
Get PDFPeople often must plan for the worst. They purchase product warranties, insure their homes, and proactively make backup plans. People should be willing to pay more to hedge against bad outcomes that are more likely to happen. Across 14 studies (N = 5,591) we find that decision-makers instead behave as though they almost fully ignore probability information when hedging against bad outcomes. As a result, they dramatically underinvest in hedges they are likely to need, while overinvesting in those that are unlikely to be helpful. This behavior occurs in both abstract settings (e.g. hedging lotteries) and naturalistic ones (e.g. buying warranties or insurance), occurs with fully incentive-compatible decisions, and is robust across a wide range of probabilities and outcomes. In our studies, we test a variety of possible explanations. Ultimately, we find support for an account in which decision makers focus almost solely on the bad outcome they are hedging against, while ignoring how likely that bad outcome is to occur. Interestingly, when the same participants invest to make a good outcome better (rather than hedge to make a bad outcome less bad), they were sensitive to probabilities. Leveraging this result, we find that a reframing of hedges which makes them appear more like investments can make hedging decisions better calibrated to the likelihood of an outcome
Computational Statistics for Medical Diagnostics: From RT-qPCR to Pathological Imaging
Get PDFPublic health challenges require sophisticated analytical approaches to handle massive, heterogeneous datasets spanning genetic, behavioral, social, and clinical domains across diverse populations and geographic regions. While classical statistical methods struggle with such complex data structures, deep learning models demonstrate significant potential in delivering more accurate predictive solutions with big data. This study addresses key concepts in computational statistics including prediction performance improvements, ground truth quality, big data approaches, and model interpretability through two distinct applications in public health diagnostics. We present SPARK, a deep learning model for RT-qPCR curve analysis that significantly outperforms current practice in SARS-CoV-2 testing. SPARK achieves a false negative rate of 1% with only 4.38% false positives, compared to over 60% false positives with conventional thresholding methods. Critically, we train SPARK using true ground truth labels from RT-qPCR quality control curves rather than human-generated labels, enabling the model to learn beyond current thresholding practice and differentiate itself from other deep learning approaches trained on human labels. In pathological image analysis, we demonstrate deep learning applications in two clinical contexts: predicting breast cancer development from benign breast disease whole slide images and detecting pancreatic malignancy from fine-needle aspiration cytology images. One of our models significantly outperforms estimated clinician prediction for breast cancer risk (AUROC: 0.577 vs. 0.472, p ≤ 0.05) and the other model achieves high accuracy in pancreatic malignancy detection with an average precision of 0.98. Our methods enable accurate diagnostic guidance through attention-based interpretability methods that highlight relevant image regions for clinical decision-making. These applications demonstrate how computational methods can transform public health practice by leveraging high-quality ground truth data and deep learning architectures to improve diagnostic accuracy and enhance clinical decision support
Indirect Detection as a Window into Physics Beyond the Standard Model
Get PDFWhile the Standard Model has proven remarkably successful in describing nature at a fundamental level, several key questions remain unanswered, such as the origin of neutrino masses and the nature of dark matter. As a means of probing physics beyond the Standard Model, we present a comprehensive study of indirect detection strategies, focusing on several well-motivated extensions: sterile neutrinos, weakly interacting massive particles (WIMPs), higgsinos, and axions. For sterile neutrinos, WIMPs, and higgsinos, we conduct analyses under the assumption that these particles constitute the dark matter of the universe. We probe their annihilation or decay signatures using astrophysical observations from X-ray and gamma-ray telescopes, leveraging the dense dark matter environments of the Milky Way. For axions, we investigate their indirect signatures in two settings: topological defects known as axion strings, and astrophysical environments with abundant axion production, such as core-collapse supernovae. We show that axion strings can deposit a fraction of their energy into Standard Model particles, leading to observable effects in cosmological data such as the cosmic microwave background and big bang nucleosynthesis. In high-production environments, we explore the potential for axion indirect detection through axion-photon conversion in magnetic fields or through the decay of heavier axions. By combining observational data with theoretical modeling, we derive new limits from the non-observation of a signal and demonstrate the power of indirect detection in the search for physics beyond the Standard Model
Machine Learning for Inorganic Materials Synthesis from Scientific Literature
Get PDFScientific literature contains decades of “how-to” knowledge on making inorganic materials, yet the information is locked in various forms of unstructured text, figures and tables. This dissertation develops machine-learning pipelines that convert unstructured text into structured, machine-readable data and then leverage it for synthesis science. First, I quantify how domain-specific pre-training and fine-tuning improve named-entity recognition and relation-extraction for materials science text, in MatBERT NER model development. Building on these models, I present two large, curated datasets and downstream analyses - seed-mediated gold nanoparticle synthesis and solid-state synthesis with impurity phases. For seed-mediated gold nanoparticle synthesis, a hybrid (rule-based and Machine Learning) parser, yields 492 fully validated recipes spanning spheres, rods, stars and other shapes. Statistical and interpretable ML models recover known morphology drivers - most notably the dominant role of the seed capping agent - and expose previously overlooked variability in reported aspect ratios. For solid-state inorganic synthesis with impurity phases, few-shot Large Language Model (LLM) extraction pipeline extracted 80,823 recipes, of which 18,874 explicitly report impurity (i.e., “failed”) products. The inclusion of negative outcomes provides new insights in synthesis routes and factors affecting phase pure syntheses. Together, these contributions show how LLMs can transform scattered resources of experimental data into comprehensive data, and learn rules and insights for targeted synthesis. Exploring many methods from rule-based Natural Language Processing (NLP) to fine-tuned LLM, this work advances data-driven materials discovery and opens new avenues for accelerating the design of inorganic materials
Traffic and the Twilight of British Liberal Governance, 1870s to 1939
No full textThis is a dissertation about the “civilization” of a new technology, the attempt to get millions of people to change their norms and habits in response to new conditions, and a new threat. When modern automobiles first appeared on British roads at the end of the nineteenth century, there were practically no rules to guide or constrain their drivers. The motor car threw long-established patterns of street movement into chaos, and soon brought violent death to thousands of people every year. Unlike other disruptive technologies, however, the automobile’s violence could not be contained at the point of production: early-twentieth-century Britons considered the motor car to be an extension of self, and driving an expression of character. To them, traffic violence was a social and ethical problem, not a systemic or technical one. Nor could the British state do much to force a sense of responsibility on motorists: they were too numerous, too mobile, and too socially powerful to be effectively disciplined by police. Administrators quickly realized, then, that the integration of automobiles into civil life would require the civilization of drivers themselves. Road users would have to internalize new limits on their bodily instincts and desires, new expectations of the street, and a new, visceral understanding of responsibility and risk.
This thesis follows that civilizing mission in Britain, from the early days of cycling in the 1870s to the dawn of mass automobility in the years before WWII. It centers on the rise of modern traffic rules – red lights, stop signs, speed limits, crosswalks, etc. – and argues that these unglamorous legal tools are, in fact, living artefacts of the interwar crisis of liberalism. In 1920s Britain, elite motorists and their allies in government believed that hallowed liberal virtues – civility, fraternal sociability, fair play, common sense, and the responsible exercise of liberty – could fix the traffic-safety crisis without recourse to the criminal law. Consequently, early traffic rules in Britain were meant to function as soft social paradigms, not hard legal requirements. By the mid-1930s, though, this image of the driver as a sociable and self-regulating gentleman had fallen apart: gentlemen had proven themselves unable to restrain their murderous driving, and driving was no longer the exclusive preserve of gentlemen. The failure of a liberal approach to traffic “civilization” made way for a new technocratic project of control. The latter part of this thesis observes administrators struggling to repurpose those old liberal traffic tools, which were designed to help elites manage themselves, into scalable procedures for managing masses. The result was a complex accord between discretion and rule, and a major shift in ordinary Britons’ experience of legal authority: by the end of the 1930s, traffic violations were the single most common way they encountered the criminal justice system. The technocratic turn in traffic control was ambivalent, uneven, and incomplete. For that very reason, it is a powerful expression of the unsteady triangulation at the heart of British governance after 1919. On one side stood the oligarchic structures of “club government,” where good administration was synonymous with the discretion of men of character. On the other side stood planning, technical expertise, and the rule-based decision-making procedures of “modern” scientific government. Standing between these two styles of governance is a new mass democracy, being courted by both camps, and struggling to decide where to place its hopes for fairness, freedom, and safety
Cargo Cult: Logistics, Labor, and the Secret Life of Supply Chains
Get PDFThis dissertation examines the contemporary labor politics and colonial history of the global logistics industry, focusing primarily on the Ports of Los Angeles and Long Beach. It argues that logistics is not a recent technical innovation born of midcentury containerization, but rather a global political project with roots in early modern empire, whose material and conceptual infrastructure emerged through centuries of maritime conquest, racialized labor regimes, and territorial expansion. Moreover, it contends that logistics should not be understood as a neutral or purely economic system for the movement of goods, but rather as a mode of governance—a way of doing politics with technics—aimed at organizing space, time, and subjectivity in service to capital accumulation. Drawing on 14 months of collaborative ethnographic fieldwork with longshore workers at L.A./Long Beach, interviews with key industry actors, and archival research spanning three continents, the dissertation brings together empirical investigation with historical and theoretical analysis to suggest that today’s logistics industry functions as a kind of cargo cult: an ideological formation obsessed with optimization, fixated on the commodity form, and haunted by the colonial past it seeks to obscure. In framing logistics this way, I aim to unsettle dominant narratives of technological progress and economic rationality, and to foreground instead the hidden histories, violences, and contestations that continue to shape global supply chains in the present.
The dissertation consists of an introduction, three chapters, and a brief coda. Chapter one examines a conflict between L.A./Long Beach dockworkers and the shipping company Maersk over automating Pier 400, the largest container terminal in the Western Hemisphere. It argues that rather than replacing longshore labor outright, port automation displaces and conceals labor, and thus functions as a ruse. Chapter two offers a material genealogy of containership design to show how imperial logics are embedded in logistical infrastructures. Through archival work tracing the bulkhead from ancient China to the British Empire to South Korean shipyards, the chapter theorizes logistics as a colonial formation. Chapter three explores how automation transforms the sensory experience of longshore work. Drawing on Jacques Rancière’s notion of the “distribution of the sensible,” it argues that logistical technics discipline perception in ways that produce both disorientation and new forms of political subjectivity, including worker paranoia. The coda speculates on these findings, outlines some lacunae, and describes directions for future research
Meconium-Related Obstruction: Contemporary Experience in a Multi-Institutional Consortium
Get PDFPurposeNeonatal bowel obstruction secondary to inspissated meconium has been historically associated with cystic fibrosis. Increasingly, meconium-related obstruction (MRO) has been observed in preterm infants. We conducted a multicenter mixed-methods study to better characterize the contemporary experience with MRO.MethodsA retrospective cohort study of infants with MRO was performed at seven children's hospitals from 2018-2022. Chi-squared tests, Kruskal-Wallis tests, and logistic regression were used to assess the association of cystic fibrosis, Hirschsprung disease, and prematurity with treatment strategies and clinical outcomes of MRO. Providers were surveyed regarding their management of MRO of prematurity.ResultsWe identified 105 infants treated for MRO, including 54 (51%) with MRO of prematurity, 16 (15%) with Hirschsprung disease, 6 (6%) with cystic fibrosis, and 29 (28%) with MRO of the term infant. Overall, 32% (n=34) received glycerin suppositories, 25% (n=26) rectal irrigation, 79% (n=83) contrast enemas, and 6% (n=6) retrograde or antegrade N-acetylcysteine. Twenty-seven (26%) infants required surgery for MRO, of whom 21 (78%) had MRO of prematurity. For infants with MRO of prematurity, a one-week increase in gestational age was associated with a 23% decrease in the odds of requiring surgery (OR=0.77; 95%CI=0.64-0.93). Survey responses from 42 providers suggested limited institutional treatment algorithms for MRO.ConclusionIn this multi-institutional study, most cases of MRO were associated with prematurity. Extent of prematurity was associated with a higher likelihood of requiring surgery. Results from this contemporary cohort study and survey provide a framework for developing a prevention and treatment algorithm for MRO of prematurity
Multisensory Temporal Integration of Simple, Neutral, and Trigger Stimuli in Misophonia
Get PDFMisophonia, a condition marked by decreased tolerance to everyday sounds such as chewing or slurping, can cause severe impairment in social and private settings. The present study examined whether the temporal binding window (TBW) differs between individuals with misophonia and controls for simple flash-beep, complex trigger, and complex neutral stimuli. We measured participants’ TBWs using a simultaneity judgment task(SJ) with simple, trigger, and neutral audiovisual stimuli. Our findings showed that the overall width of TBW varied as a function of stimulus complexity in both groups, with trigger stimuli having the widest TBW, followed by neutral stimuli. Our results also showed that the right TBW was significantly narrower for the trigger PAVS stimuli in the misophonia group, suggesting more temporal precision for misophonics when the video preceded the trigger sound. We also report a novel finding that the benefit of PAVS stimuli is maximal when they are synchronized compared to when the trigger precedes the video or vice versa. An unexpected finding was that misophoics were less susceptible to the McGurk illusion. Overall, this study addressed a novel theoretical question regarding the temporal constraint of multisensory processing of simple, neutral, and trigger stimuli in misophonia
High Power Density Partially Superconducting Electric Motors for Electric Aircraft Applications
Get PDFThe transition toward electrified aviation demands electric motors with exceptional power density, efficiency, and thermal stability. This dissertation introduces a series of innovative designs for double rotor flux-switching motors (DRFSMs) optimized for electric aircraft propulsion. By integrating high-temperature superconducting (HTS) YBCO field coils, aluminum Litz armature windings, and advanced cryogenic thermal management systems, the proposed topologies achieve substantial improvements in gravimetric power density and overall system efficiency. Three motor configurations are explored: DRFSMs with HTS field coils, DRFSMs with HTS field coils and superconducting magnetic shields, and a flux-reversal machine (DRFRM) with HTS field coils and magnetic shield. Finite element analysis (FEA) and parametric optimization maximize performance, resulting in designs achieving up to 100 kW/kg and efficiencies exceeding 99.5%. The results demonstrate the feasibility of employing superconducting technologies to meet the stringent performance targets of next-generation all-electric aircraft