150813 research outputs found
Sort by
Learning to Solve Long-Horizon Robot Manipulation Problems
No full textIf we want mobile robots that perform multi-step tasks in visually diverse and geometrically complex environments, we need them to quickly decide what to do and how to do it. Manipulating multiple objects in environments with movable and articulated obstacles over time requires the robot to satisfy constraints like collision-freeness, reachability, and action feasibility. For problems with large state spaces, continuous action spaces, and long decision horizons, the hybrid constraint satisfaction problems induced by planners become combinatorially difficult to solve. In this thesis, I will discuss strategies for using offline learning to speed up deploymenttime planning, i.e., using a plan feasibility predictor, a subgoal generator, or a compositional joint continuous constraint solver. I will also present strategies for chaining policies learned from demonstrations using conditional inputs, such as key poses and natural language, for generalization in real-world environments. With the resulting efficient long-horizon manipulation planning system, we can solve complex robotic manipulation problems faster at deployment time. It can also be used to generate diverse large-scale whole-body trajectories as part of the data mixture for training robot foundation models in embodied reasoning, planning, and acting.Ph.D
Interplay of ALP couplings at a muon collider
No full textAxion-like particles can couple to Standard Model gluons, electroweak gauge bosons, and massive fermions. A future multi-TeV muon collider provides a favorable environment to probe axion-like particles through multiple production channels, including vector boson fusion via electroweak gauge boson couplings and the top-associated production mediated by direct fermionic couplings. Motivated by the quality issue of the QCD axion, we focus on axion-like particles with masses and decay constants around the TeV scale. We explore how different axion-like particle couplings shape its production and decay modes, revealing a rich and intricate phenomenological landscape
DOE selects MIT to establish a Center for the Exascale Simulation of Coupled High-Enthalpy Fluid–Solid Interactions
No full textThe research center, sponsored by the DOE’s National Nuclear Security Administration, will advance the simulation of extreme environments, such as those in hypersonic flight and atmospheric reentry.Institute for Soldier Nanotechnologie
SquareLoop: Explore Optimal Authentication Block Strategy for ML
No full textHASP 2025, Seoul, Republic of KoreaOff-chip memory in ML accelerators is vulnerable to both hardware
and software attack, which needs encryption and authentication.
Precise performance modeling of it requires (1) representation of
authentication blocks (AuthBlock) to cover the full design space of
shapes and orientations, and (2) precise memory behavior modeling,
as encryption and authentication mainly increase memory traffic.
This paper introduces
2Loop, a framework that resolves these
challenges by introducing (1) flexible, all-level partitioning based
AuthBlocks for ensuring full coverage of the entire design space, (2)
a realistic layout-based memory model, and (3) an Mapping-LayoutAuthentication co-search algorithm to explore the drastic combinatorial design space to figure out optimal mapping, layout, and
AuthBlock shape choice for multi-layer workloads. SquareLoop’s
detailed memory model helps find better mapping to achieve 1.32×
speedup on ResNet18 compared to the SotA SecureLoop, and our
latency predictions are validated to within 7.3% of an RTL implementation.
2 also achieve up-to 1.08×/1.82× overall speedup for
authenticated ResNet18/MobileNet-V3 on various accelerators with
AuthBlock and Mapping co-searching. We open-source
2Loop to
provide a powerful and validated tool for designing efficient, secure
accelerators at https://github.com/maeri-project/squareloop
Remove hydrogen and store it too: an acid-in-clay based electro-chemical solution
No full textExtracting hydrogen from metallic components can open up a new pathway for preventing hydrogen embrittlement. To this end, we propose an electrochemically driven, all-solid method for hydrogen control, capable of both extracting and storing hydrogen simultaneously. In this approach, we employ acid-in-clay as a proton conducting electrolyte at room temperature. Through this electrochemical treatment, hydrogen is efficiently extracted from pre-charged steels, thereby restoring their tensile properties and preventing embrittlement. Moreover, it has been confirmed that the extracted hydrogen can be efficiently collected at the counter electrode, demonstrating the significant advantages of the process
Need Help? Designing Proactive AI Assistants for Programming
No full textCHI ’25, Yokohama, JapanWhile current chat-based AI assistants primarily operate reactively, responding only when prompted by users, there is significant potential for these systems to proactively assist in tasks without explicit invocation, enabling a mixed-initiative interaction. This work explores the design and implementation of proactive AI assistants powered by large language models. We first outline the key design considerations for building effective proactive assistants. As a case study, we propose a proactive chat-based programming assistant that automatically provides suggestions and facilitates their integration into the programmer’s code. The programming context provides a shared workspace enabling the assistant to offer more relevant suggestions. We conducted a randomized experimental study examining the impact of various design elements of the proactive assistant on programmer productivity and user experience. Our findings reveal significant benefits of incorporating proactive chat assistants into coding environments, while also uncovering important nuances that influence their usage and effectiveness
Expansion Microscopy of Extracellular Space for Light Microscopy-Based Connectomic Analysis
No full textIn this dissertation, we present an exploratory methodology, termed expansion microscopy of extracellular space (ExECS), designed to enhance the visualization of the extracellular space (ECS) within aldehyde-fixed tissue. This technique leverages the principles of expansion microscopy (ExM), a method that facilitates nanoscale imaging on conventional microscopes through physical magnification of specimens, thereby supporting improved visualization of various cellular and tissue components including proteins, nucleic acids, and lipids 1. The ECS forms a continuous environment between cells2. Its presence throughout neural tissue makes it an attractive target for contrast-based techniques such as shadow imaging, where the ECS is selectively labeled to produce negative contrast, revealing cell shapes and boundaries as unlabeled silhouettes within a labeled background. Although ECS delineation in fixed tissue is limited by the fidelity of fixation and may not fully reflect its live-state structure, the resulting contrast with the intracellular environment may offer useful contrast for investigating neural morphology and connectivity, offering a useful approximation of network organization. A key component of the ExECS methodology is the introduction of a customengineered ECS Filler solution. This formulation, detailed later, includes a macromolecular probe intended to serve as a proxy for the ECS. When applied to aldehyde-fixed tissue, the filler is designed to diffuse throughout the sample, preferentially occupying extracellular compartments while remaining largely excluded from intracellular regions. This selective distribution is expected to persist even in areas where aldehyde fixation may have increased membrane permeability. This diffusion behavior is presumed to result from a combination of size-based exclusion and intermolecular interactions between the hyaluronan polymers, which form the main component of the filler solution, and the plasma membrane. The constituent hyaluronan is functionalized with amine groups to enable covalent crosslinking and with azide groups to allow fluorescent tagging via click chemistry. These modifications are intended to enable the ECS filler to act as a contrast agent by labeling the extracellular space, providing a foundation for a shadow-based imaging strategy to delineate morphology of cellular structures. In parallel, we introduce a lipid-targeted form of ExM, termed membrane expansion microscopy (mExM). This approach employs a custom chemical tag that enables nanoscale optical imaging of lipid membranes using a lipid-optimized expansion protocol. mExM, via a novel post-expansion antibody labeling protocol, enables protein-lipid relationships to be imaged in intracellular organelles. This technique may offer new opportunities to examine aspects of neural circuitry by linking cellular morphology with molecular identity. Together, ExECS and mExM offer a potential basis for a light microscopy-based framework for connectomic reconstructions. Unlike traditional electron microscopy approaches, which are labor-intensive and low-throughput3, this strategy aims to improve throughput in mapping of neuronal morphology with enhanced resolution that surpasses diffraction limitations. With the aim of bridging the gap between tissue ultrastructure and optical accessibility, this work may contribute to efforts toward scalable, high-resolution analysis of neural tissue organization.Ph.D
Mass and Distance Estimation Simulations for the Nancy Grace Roman Space Telescope Using PyLIMASS and a Case Study on Intellectual Property Frameworks in Space Collaborations
No full textGravitational microlensing is a phenomenon in which a foreground star or planet briefly magnifies light from a more distant background star. This effect enables the discovery of exoplanets that are otherwise undetectable, including those orbiting faint hosts and at large separations. Microlensing is well suited to characterizing exoplanets beyond the snow line, revealing mass ratios and orbital geometries inaccessible to transit or radial velocity methods. The Nancy Grace Roman Space Telescope will carry out the Galactic Exoplanet Survey to detect thousands of microlensing events with the cadence and precision necessary for statistical exoplanet population studies. To verify Roman’s ability to meet its core science requirement, recovering the lens mass and distance in at least 40% of planetary events with better than 20% uncertainty, targeted simulations are essential. Using the pyLIMASS inference framework and Fisher matrix-based uncertainty propagation, I demonstrate that for the well-characterized event OGLE-2013-BLG-0132Lb, the lens mass can be constrained to within 18.7% uncertainty, validating the feasibility of Roman’s requirement on a case-study basis. This thesis also addresses the legal and policy foundations needed to ensure global access to these simulation tools. By advancing open-source software models and proposing a space IP framework for equitable knowledge sharing, it supports collaborative scientific infrastructure for future international space missions.S.M
Toward Scalable Learning-Based Optical Restoration
No full textAPNET 2025, Shang Hai, ChinaThe increasing scale and dynamic nature of modern optical networks present significant challenges to the scalability and adaptability of fault recovery. Existing state-of-the-art (SOTA) optical restoration methods rely primarily on offline pre-computation for each fault scenario, followed by online traffic reallocation. Their scalability to large network topologies is limited by the reliance on traditional solvers and imprecise modeling of potential faults.
This paper proposes LBOR, an optical restoration system built on multi-agent reinforcement learning (MARL) and integrated with a traffic allocation framework. We introduce a sequential restoration workflow for each failed IP link, employing two agents dedicated to path selection and wavelength assignment, respectively. In addition, we develop a randomized assignment ordering strategy to mitigate premature convergence to local optima and an action masking mechanism to prune the MARL search space. Experiments conducted on a large topology with 70 nodes indicate that LBOR achieves up to a 1000 × speedup compared to the SOTA approach, with only a slight reduction in allocation precision
Metrics, Muons, Moments, Models, Machine Learning, Measurements, and More: A Manifesto on Collider Physics
No full textThe interface between particle theory and particle experiments is essential to improving our understanding of the Standard Model and looking for new physics beyond it. At this interface lies a complicated web of complex and expensive simulations that cannot fully be trusted, experimental and theoretical uncertainties, overwhelmingly large amounts of data, all while we have yet to find any deviations from the Standard Model.
In this thesis, we propose strategies for improving the theory ↔ experiment pipeline at all stages. We first show how modern Machine Learning and statistical techniques can be used to improve the calibration and resolution of particle detectors in a robust way, which can lead to improved measurement precision. We then develop brand new classes of measurable observables based on the principle of infrared-and-collinear-safety, geometry, and machine learning, which come with guarantees about their theoretical calculability and interpretability, in turn motivating measurements at collider experiments. Finally, we then present two complementary approaches to search for new physics: one, in the form of an experimental proposal for a muon beam dump experiment that is viable alongside a full future collider program; and the other, in the form of machine-learning based anomaly detection to search for subtle signals in already-published data.Ph.D