I investigate whether combining ultra-wideband (UWB) radar with inertial measurement units (IMUs) can produce more robust human pose estimations than using IMUs alone. UWB radar yields precise distance measurements, offering positional data that standard IMUs—sensitive mainly to angular velocity and acceleration—cannot capture. To test this approach, I built an embedded system that integrates a UWB radar module with wearable IMUs, then designed a user study involving everyday movements and targeted exercises performed by a small group of participants. This setup allowed me to collect a diverse dataset under realistic conditions. I processed these data using neural network models, including long short-term memory (LSTM) and transformer architectures, to generate accurate joint angles. I then fed those angles into a 3D skeleton reconstruction model. My preliminary findings suggest that the additional distance data from the UWB radar substantially improves tracking accuracy and reduces ambiguity in limb positioning. This enhanced estimation could lead to more realistic virtual reality avatars, improved fitness tracking, and better physical therapy tools. By overseeing the hardware design, data collection, and model development, I actively demonstrate how interdisciplinary methods can advance human-computer interaction through more precise and accessible pose estimation.

4.5 million adults in the United States are diagnosed with chronic liver disease. Over time this can lead to cirrhosis, an end-stage condition in which scarring occurs in the liver. Reduced liver function from cirrhosis results in accumulations of neurotoxic substances that induce a spectrum of neurological impairments known as hepatic encephalopathy (HE). The critical flicker frequency (CFF) test is a well-established screening test for HE. Previously we developed Beacon, a novel and portable CFF measuring device that can be administered at home via smartphone app, as an accessible alternative to current CFF measurement devices that are large, expensive, and not intended for at-home use. We found that Beacon produced a CFF measurement that aligned with commercially available devices. While the current Beacon reflects current commercial devices, the efficiency of measurement is bottlenecked by the fact that pairs of flickering light stimuli can only be presented sequentially due to the singular light source. We therefore propose a dual headed version of Beacon that gives the option of flashing two frequencies simultaneously. I designed and developed a version of this dual-headed Beacon with sliding heads as well as an accompanying user interface before conducting a series of user studies, beginning with a pilot study on healthy individuals and progressing to a clinical trial on chronic liver disease patients, to evaluate the impact of the number of light sources and the distance between them on CFF measurement time and repeatability. I hypothesize that the two-headed Beacon will produce a CFF measurement more quickly than the original Beacon and that a closer distance between heads will also produce quicker and more consistent measurements. These findings will help inform the development of future iterations of the Beacon, leading to improved outcomes for chronic liver disease patients.

Large Language Models (LLMs) have demonstrated remarkable performance in code generation and problem-solving, but their effectiveness varies across problem domains and complexities. In this study, we assess the problem-solving capabilities of different LLMs on a subset of LeetCode problems, categorized by difficulty, topic, and computational complexity. Through a systematic analysis of LLM performance across question topics—such as dynamic programming, matrix, and binary search—we identify trends in their strengths and weaknesses. Our findings reveal that LLMs excel at problems likely included in their training data but struggle significantly with novel or unseen problems. By evaluating their performance across different algorithmic domains, we offer insights into the potential limitations in specific algorithmic domains and implications for AI-assisted coding, competitive programming, and software engineering education. 

Robots must be able to pick objects from densely packed shelves in order to automate industrial warehouses. Dense packing gives rise to challenges in grabbing targeted objects efficiently as the shelves can be clustered, objects can be stacked, and the target object can be obstructed from direct reach. The goal of this research project is to create a new gripper combined with reinforcement learning to manipulate objects within a shelf without multiple attempts or repositioning of the robot arm. The new gripper design includes four fingers that are linear actuators with vacuum units and suction cups attached to the ends of each finger. Additionally, each finger contains a time-of-flight sensor at the tips which provide spatial information for different objects within the shelf. I integrated time-of-flight sensors into the multi-fingered gripper and filtering algorithms for the sensor’s data. I modified the previous vacuum ejector unit which only provided support for one unit to four vacuum ejector units. I also conducted a series of experiments that provided cases where the new gripper design proved to be better than the previous design. We also collected suction cup vacuum data and trained a neural network to predict the success rate of suction cup attachment. The results of this project will inspire new designs for increasing the success rate of robotic grasps within densely packed environments.

Fluorophore quencher pairs are commonly used as reporters for DNA reactions due to their low background signal when untriggered and their ability to detect low DNA concentrations. However, these modifications are expensive and require a fluorescent plate reader to detect the signal, limiting their accessibility for point-of-care or low-resource settings. We are developing an alternative reporter using G-quadruplexes, which are guanine-rich DNA sequences with enzymatic activity in vitro. These structures can be utilized in detection assays due to their well-characterized peroxidase activity. Current approaches utilizing G-quadruplex structures have limited sensitivity due to high levels of background activity. Our approach combats this problem using altered G-quadruplex sequences inactivated by sequence-mismatched complexes, later activated by downstream reactions. By making these modifications, we detect DNA concentrations as low as 2 nanomolar. We hope this inexpensive approach can be utilized as an accessible alternative to traditional detection assays due to its colorimetric properties, leading to detection by the human eye and effective point-of-care detection.

The most compelling challenges in robotic behavior cloning arise when agents must perform precise, complex tasks - especially those that challenge even the best human expert demonstrators. The key question here is as follows: how can we most optimally utilize human-collected demonstration data in such domains? There are many ways to tackle the issue of robust, data-efficient robotic behavior cloning; we explore three: leveraging learned system dynamics to generate synthetic corrective data with assumed Lipschitz continuity, exploiting local structure by utilizing a cloud of distance-aware neighboring data points and their predicted actions, and ensembling past predicted action trajectories conditioned on their confidence to produce ideal, outlier-robust actions and even predict when an agent needs guidance and correction from a human expert. The first of the three has already been published as a series of works under the acronym CCIL and have shown large improvements in both simulated imitation tasks and real-world robotic fine manipulation, showing particular promise in low-data regimes. The latter two are ongoing research projects; the first, utilizing local neighborhood information, has shown promising results on simulated tasks, and work to transfer this algorithm to the real world is currently under development. The final of the three has shown promise on real-world robotic tasks as an out-of-distribution detector and confidence measurement tool, and research is underway to utilize this information for the purposes of robustness and corrective data collection. The projects and their findings all contribute towards the common goal of optimizing data usage in a robotic behavior-cloning paradigm, opening the door for robots to complete more and more complex and data-scarce tasks performed by humans.

Welcome all young and old to the future of movie magic! 2D animation remains a powerful storytelling medium, yet its resource-intensive nature has made it increasingly rare in today’s industry. What if we could change that? What if artificial intelligence (A.I) can work with, rather than against artists, making 2D animation more accessible? Could a small studio implement this and revive this beloved genre? Join international award-winning filmmaker G Alvarado as we explore cutting-edge image generation and video interpolation A.I models. Along with an enhanced 2D animation pipeline that preserves artistic integrity using customly trained models. Early findings suggest that this can significantly reduce production time, transforming what once took years into mere months. Come all far and near to see our research results in action and peek behind the curtain. For once you do, you will find through science comes art, and through innovation, a new era of storytelling begins!

DNA concentration sensing is important for accurate reagent input measurement and output data collection for various molecular biology applications, such as genomics, biotechnology, and clinical diagnostics. Common use cases for DNA concentration sensing include polymerase chain reaction (PCR), gel electrophoresis, and enzymatic assays. Off-the-shelf spectrophotometry systems, used today to measure DNA concentration, require an aliquot of DNA to be pipetted onto a sensor. The sample is then discarded to avoid contamination. Our goal is to develop a novel, cost-effective, and contactless method of containing and directly measuring DNA concentration in individual microliter droplets in real-time. Advantages of contactless containment are: (1) no sample is lost to adhesion to the container, (2) no spectral signature from the container material is added to the sample’s spectrum, and (3) samples can be weighed without contact for closed loop control of sample mass. To contain the droplets of DNA without contact, we use an acoustic levitation system. This system emits focused ultrasonic sound to lift, move and contain liquid droplets in air without making direct contact. Since DNA absorbs ultraviolet (UV) light at a wavelength of 260 nm, we use a low-cost, off-the-shelf spectroscopy sensor to build a portable DNA concentration measurement system within the levitation system to measure the amount of 260 nm UV light absorbed by the DNA. Preliminary results show that the device can distinguish samples containing different concentrations of DNA. Further research will focus on enhancing the device’s sensitivity and expanding its application to other fields related to biology.

Robotic manipulation systems operating in diverse, dynamic environments must exhibit three critical abilities: generalization to unseen scenarios, multitask interaction, and spatial memory. While significant progress has been made in robotic manipulation, existing approaches often fall short in addressing memory-dependent tasks and generalization to complex environmental variations. To bridge this gap, we introduce SAM2Act, a multi-view robotic transformer that leverages multi-resolution upsampling and visual representations from large-scale foundation models. SAM2Act achieves a state-of-the-art average success rate of 86.8% across 18 tasks in the RLBench benchmark, and demonstrates robust generalization on The Colosseum benchmark, with only a 4.3% performance drop under diverse environmental perturbations. Building on this foundation, we propose SAM2Act+, a memory-augmented architecture inspired by SAM2, which incorporates a memory bank and attention mechanism for spatial memory. To address the need for evaluating memory-dependent tasks, we introduce MemoryBench, a novel benchmark designed to assess spatial memory and action recall in robotic manipulation. SAM2Act+ achieves competitive performance on MemoryBench, significantly outperforming existing approaches and pushing the boundaries of memory-enabled robotic systems. Project page: sam2act.github.io.

3D Gaussian Splatting (3DGS) is a real-time rendering method that models 3D scenes as thousands of anisotropic Gaussians, which are projected and rasterized ("splatted") onto 2D screens to synthesize images. A core challenge is ensuring these splats render in correct depth order to resolve occlusions, traditionally requiring computationally expensive sorting. We propose eliminating sorting entirely to accelerate 3DGS while preserving visual quality. To achieve this, we explored clustering algorithms to group spatially coherent splats, avoiding explicit sorting. Among these, sequential k-means clustering emerged as a promising solution, achieving near-identical image reconstruction to ground truth while reducing computational overhead. By grouping splats into depth-ordered clusters, we bypass per-frame sorting without sacrificing accuracy. We are currently reimplementing the CUDA-based forward and backward rendering passes to integrate this cluster-first approach. This work demonstrates the potential of algorithmic redesign to unlock efficiency gains in modern graphics pipelines, with implications for scaling 3DGS to dynamic scenes in VR, simulations, and gaming.

Current at-home, minimal cost viral test kits are often limited to human-visible (colorimetric) readout methods which lack the same sensitivity achievable in laboratory settings that use complex equipment. We aim to develop a more accessible alternative by leveraging smartphone touchscreens to detect viral presence. Touchscreens emit an electrical field that changes when conductive materials interact with them. DNA has been shown in prior work to exhibit conductive properties based on its negative charge. Our approach utilizes a DNA replication reaction involving a thermostable polymerase, primers, dNTPs, and viral RNA as a template. If the template is present, amplification occurs, altering the capacitive response compared to a negative control. To validate this, we are testing the reaction on a vector network analyzer (VNA), measuring capacitive output changes directly on the sensor. We are also building and testing low-cost temperature controls to enable isothermal amplification. With the use of a Peltier heater, a temperature control sensor, we aim to speed up the reaction times and the use of a Pulse Width Modulation (PWM) power control system to ensure consistent reaction temperature. We are currently comparing active polymerase reactions to controls and plan to eventually transition these tests onto phone screens, creating a cost-effective, widely available diagnostic tool.

Molecular computing, which harnesses biomolecules such as DNA for computation, has rapidly advanced in the past two decades. DNA Strand Displacement (DSD) is a key molecular primitive used to implement molecular circuits. DNA’s predictable A/T C/G base-pairing enables precise control over molecule interactions. However, visualizing DSD processes remains a challenge. Current tools generate only static representations, making it difficult to illustrate reaction pathways and communicate complex molecular interactions effectively. This lack of clear visualization hinders collaboration among researchers and makes it difficult to communicate to those outside the field about the principles and potential of molecular computing. To address this, we have developed a Python package that automates the visualization of DSD reactions, generating both static and animated representations of DNA/RNA secondary structures. Using the Manim library from creator 3Blue1Brown, our tool takes as input DNA/RNA structures written in the widely-used dot-parenthesis notation and produces layouts and animations of the displacement events. Users can toggle between different layout and color modes that highlight features such as sequence and bonding probabilities, providing flexible options for different needs.

The Traveling Salesman Problem (TSP) is a classic NP-hard combinatorial optimization task with numerous practical applications. Classic heuristic solvers can attain near-optimal performance for small problem instances, but become computationally intractable for larger problems. Real-world logistics problems such as dynamically re-routing last-mile deliveries demand a solver with fast inference time, which has led to specialized neural network solvers being favored in practice. However, neural networks struggle to generalize beyond the synthetic data they were trained on. In particular, we show that there exist TSP distributions that are realistic in practice, which also consistently lead to poor worst-case performance for existing neural approaches. To address this issue of distributional robustness, we present Combinatorial Optimization with Generative Sampling (COGS), where training data is sampled from a generative TSP model. We show that COGS provides better data coverage and interpolation in the space of TSP training distributions. We also present TSPLib50, a dataset of realistically distributed TSP samples, which tests real-world generalization ability without conflating this issue with instance size. We evaluate our method on various synthetic datasets as well as TSPLib50, and compare to state-of-the-art neural baselines. We demonstrate that COGS improves distributional robustness, with most performance gains coming from worst-case scenarios.

Controlled and untethered Micro Aerial Vehicles (MAVs) near 1 gram offer transformative potential in applications like disaster response, inventory inspection, and precision agriculture, offering reduced costs and minimal hazards compared to larger drones. However, MAVs of this size face significant challenges in achieving both flight stability and maneuverability, particularly due to difficulties in generating sufficient lift and controlling multiple degrees of freedom mid-flight. While recent advancements have addressed various aspects of untethered flight, there has yet to be a MAV near 1 g that has also demonstrated stable hover and autonomous navigation. We introduce Coin-copter, a dual-rotor helicopter designed to overcome these limitations. We present three Coin-copter sizes, ranging from 0.8 g, to 1.1 g, and 1.8 g that leverage a foldable flybar-propeller mechanism for achieving passive stability and a feedback-controlled tail motor for yaw-axis control. Our prototypes achieve free-flight stabilization with payload capacities of up to 0.3 g, 2 g, and 5 g respectively, and evaluate the operational efficiency of each design to determine the optimal Coin-copter size for maximizing duty cycled flight time under practical energy harvesting scenarios.

DNA computing utilizes the unique properties of DNA molecules to process information while still in molecular form and enable the programmable control of matter at the nanoscale. However, a major limitation is the low reading bandwidth of DNA circuit outputs with fluorescent-based reporters, which hinders scalability and practical applications. Nanopore sequencing is an advanced DNA sequencing technology capable of rapidly detecting single molecules of DNA as they pass through a nanoscale pore, unlike traditional sequencing methods that require amplification. My research seeks to overcome this barrier by integrating DNA computing architectures with nanopore sequencing technology to achieve high-throughput readout and real-time monitoring of circuit kinetics. I am designing DNA-based reporters that encode DNA circuit outputs in a format compatible with nanopore sequencing. These reporters have distinct sequence signatures that can be efficiently read by Oxford nanopore sequencing devices, enabling high-throughput, real-time parallel sequencing. My work involves designing and engineering these reporters, validating their function through experimental assays, and optimizing their compatibility with nanopore platforms. By bridging DNA computing with nanopore sequencing, this research has the potential to expand the capabilities of molecular computing, making it more practical for real-world applications. Beyond computing, this approach could enhance biosensing and diagnostic technologies by enabling rapid and precise detection of molecular signals. For example, DNA circuits could detect specific disease biomarkers, with nanopore sequencing providing an immediate readout. Since nanopore sequencing is a more accessible and portable technology, it could be better deployed in low-resource settings, broadening DNA computing's impact on global healthcare and research. Ultimately, this work not only advances DNA computing but also has implications for broader fields such as DNA nanotechnology and personalized medicine.

Radio Detection and Ranging (RADAR) uses radio waves for object detection in applications such as air traffic control, radio astronomy, and defense systems. This project explores the feasibility of performing RADAR using Modulated Johnson Noise (MJN), which leverages the thermal noise inherent in electrical conductors to transmit information without the use of a conventional radio frequency (RF) carrier. Unlike traditional RADAR, MJN enables stealthier, low-interference operation and ability to function in areas with no ambient radio frequency. In this project we test the hypothesis that RADAR can be performed with MJN by transmitting a square wave signal made with two different noise levels and timing its reflection. To establish a proof of concept, the project follows a multi-phase approach. First, prior MJN research is reproduced by implementing a noise-modulated transmitting system using a Raspberry Pi, an RF switch board, and a Software Defined Radio (SDR) in an anechoic chamber. Next, signal control (transmit) and processing (receive) are integrated into a single microcontroller unit for synchronized operation. The electrical components for the receiving system are validated for amplification and filtering of the reflected signal. The antennas for transmitting and receiving the signal are selected based on their radiation pattern and optimal placement for the RADAR application. Once the transmit and receive systems are finalized, a microcontroller (ie. STM32 Nucleo board)  is used to synchronously transmit and receive reflected signals. Then, indirect time of flight methods are used for distance measurement by analyzing the phase shift between the transmitted and the received signal. The findings will contribute to the development of a RADAR system suitable for resource-constrained environments such as remote locations on Earth or in space and is beneficial for stealth operations where the object emitting the signal must be unidentifiable.