Cells dynamically organize, sort, and synthesize biomolecules through constant crosstalk between organelles, largely governed by their electrostatic and chemical interactions. Such organelles can be reconstituted synthetically, using DNA structures to mimic cellular function. DNA nanotechnology uses sequence programmability and base-pairing to create synthetic systems that mimic cell machinery. In this project, we focus on two classes of DNA-based nanostructures with comparable physical properties to those in living cells: condensates, mimicking membraneless organelles, and nanotubes, mimicking cytoskeleton filaments. Condensates form by the interaction of DNA branched motifs, nanostars, while nanotubes consist of tile motifs. A critical step towards biomedical applications of this technology is understanding how these structures interact and organize in cells. We explore the electrostatic interactions of condensates and nanotubes in protocells, i.e., emulsion droplets, to investigate how these structures form and interact in a cell-like environment. We encapsulated DNA tiles and nanostars in water-in-oil droplets and observed them over time with varied tile, nanostar, and divalent salt concentrations. We also introduced a UV-responsive hairpin linker that blocks sticky end hybridization until irradiation, enabling on-demand, light-controlled assembly of our nanostructures. From this, we discovered condensates can improve the assembly kinetics of nanotube networks. Nanotube morphology also changes, forming branched networks in larger droplets and rings in small droplets. However, increasing salt or tile concentration induces nanotube network branching in smaller droplets. Understanding how nanostructures function in different conditions advances the development of nucleic acid assemblies for future applications including in synthetic organelles, molecular scaffolds, and engineering cell function.

End-to-End Robot Learning has emerged as an effective paradigm in contemporary robotics, largely propelled by advancements in computational power, computer vision, and AI architectures. Imitation Learning(IL) enables teaching robots intricate, human-like skills. Nonetheless, conventional IL are confronted with challenges, notably multimodality, covariance shift, and error compounding. Strategies from prior research like multi-head transformer ACT and diffusion policies typically impose stringent requirements for high-quality demonstrations—often necessitating hundreds of precise and diverse expert trajectories. Furthermore, their applicability is frequently restricted to sequential actions, and their performance notably degrades as the task horizon expands.    Inspired from the human learning principles: complex tasks are decomposed into series of manageable subtasks and learning involves dynamic, selective attention to environmental cues while filtering out irrelevant information, we propose a novel and robust approach to robot imitation learning: gaze-attention-guided policy learning. Our methodology starts by directly capturing human gaze during expert demonstrations. We train a powerful gaze generative model, which predict precisely where the robot should focus its attention(gaze) at inference. We introduce a powerful mechanism that integrates this predictive gaze information, guiding the robot's policy model to discern the specific subtask currently being executed. Our experimental results demonstrate that the incorporation of our generative gaze model enables the learning of non-sequential actions from a more flexible set of demonstrations, and show better performance on long-horizon tasks.    Key Words: Robotics, Imitation Learning, Human-Robot Interaction, Egocentric Vision

Nanoparticles of metals such as gold and platinum are prevalent in biosensing technologies due to their abilities to enhance electrical conductivity and facilitate aptamers to react with target molecules. They immobilize the aptamers (DNA strands that attach to target drug molecules) onto the microneedle tip, and allow for greater amounts of reaction to occur due to the enhanced surface area. In this project, we investigated the material properties of the nanoparticles using electrochemical techniques of amperometry, cyclic voltammetry and square wave voltammetry. After depositing metal nanoparticles onto the microneedles, we measure the current activity by inserting them into a redox solution. Firstly, we evaluated the consistency of performance of the microneedles to sense electricity by measuring the current with different concentrations of analyte. Then, we compared the sensing abilities for microneedles that were acid cleaned and non-acid cleaned using square wave voltammetry, evaluating current peak differences over multiple concentrations. Lastly, we sought for the ideal scan rate (applied voltage per second) for cyclic voltammetry, by identifying the range that fits the Randles-Sevcik equation, where peak current is proportional to the square root of scan rate. Results have shown that current was proportional to the chemical activity of the analyte, showing consistent results across multiple trials. In addition, the acid cleaned needles have exhibited significant increases in the peak differences as concentration increased, proving to be effective in metal microneedles. Lastly, the higher scan rates were proven to be more effective, with a more linear relationship between the current and square root of scan rate (R2>0.98).   

More than 5 million people in the United States live with paralysis and many others struggle with  various types of motor disabilities. Shared autonomy enables individuals with disabilities to  perform everyday tasks by combining human input with robotic assistance. However, current  systems often depend heavily on gaze tracking and brain–computer interfaces, which can cause  user fatigue and struggle to generalize across diverse tasks. This study introduces a shared  autonomy framework that integrates gaze-based robot navigation with electromyography (EMG)  signals to trigger basic robot actions, aiming to reduce fatigue and improve adaptability. A  prototype was evaluated on a multi-step task: picking up bowls, pouring ingredients into a pot,  and returning the bowls. Compared to a purely gaze-based baseline (91.7% success rate), the  proposed system achieved a 58.3% success rate but reduced task completion time by 43.7%.  While performance improvements are needed, especially because the pipeline is not finalized,  the results indicate potential for faster task execution and reduced user strain, highlighting the  promise of combining gaze and EMG inputs in assistive robotics.