Diagnosing hidden failures,
then guaranteeing against them

Research trajectory

I am an undergraduate researcher in Applied Computer Science and Mathematics (Honors) at Iowa State University. Across robotics, computer vision, and reinforcement learning, my research is driven by a single question: how can we trust an autonomous system?

That question has two complementary parts. First, can a robot trust what it perceives? Second, can we trust the decisions it makes? My work explores both problems because they ultimately depend on one another. Reliable autonomy requires trustworthy perception as much as it requires safe control.

I did not begin my research asking these questions explicitly. I began by building systems that had to work.

As a high school researcher in Drake University's computer vision laboratory, and later as a first-year researcher at Iowa State University's Virtual Reality Applications Center (VRAC), I developed a digital twin of a robotic fringe projection profilometry (FPP) scanner in NVIDIA Isaac Sim. That work became VIRTUS-FPP, a physics-faithful virtual sensor published in IEEE Sensors Journal that generates realistic synthetic data for optical metrology. More importantly, it taught me the value of simulation as a scientific instrument. When every physical parameter is controllable and every measurement has known ground truth, simulation becomes a powerful environment for understanding, not just evaluating, machine learning systems.

Working in that setting revealed a deeper problem. During our study of long-range single-shot FPP, we discovered that a neural network producing strong quantitative results had learned to rely primarily on object shape priors rather than the projected fringe information it was intended to interpret. The model appeared successful, but for the wrong reason. I led work to diagnose this shortcut and restore the underlying imaging physics by integrating differentiable physical models directly into the learning pipeline. That experience fundamentally changed how I think about machine learning: high benchmark performance alone is not enough if we cannot explain why a model succeeds.

While that work focused on perception, it naturally raised the complementary question of control. During my internship at the U.S. Naval Research Laboratory, I shifted toward reinforcement learning for autonomous high-altitude balloon swarms. There, I worked on communication-efficient safe reinforcement learning, developing a framework that jointly learns both control actions and adaptive sampling intervals while remaining protected by a Lyapunov-based run-time assurance shield. The resulting system reduced control updates by 3.51× compared to classical approaches while preserving formal safety guarantees. Although the application differed, the underlying philosophy remained the same: rather than hoping a learned system behaves correctly, design it so that critical failures are impossible by construction.

These experiences gradually brought the two halves of my research together. Today my work sits deliberately at the intersection of trustworthy perception and assured control, combining reliable representations of the world with controllers whose behavior can be formally bounded. Seventeen publications, invited talks, and collaborations across academic and defense laboratories have all reinforced the same conviction: trustworthy autonomy requires understanding both what autonomous systems perceive and how they act.

How perception and control connect

Although perception and control are often studied independently, they are fundamentally inseparable.

A controller can only be as reliable as the information it receives. A run-time assurance shield may certify that a policy is safe for a given state, but if the perception system provides an incorrect state estimate, the certification no longer reflects reality. Conversely, even perfectly accurate perception cannot guarantee safe behavior if the controller itself lacks formal safety properties.

This dependency motivates my current research. I investigate methods that improve the reliability of learned perception while simultaneously developing control algorithms with provable safety guarantees. Current projects include language-conditioned robot navigation that aligns vision-language representations with the latent space of privileged reinforcement learning policies (LCLA), as well as safe offline reinforcement learning from preferences generated by vision-language models.

Across both areas, the guiding principle is consistent: identify the hidden ways learned systems fail, then design architectures that prevent those failures rather than merely measuring them afterward.

Motivation & long-term goals

Autonomous systems are rapidly moving beyond controlled laboratory environments into applications ranging from aerial robotics and environmental monitoring to autonomous transportation and collaborative manufacturing. As these systems become increasingly capable, the central challenge is no longer whether they can perform a task; it is whether they can do so reliably under uncertainty.

My long-term research goal is to develop a unified framework for assured autonomy, combining perception whose failure modes are understood and mitigated with controllers whose behavior satisfies formal safety guarantees. I hope to pursue this agenda through a PhD and a research career at the intersection of robotics, machine learning, and control theory.

Research philosophy

• Trustworthiness over leaderboard scores. A model that performs well for the wrong reason is ultimately unreliable. Understanding failure modes is often more valuable than achieving marginal improvements on benchmark metrics.
• Guarantees over heuristics. When safety matters, guarantees should be embedded within the system itself through physical models, Lyapunov certificates, and run-time assurance rather than relying solely on empirical performance.
• Simulation as a scientific instrument. Physics-faithful simulation provides controllable experiments, repeatable evaluation, and trustworthy ground truth, making it indispensable for studying learning-based systems.
• Autonomy is an integrated system. Perception and control cannot be treated independently. Reliable autonomous systems require guarantees that compose across the entire pipeline, from sensing to decision-making.

Collaborations

My research has been shaped by collaborations with mentors across academia and government research laboratories, including Dr. Cody Fleming (Iowa State University, safe reinforcement learning and autonomy), Dr. Beiwen Li (Iowa State University, optical metrology and computational imaging), Dr. Erick J. Rodríguez-Seda and Dr. Tristan Schuler (U.S. Naval Research Laboratory, run-time assurance and autonomous systems), Dr. Peng Li (DEVCOM Soldier Center, human body modeling), Dr. Alimoor Reza (Drake University, computer vision and time-series analysis), and Dr. Soumik Sarkar (Iowa State University, vision-language robotics).

Background

I entered Iowa State University with junior standing after completing two associate degrees during high school, and I have been continuously involved in research since my first semester. My research journey began in Drake University's computer vision laboratory while I was still in high school and has since expanded across multiple universities, government laboratories, and interdisciplinary collaborations.

Beyond my own research, I serve as an Undergraduate Research Ambassador, mentoring students who are beginning their research careers and encouraging undergraduates to engage in meaningful scholarly work early in their education.

Whether writing papers, developing simulation environments, or designing learning algorithms, I try to approach every project with the same question that first motivated this research trajectory: how can we build autonomous systems that deserve to be trusted?