Over the course of the 2025-26 academic year, six faculty members across the UMass Amherst campus were named the recipients of prestigious five-year U.S. National Science Foundation (NSF) CAREER awards.

Image

The Faculty Early Career Development (CAREER) Program is a foundation-wide activity that offers NSF awards in support of early career faculty who have the potential to serve as academic role models in research and education and to lead advances in the mission of their department or organization.

Manning College of Information and Computer Sciences (CICS) professors Donghyun Kim, VP Nguyen, Marco Serafini, Juan Zhai and Chuang Gan were awarded CAREER grants for their work on making robots more graceful, brain-activity monitors more useful, AI data processing more efficient, AI agents more flexible, and AI-generated software more reliable, respectively. The awards for Gan, Kim, Nguyen, Serafini and Zhai bring the cumulative number of CAREER awards for CICS to 45.

The Riccio College of Engineering was awarded one CAREER grant this year, for Mariana Lanzarini-Lopes’s work on studying the properties of light, which brings the college’s cumulative total to 47 awards.
 

2025-26 NSF CAREER Grant Recipients
 

Image

Chuang Gan, assistant professor of CICS, has been awarded $599,814 to enhance situational flexibility in virtual assistants, robots and other embodied, artificially intelligent agents that interact with humans.

The current generation of embodied agents are good at what they have been pre-programmed to do, but they can only perform successfully in specific preconceived environments and situations. They like things neat and orderly. The human world, on the other hand, is messy and unpredictable, and that unpredictability can confound AI agents.

There are two main reasons for this: The agents don’t fully understand physical space, and open-ended social interactions tend to trip them up. This means they have a hard time learning physical skills and dealing with multiple people at once.

Gan is working to empower embodied agents with social and physical intelligence by incorporating physics engines and generative foundation models. Physics engines are software that simulates the physical laws of the real world, such as gravity and friction, and generative foundation models are neural networks inspired by the human brain that enable flexibility in machine learning.

“Our approach draws inspiration from cognitive science, which attributes human generalization and adaptability to internal mental models,” Gan said. The goal is to build a world model that allows embodied agents to learn from—and ultimately adapt to—complex situations.
 

Image

Donghyun Kim, assistant professor in CICS, was awarded $750,000 to help humanoid robots move more efficiently through intelligent control of an articulated toe.

The project centers around developing control algorithms to maximize energy efficiency and agility for robots with actuated toes. “Energy efficiency means that, by incorporating toes and using them effectively, humanoid robots could potentially operate much longer on the same battery charge,” Kim said. “If our hypothesis is correct, robots may be able to walk three times farther than they can today.”

That achievement could put currently out-of-reach real-world applications on the table. For example, smooth-moving robots with articulated toes could be used in dangerous or physically demanding areas such as disaster response, inspection of infrastructure and field operations.

“We are investigating a specific hypothesis: Humans can walk up to four times more energy-efficiently when they use toe push-off compared to walking without it. Similarly, humans can run more than twice as fast because toe contact at the end of push-off effectively extends leg length and increases propulsion,” Kim said. He is working to harness that push-off movement for robotics.

Past projects in the field tried to achieve energy efficiency using actuated toes, but adding mechanisms to the toes weighed down the robots’ legs and made them too complex, Kim noted. “This highlights a fundamental challenge in robotics: A good idea does not automatically translate into better performance,” he said. “Robots are complex systems that combine mechanics, electronics, control and intelligence. Only when all of these components are carefully integrated can the true value of an innovation be realized.”

In the end, this project could enable more graceful humanoid robots as well as a greater understanding of human locomotion and design, learning and optimization for robotic systems.

“System-level research in humanoid robotics is inherently difficult,” Kim said. “It is somewhat like saying, ‘I want to win a Formula 1 race, so I’ll start by designing an engine.’ Success depends on much more than any single component; every part of the system must work together seamlessly.”

It’s a team effort, Kim noted. “Many students and collaborators have invested enormous effort into making this ambitious project possible,” he said. “I am especially grateful for CICS’s support of this nontraditional computer science research, which often appears more like engineering because of its strong emphasis on hardware development.”
 

Image

Having developed a special glass that uniformly emits low levels of UV radiation capable of keeping underwater surfaces clear of biofilm, Mariana Lanzarini-Lopes, assistant professor of civil and environmental engineering, has been awarded $709,645 to turn her technology into a platform that explores how light interacts with cells.

One of the challenges of light-driven research is controlling the light itself. Lanzarini-Lopes’s invention of UV-emitting glass can prevent biofilm formation by 98% on marine surfaces like lenses or ship hulls. While this particular work is being advanced as a patent, she realized that she also has a platform that can fine-tune the properties of light—type, exposure frequency dose and time, wavelength, cycling, intensity—across a uniform surface.

“We’ve demonstrated this cool tech, but now we’re taking a step back and saying, ‘Let’s spend the next five years really understanding what’s fundamentally going on so that we can drive a much wider movement of light-based treatment forward,’” says Lanzarini-Lopes.

With this CAREER award, Lanzarini-Lopes will develop a platform that allows the precise control of the properties of light. This will answer fundamental questions of how light interacts with cells, both to eradicate biofilms as well as promote cell growth, with applications across environmental engineering, phototherapy and photobioreactors.

The ultimate goal is to develop more effective methods for manipulating cell behavior with light. Lanzarini-Lopes’s research has focused on the inactivation and prevention of biofilm, so she gives preventing biofouling as an example. “There’s no one-size-fits-all solution, but knowing the limitations and why things are or are not working allows us to then take other methods and use them in synergistic manners,” she says.
 

Image

VP Nguyen, professor in CICS, was awarded $686,250 to develop wearable brain monitors that collect data on brain activity in everyday life, which will be especially useful for detecting and managing medical conditions such as seizures and Alzheimer’s.

“To understand what is happening in someone's brain today, a patient usually spends days in a hospital, wired to bulky equipment, waiting to capture an event like a seizure,” Nguyen said. “It is uncomfortable, expensive, and it records only a short and artificial slice of a person’s life.”

And if a seizure doesn’t happen during that time, the data might not even be useful for clinical purposes.

Nguyen is looking to change that by building wearable devices that continuously track physiological signals throughout a regular day—and for months at a time, giving clinicians more data to diagnose and manage neurological conditions.

Traditional computing models rely on processors that “constantly fetch, compute and store, which is wasteful for the sparse, bursty signals the body produces,” Nguyen said.

“Processing this kind of data in real time normally drains a battery within hours, and sending it off to the cloud to make a decision adds delay that matters when you are trying to catch a seizure as it happens,” he said. “We solve both by designing computing that works the way the brain does, staying quiet until something meaningful happens.”

This biologically inspired process is called neuromorphic computing: integrating processing and memory, allowing the system to process data only when needed instead of running continuously.

Ultimately, the monitor will enable “continuous, comfortable physiological monitoring in everyday life,” Nguyen said. “That means catching seizures, or warning of them before they happen, so a person or caregiver can prepare. It means spotting early signs of Alzheimer’s and cognitive decline sooner than a clinic visit would. And because the device makes decisions on its own, this can work with real-time responsiveness and privacy.”

Nguyen is collaborating with industry and national labs including NextSense Inc., SynSense, Dolby Laboratories (Sight Labs), Oak Ridge National Laboratory and Earable Neuroscience to create and test the brain-activity monitors for clinical use. Clinical partners include Felicia Chu at UMass Chan Medical School, Jiang Yang at the University of Kentucky College of Medicine, and Laura Brattain at the University of Central Florida College of Medicine.
 

Image

Marco Serafini, associate professor in CICS, was awarded $666,930 to develop new software tools that will make AI-powered information processing easier and faster.

Data—and, by extension, data processing—is essential to just about every human endeavor in our modern world: Consider medical records, customer data, legal documentation, scientific data, and so on. Processing all that data is more important every day…and more unwieldy. Artificial intelligence can help, but applying AI to large databases can be slow and computationally costly. This project aims to remove this major barrier to the adoption of AI.

“A lot of valuable data is stored as natural language text, images, or other types of ‘unstructured’ data that computers have traditionally struggled to understand,” Serafini said. “This is why we want to use AI for data processing.”

Serafini’s project will build software that can be incorporated into existing databases to speed up large-scale data processing by taking advantage of AI. “The project will address three complementary directions: making predictions as data changes over time, finding data using text or images as queries, and answering more complex natural language queries that may require correlating information from multiple datasets,” he said.

Traditional database management systems use query languages such as SQL to make data analysis accessible to non-experts, Serafini explained. “With SQL, users only need to describe the information they want and delegate the efficient execution of their queries to the database system,” he said. “This project will develop tools that extend SQL-like languages with AI-powered operators and automatically choose an efficient and scalable way to execute queries.”

This is a new way of thinking about data analysis, Serafini said. “AI models require organizing, storing and processing data in a way that is fundamentally different from traditional data analysis tools,” he said. “Rather than representing data explicitly, they use numerical representations that capture important features of the data. Managing this type of data at scale requires novel solutions.”

The hardware used for such computations is specialized, Serafini added, so it’s a challenge to write software that optimizes the hardware and lowers costs in terms of money and energy.

“These challenges require a new kind of data infrastructure that rethinks traditional system design principles,” he said.

Ultimately, this project will be released as open-source software for use in a variety of applications.
 

Image

Juan Zhai, assistant professor in CICS, was awarded $689,531 to make computing systems more trustworthy by ensuring reliability and security in modern software, machine learning systems and agentic AI systems.

Software developers today employ autonomous AI systems to create, edit and assess code. Zhai compared the process to using an AI construction team to build a house: “AI coding tools can build quickly, but fast does not always mean correct or safe,” she said. “My project helps AI tools take what a person wants to build, create clear software ‘blueprints’ and ‘inspection rules’ from it, and use those rules to write and check code so it behaves as intended.”

Checking the code has been a difficult step so far. Engineers currently struggle to determine whether the software behaves as intended, because formal specifications—which use rigorous math and logic to define a system’s behaviors and requirements—are difficult to generate and maintain. 

“Formal specifications are hard because they need to describe complex software behavior precisely and concisely,” Zhai said. “They must be detailed enough to guide the AI, but not so detailed that they become another version of the code. This is especially hard when AI agents quickly change many connected parts of a project, because the specifications must keep up with the code, documentation, dependencies and developer intent.”

Zhai is working to build a resource that will link plain-language descriptions with code and precise rules for system behavior; those rules will be automatically evaluated and updated as the software evolves, making AI-generated code more dependable.

“AI is changing how software is built,” she said, “but the key question is still the same: Does the software do what we meant it to do? I see formal specifications as a bridge between human intent and AI-generated code. They can help make AI-assisted development more precise, reliable, and trustworthy.”

Zhai’s work will make autonomous AI-based systems safer, with fewer expensive failures. This will lead to more reliable software systems in areas such as finance and healthcare, where public trust is especially important.

It will also make software development more accessible. “Students, small businesses, public agencies and community organizations could use AI coding tools with more confidence,” Zhai said. “Many people have useful software ideas, but not the technical background to build reliable systems. I hope this project can help AI turn those ideas into software people can actually trust.”