The evolution of mobile and wireless networks is being driven by three major technological advances: a move to higher carrier frequencies (e.g., mmWave/THz), the increased density of task-oriented network entities, and a higher level of intelligence. To guarantee consistently high-quality services for users amid this evolution, AI-native network resilience and self-reconfigurability are crucial. This project pursues a novel resilience-native network paradigm called CatFly, inspired by the way a cat behaves during a crash landing. If a network system is co-designed with an analogous ability to predict disturbances or failures ahead of time, automate prepared countermeasures, and reconfigure physical structure to prevent losses, it can increase self-resilience to sustain performance in disruptive situations. To this end, the project explores hybrid digital-physical intelligence that uses an evolutional digital replica of the network with AI-empowered analysis to uncover multi-dimensional physical reconfigurability.

This project introduces a framework that leverages large language models (LLMs) to automate and orchestrate multi-layer network simulations traditionally performed with distinct and complex tools. By integrating LLMs as agentic intermediaries, the framework simplifies user interaction across the physical, packet, and protocol layers of simulation. Each simulator is paired with a customized network-oriented LLM, trained via parameter-efficient fine-tuning and retrieval-augmented generation, to interpret natural language commands. A centralized LLM-based scheduler determines task execution order while an orchestrator oversees end-to-end coordination and data flow. Using NVIDIA’s Sionna and complementary simulators, the solution enables researchers to model intricate networks without writing scripts or mastering tool-specific configurations.

As an advanced layer of cyber-physical systems, hybrid data-driven emulators, predictors, and optimizers play a crucial role in the National Artificial Intelligence Research Resource initiative by acting as intelligent models that replicate physical systems and processes. These virtual experimental counterparts allow researchers to optimize complex networked systems in a risk-free environment, accelerating discovery and innovation. This project lays the foundations of a test platform by developing multi-dimensional data generation and continuous modeling approaches that merge tools from machine learning, communication theory, and distributed optimization to advance AI-driven cyber-physical systems across different areas.

This project examines how children can learn to recognize AI hallucinations by creating their own chatbots. By engaging children in chatbot design and testing, the study explores how hands-on AI creation can support children’s AI literacy, critical evaluation skills, and understanding of when AI systems may produce inaccurate or misleading responses.

This project focuses on developing autonomous AI agents that enhance safety and trust across digital platforms. The research explores how AI systems can proactively detect and investigate platform-specific policy violations, including AI-generated scams on social media, counterfeit products on e-commerce marketplaces, and other forms of deceptive or harmful content. By combining multimodal content analysis, automated reasoning, and platform-aware decision making, these agents can assist operators in identifying emerging threats, enforcing platform policies, and improving the overall integrity of online ecosystems.

This project explores the design of AI-powered assistants that help users securely manage and configure smart home devices. The goal is to enable non-expert users to better understand the security and privacy implications of connected technologies through personalized guidance, risk assessments, and actionable recommendations. These assistants can identify device vulnerabilities, recommend appropriate privacy and security settings, answer technical questions, and help users make informed decisions that align with their preferences and risk tolerance.

This project examines the privacy implications of integrating generative AI technologies into applications, services, and connected ecosystems. We investigate how user data is collected, shared, and processed across AI-enabled systems, identify emerging privacy risks, and develop techniques to improve transparency and user control. Our goal is to create privacy-preserving architectures and design principles that enable organizations to build trustworthy AI ecosystems while protecting sensitive user information.

This project is based on the 1:10-scale autonomous racing platform. We design machine learning-based and model-based control algorithms for racing, analyze scheduling algorithms for ROS2 and the CPU-GPU platform, and verify system correctness via timing analysis and model checking. A series of works have been published in premier venues such as RTSS, EMSOFT, RTAS, DAC, and TIST. Our F1-Tenth team ranked 4th in time-trial and 5th in head-to-head racing at ICRA 2022, and 1st in total number of laps and 3rd in time trials at CPS-IoT Week 2023. This project has been supported by NSF CNS, NSF SHF, NSF CPS, NVIDIA, and internal grants at NC State.

This project aims to integrate wireless sensing, tiny wearable sensors, machine learning, real-time control, and robotics into a ubiquitous healthcare system. Various cyber-physical systems core problems and techniques are correlated, such as motion capture, motion prediction, environmental sensing, real-time machine learning, data transfer, and knowledge distillation. For the short term, we target people with lower limb amputation or heart disease. A series of works have been published in premier venues such as ICCV, AAAI, IJCAI, JBHI, and TNNLS. This project has been supported by NSF FRR, NSF FW-HTF, NHK, and internal grants at NC State.

The overall aim of this work is to devise a co-simulation environment for autonomous vehicle certification and control with well-defined requirements and standards. The objective is twofold: to create an environment supporting seamless transitions from lab testing through development in a hybrid environment to fully deployed road testing; and to systematically identify minimum requirements and testing scenarios for lab, hybrid, and physically deployed autonomous vehicles that support development cycles and facilitate certification.

Animal-Assisted Interventions (AAIs) are goal-oriented programs that intentionally incorporate animals, such as dogs, for therapeutic benefits. AAIs are widely used in a variety of settings, including for cancer patients and veterans with post-traumatic stress disorder. The overall objective is to develop, test, and evaluate an Internet of Things software and hardware system for dyadic physiological monitoring of humans and animals in AAI settings, and to innovate in analytic methods for interpreting the data. Our team has developed and tested a platform of wearable, wireless sensors that simultaneously gather physiological data from both humans and dogs involved in AAIs, combined with psychological and symptom data collected from the patient and dog handler.

Since 2014, we have been working with partners in the guide dog industry in the U.S., Canada, and Australia to collect objective measurements of candidate guide dog behavior. Using these measurements, we have developed analytic techniques to predict temperament across 29 dimensions known to be relevant for service work.

We aim to address three key limitations in Potential-Based Reward Shaping (PBRS), a technique used in reinforcement learning to improve learning in sparse-reward environments without altering optimal policies: its limited scope in covering all optimality-preserving shaping functions, its practical challenges in hand-designing effective potential-based rewards, and its inapplicability to reinforcement learning with human feedback scenarios where the true reward is unknown and being estimated. We are developing a class of plug-and-play methods that require no manual design of shaping rewards and are easily adaptable to existing shaping rewards or intrinsic motivation terms.

The NSF AI Institute for Engaged Learning creates AI-driven narrative-centered learning environments, adaptive collaborative learning technologies, and multimodal learning analytics through advances in natural language processing, computer vision, and machine learning.

This project develops causal models to represent the complex, nonlinear dynamics of open-ended learning environments, where learners face loosely defined problems with many possible solution paths. Using a novel framework called the Formative Exploration Support Template (FEST), the causal model captures learner actions, evolving system states, and learning goals as structured nodes, and serves as the computational foundation for generating targeted formative feedback. This is among the first work to use causal modeling to support and scaffold exploration in open-ended, complex systems learning environments.

This project applies computer vision and machine learning to a socio-environmental board game simulating real-world climate policy deliberations, called FutureScape. The project designs an AI pipeline that detects and analyzes player actions in real time from video data, extracts decision-making intelligence, and integrates it with established climate models to visualize the cascading spatio-temporal consequences of collaborative choices. The long-term goal is to use this dataset of human decision-making in high-stakes environmental contexts to train a custom human-in-the-loop AI system.

BioVerify is an AI-driven vision pipeline that validates species identification in crowdsourced predator-prey interaction images to support the construction of large-scale ecological networks. Developed in conjunction with the Who Eats Whom project on iNaturalist, it addresses a core challenge in citizen science data: photographs of feeding interactions contain multiple species, often partially visible, that existing models cannot reliably identify. BioVerify combines open-set object detection and hierarchical taxonomic classification to localize and verify each organism before admitting the observation into the food web. Rather than discarding uncertain cases, the pipeline triages them for expert review, preserving ecologically significant records at scale.

The overall aim of this work is to devise a co-simulation environment for autonomous vehicle certification and control with well-defined requirements and standards. The objective is twofold: to create an environment supporting seamless transitions from lab testing through development in a hybrid environment to fully deployed road testing based on AI-driven analysis and controls; and to systematically identify minimum requirements and testing scenarios for lab, hybrid, and physically deployed autonomous vehicles that support development cycles and facilitate certification.

This project aims to devise efficient tensor networks, especially for sparse data, which are prevalent in many real-world applications, including machine learning. The impacts of the project encompass improving data compression, computation, memory usage, and interpretability of tensor networks; fostering enduring and collaborative partnerships among academia, national research labs, and industry; and broadening education avenues by designing relevant new courses, training undergraduate and graduate students, organizing workshops, and enhancing K-12 outreach.

The primary goal of the QACTI quantum system and technology demonstrator is to build an advantage-class trapped-ion quantum computer capable of being used by the broader scientific community remotely. The secondary goals are to discover algorithms suited for near-term quantum computers, including quantum machine learning; to improve and democratize ion trap quantum technology; and to develop a workforce capable of utilizing and building advantage-class machines. These goals are achieved by performing device-oriented experiments at fine-grained control levels not available on commercial platforms, contributing to a combined hardware and software stack in an open-source manner.

Typos is a computer science education platform designed to provide students with lower-level practice opportunities while learning computer science. Research opportunities include developing recommendations for struggling students and data analysis on student clickstream data.

This project develops multimodal AI for precise, hands-free gesture understanding. It combines acoustic and optical sensing, then uses learned fusion methods to align and integrate the two streams so the system can distinguish subtle finger motions that are difficult for either modality alone. Focus areas include multimodal representation learning, temporal alignment, and robustness to weak observations, occlusion, and user-to-user variation.

This project develops AI-enabled anti-spoofing for fingerprint authentication. It reconstructs a 3D haptic response from a short sequence of fingerprint images using photometric stereo, then combines rotation-invariant texture descriptors, fingerprint minutiae retrieval, and supervised and unsupervised classifiers to learn the difference between live fingertips and fabricated spoofs across materials and attack conditions. Focus areas include biometric anti-spoofing, 3D haptic sensing, photometric stereo, texture-based machine learning, and live-versus-spoof classification.

This project develops mmWave speech sensing for voice interfaces by learning speech-relevant patterns from RF reflections caused by vocal articulation. It combines mmWave signal processing with AI to convert weak throat and mouth motion into robust speech representations, improving inference when microphones are degraded by environmental noise. Focus areas include RF representation learning, articulatory motion modeling, noise-robust inference, and voice user interfaces.

This project co-designs AI workflows and satellite onboard systems to enable real-time Earth monitoring and analytics, and investigates the design and operations of future orbital data centers operated by companies like SpaceX and NVIDIA. Orbital data centers are next-generation AI computing infrastructure that harnesses solar energy in space and delivers real-time intelligence to global users.

This project aims to investigate cutting-edge quantum technologies, design next-generation quantum computing and communication infrastructures, and enable quantum intelligence for classical and quantum problems. The project combines quantum information science, quantum system engineering, and AI, and aims to solve the hardest problems with near-term quantum resources.

This project develops efficient and adaptive inference methods for large language models and generative AI systems. A representative system, Adaptix, accelerates LLM decoding through adaptive draft-verification without requiring model fine-tuning, enabling faster generation while maintaining output quality. The broader goal is to make large generative models more practical for real-world deployment under latency, resource, and computational constraints.

This project develops agentic AI platforms for scientific discovery, with an emphasis on microscopy, imaging, and complex scientific data analysis. A representative system, SIMBA, integrates large language model-based agents with single-molecule localization, spectral processing, deep learning-based denoising, and automated workflow orchestration. The project aims to make scientific image analysis more scalable, interactive, reproducible, and accessible through language-guided AI agents.

This project builds trustworthy generative AI systems that support teaching and learning while maintaining transparency, accountability, and educational rigor. A representative platform, MerryQuery, provides AI-powered support for educators and students through retrieval-augmented and human-centered interaction workflows. The project studies how generative AI can be responsibly integrated into real educational settings to provide personalized assistance, improve learning support, and support instructors at scale.

Measuring and visualizing sentiment in social media posts from numerous visual perspectives, allowing the general public to query social media sites for posts by keyword, then presenting those posts and the relationships they form using a web-based visualization dashboard.

We are developing Discrowd, a model of groups and crowds of rich individual virtual characters with small LLM-based implementations. Each individual agent in a Discrowd simulation includes a personality model based on 16 personality types, back-stories, and roles. The social agent framework is deployed through a Discord server where initialization adds individual bots and assigns them locations. An agent-based social simulation makes individual agents interact with other agents in the same location using social cues. We demonstrate the framework through a prototype simulation of a small coastal town and its response to hurricane and flood alerts, tracking opinion changes about emergency response across different message types.

Safe reinforcement learning commonly expresses safety requirements through constraints, which are then incorporated into the agent via Lagrangian relaxation. We formally demonstrate that this relaxation inherently biases the agent toward trajectories that keep it safe at the cost of optimality, an issue we name the Lagrangian hack. To address this limitation, we introduce a modified safe boundary objective and an approach based on reward shaping, SaBER, that modifies the reward perceived by the agent to favor enforcement of the constraint over reward maximization. Agents evaluated with SaBER consistently find the optimal policy near the cost threshold with fewer constraint violations.

SimStudent is an interactive learner that applies inductive logic programming to learn cognitive skills from examples. SimStudent is customizable, or programmable, so that we can test various hypotheses to identify factors that affect learning and test when and how such factors affect learning. The learning outcome, a set of learned cognitive skills, is represented as production rules that are also human readable.

The goal of the PASTEL (Pragmatic methods to develop Adaptive and Scalable Technologies for next generation E-Learning) project is to develop evidence-based methods for efficient and practical learning engineering. In particular, we are interested in developing advanced technologies to build adaptive online courseware.

This project develops an AI-driven wearable cardiac monitoring framework that converts photoplethysmography (PPG) signals collected from wearable devices, such as smartwatches, into multi-lead electrocardiogram (ECG) signals. The system uses a conditional diffusion model to generate single-lead ECG from PPG and an LSTM-based deep learning model to predict multi-lead ECG signals, enabling richer cardiac information from passive, continuous wearable sensing. By combining generative AI, physiological signal modeling, and wearable health sensing, the project aims to support daily-life cardiac monitoring and assist in detecting abnormal cardiac patterns.

This project develops an adaptive multimodal AI framework to monitor medication adherence in Parkinson’s disease using smartphone sensor data. The system analyzes gait and voice signals, transforms them into spectrogram-based representations, quantifies the reliability of each sensing modality, and dynamically fuses multimodal information to infer whether a patient has taken medication. By reducing reliance on self-reporting and single-modality sensing, the project aims to support more accurate, passive, and scalable medication adherence monitoring in patients’ daily lives.

This project develops an AI-driven brain-computer interface that reconstructs speech from low-channel EEG signals to support people with communication disorders. The system uses deep learning models, including a generator-discriminator architecture, to learn complex mappings between brain activity and speech representations. It also applies AI-based signal processing and channel optimization methods to reduce the number of required EEG channels from conventional high-density setups to as few as four to six channels while maintaining strong speech reconstruction performance.

As AI is integrated into courses in computer science and more broadly across institutions of higher education, we need a way to assess students’ literacy in using AI tools to support disciplinary work. We are developing an inventory to assess student AI literacy. This project is in collaboration with David Roberts.

Riot is our ongoing effort, building on extensive previous research, to enhance and synthesize three directions in tools for engineering multiagent systems. For coordination, it pursues declarative information protocols with verifiers and runtime support. For meaning, it pursues formal models of deontic norms with query processors. For agents, it pursues approaches to reasoning about interaction given stakeholder goals that may be realized on diverse infrastructure. Riot also pursues tools for authoring protocols and norms and for bridging to legacy agents.

This NSF project is investigating ways to develop a sociotechnical system for affordable public microtransit to connect suburban and rural populations to employment, healthcare, and other services. The goal is to identify, test, and evaluate technologically enabled, community-supported solutions for distributing travel demand for on-demand public microtransit in an equitable manner. It investigates how artificial intelligence can facilitate prosocial behavior at a low cognitive burden for the user. Our community partner is the City of Wilson, N.C., which has been operating microtransit service since September 2020.

This project, dubbed Mae for Multiagent Ethics, develops a computational model of a sociotechnical system that captures how ethical concerns arise in the mutual interactions of multiple stakeholders. These foundations would enable us to realize ethical sociotechnical systems that incorporate social and technical controls to respect stated ethical postures. The project focuses on modeling stakeholder values, multiagent simulation to understand norm emergence and equity, and participatory design of sociotechnical systems.

AI-LEAF is one of the National AI Research Institutes funded by NSF and NIFA that is specifically focused on agriculture and forestry. Our team is developing large-scale deforestation mapping approaches using deep learning and foundation models.

This project is developing deep learning and foundation model-based methods for semantic segmentation of clouds, shadows, and sea ice in remote sensing imagery, along with knowledge-guided machine learning approaches for imputing missing values.