April 2025
NSF I-Corps AI/Robotics Teams 

A4E

Enhance student learning by leveraging AI agents to streamline access to materials, improve lesson clarity, support ongoing review, and simplify quiz creation and exam generation.

 Mohamed Farag, Carnegie Mellon University, Faculty

Active AI

As AI continues to transform our world, it is critical that future leaders, policymakers, and innovators possess AI literacy to navigate its complexities and societal impact. However, there is a significant gap between advanced AI educational resources and those that can be easily implemented by K-12 educators. 

To address this challenge, we provide scalable, interactive AI literacy modules designed for K-12 teachers and students. These bite-sized online learning modules equip students with foundational AI knowledge and help teachers effectively integrate AI education into their classrooms. By making these resources accessible and engaging, we aim to prepare the next generation to understand, use, and critically engage with AI, ensuring they are ready to make informed decisions in a rapidly evolving world.

Ruiwei Xiao, Carnegie Mellon University, Graduate student
Ying-Jui Tseng, Carnegie Mellon University, Alumni

ALEO

With 65% of mothers with children under 3 in the workforce—double the rate of the 1970s—parents need reliable, intelligent support at home. Parenting is rewarding but also exhausting, and not all families can afford a full-time nanny, which costs $43,000–$49,000 per year in the U.S. Meanwhile, parents often rely on screen time for engagement, despite the American Academy of Pediatrics recommending no screen time for children under 2.

ALEO is an AI-powered nanny robot designed to assist parents and caregivers with kids monitoring, engagement, and safety. Using computer vision and AI-driven motion tracking, ALEO detects movement, recognizes behavioral patterns, and provides interactive, screen-free stimulation tailored to a child's development. It plays games like Red Light, Green Light, Hide and Seek, Tag.

Alibek Bekenov, University of Maryland, Alumni

Allervision

Allervision is an AI-powered mobile app designed to make food allergy management safer and easier—especially in environments like daycare centers where the risk is high and mistakes can be life-threatening. Our technology uses computer vision and machine learning to scan ingredient labels on packaged foods and detect hidden allergens, even those not required by law to be highlighted.

The problem we’re solving is the inconsistency, confusion, and manual burden involved in identifying allergens—particularly in shared caregiving environments. Teachers and staff often rely on binders, sticky notes, or memory to manage children’s allergies, which leads to errors and anxiety.

Varsha Sankar, Carnegie Mellon University, Alumni

Basilisk Robotics

Basilisk Robotics is developing an autonomous underwater robot that inspects and cleans midsize boat hulls, targeting biofouling—a costly and environmentally damaging problem. Biofouling increases drag, fuel consumption, and emissions, and the current solution—antifouling paint—is toxic to marine ecosystems. Our AI-powered robotic solution uses machine vision and autonomous navigation to clean hulls without human divers or harmful chemicals. By enabling safer, more efficient, and environmentally sustainable maintenance, Basilisk Robotics addresses both operational inefficiencies in maritime upkeep and the urgent need to reduce the environmental footprint of marine industries.

Trisha Jha, Carnegie Mellon University, Undergraduate student
Nick Yaeger, Carnegie Mellon University, Undergraduate student
Madison Darnaby, Carnegie Mellon University, Undergraduate student

Best Possible

Medication lists obtained in the hospital are often inaccurate. Having complete and accurate medication lists, also called Best Possible Medication History (BPMH), are crucial for clinicians to assess the appropriateness of a patient’s current therapy and to direct future treatment options. However, the process of creating the BPMH is cumbersome and time consuming. Yet our overburdened clinicians and systems are not set up for success - they often have to manually hunt for information from other providers, parse through pages of electronic and sometimes even paper records, and update and correct the history. In short, “garbage in, garbage out”  and there are no real guardrails in place to prevent patient harm. This causes adverse drug events, extra hospital utilization, and prolonged length of stay, and increased mortality. With our aging population, health worker shortage, and increased medication utilization, patients and clinicians (i.e. everyone) deserve a health care infrastructure designed for best possible medication safety.

Thomas Tam, Carnegie Mellon University, Alumni
Yifeng Wang, Carnegie Mellon University, Alumni

Blaster

Our technology is a compact 3D sensor that can scan objects and produce colored point cloud data with around 0.1mm accuracy. We have also proven that it could be applied to defect detection for FDM printing. We think that our sensor could also be further developed into an inspection addon for various additive manufacturing machines to retrieve in-process inspection data.

Cadence Labs

Studies show that over 70% of dog owners misinterpret their pet’s behavioral cues, which can lead to missed early signs of distress, unaddressed anxiety, and preventable medical conditions that, if left unchecked, may escalate into emergencies. Furthermore, over 50% of dogs exhibit signs of chronic stress or anxiety, yet most of these signals go unnoticed or misunderstood. Most owners are forced to rely on guesswork—leading to miscommunication, ineffective interventions, and a diminished quality of life for both dog and owner.

Cadence Labs is solving this problem with a new class of petcare technology. We’re developing a smart collar and mobile app powered by a fusion of machine learning and signal processing. Our paralinguistic signal processing (PSP) system analyzes vocalizations, body movements, and subtle physiological signals to translate canine behavior into real-time emotional and behavioral insights.

Art Stephen, Carnegie Mellon University, Undergraduate student
Matthew Akuamoah-Boateng, Carnegie Mellon University, Undergraduate student
Emmanuel George, Carnegie Mellon University, Undergraduate student

CRS

It consists of a system to entrain cables into concrete printed filaments during printing so that it permits the printing of raised slabs or reinforced walls or vaults.

Jose Pinto Duarte, Penn State University, Graduate student
Ali Baghi, Penn State University, Graduate student

Dwaste

DWaste is an AI-powered waste management solution designed to improve recycling and waste sorting for individuals, businesses, and institutions. The app uses image recognition and gamification to help users sort and recycle waste efficiently. By scanning and categorizing waste items, users earn rewards that can be redeemed at partnered businesses. This solution addresses the growing problem of waste contamination and low recycling rates, promoting sustainable practices while offering real-time insights and data-driven solutions to businesses and institutions for better waste management strategies.

Suman Kunwar

EcoMerc

EcoMerc is developing a compact and portable spectroscopy instrument for recycling processors and material traders, coupled with machine learning, to identify high-value metals in e-waste components. Our vision is to bring efficiency and transparency to electronic waste trading, enabling faster and more accurate identification, higher profit margins, and faster throughput for material traders.

Guillermo Gutierrez, Carnegie Mellon University, Alumni
Manish Mishra, Carnegie Mellon University, Alumni

Experimental Intelligence

Biopharma organizations accumulate large volumes of scientific data from experiments, yet much of this information is effectively forgotten. Scientists lack tools to systematically revisit and extract insights from past work, making it difficult to connect historical results to current decisions. As a result, promising leads are overlooked, redundant experiments are repeated, and strategic opportunities are missed. Existing AI tools are not built to handle scientific modalities or experimental context—they lack the judgment needed to draw conclusions or propose next steps. We are building a new interface for institutional memory: a language model agent designed specifically to understand experimental data and support scientific reasoning. The goal is to turn static, fragmented data into a dynamic source of strategic guidance—helping teams make better decisions, faster, based on what they already know.

Benjamin Wu, University of Maryland,  Graduate student
Chen-Yu Chen, University of Maryland

General Astronautics

We are building the next generation of aerospace robotics: highly available, low SWaP-C (Size, Weight, Power, Cost), fully autonomous robotic agents for the repair and assembly of on-orbit assets. We currently have a desktop prototype that mirrors the legacy systems such as the CanadArm2 on the ISS, while integrating state-of-the-art reinforcement learning, computer vision, and modern optimized control techniques.

By fusing traditional control theory with advanced machine learning, we are able to adapt to unpredictable and unseen space environments. This generalization allows us to execute complex tasks such as component replacement or structural assembly with precision and reliability. Our mission is to tackle the most critical challenges in maintaining and constructing orbital infrastructure - from repairing dysfunctional or uncooperative satellites to assembling large-scale structures such as space stations and orbital refueling depots.

Our technology addresses a massive market, as both governmental defense agencies and commercial space operators invest billions in sustainable, autonomous space operations. By reducing mission risks and lowering operational costs, our system is poised to become the universal indispensable asset in the rapidly expanding space economy, ensuring continuous asset functionality and enabling the assembly of new space infrastructures necessary for long-term space exploration and colonization

Bram Schork, California Institute of Technology, Undergraduate student
Red Whittaker, Carnegie Mellon University, Faculty
Shibo Zhou, Carnegie Mellon University, Undergraduate student

GNSS Services Market Discovery

Discover how AI can assist GNSS technology to improve servicing the geographical positioning market.

Gustavo Vecino, University of Maryland, Post-doc
Miroslaw J Skibniewski, University of Maryland

GuardAIn 5G

We are creating an intelligent network of 5G-enabled AI drones and ground based cameras to detect, localize, and track active threats around university campuses and communities to empower public safety agencies and enhance safety of civilians.

Mudit Singal, University of Maryland, Staff

Kevät

"Leveraging the latest advances in AI, including generative AI, Kevat provides accurate, fast, and energy-efficient statistical modeling for companies working with geospatial data which include nature-based carbon removal companies (CDR) that need to demonstrate the efficacy and safety of their processes and accurately quantify the CO2 removed, other environmental companies, and possibly defense and intelligence companies. Due to the uncertainty inherent in natural processes, statistical modeling is a better fit for analyzing geospatial data than the current standard practice--deterministic, physical modeling or prediction-based machine learning methods. Recent developments in neural inference (ML methods essentially) for spatial statistics within academia have now made fast, reliable statistical modeling of large amounts of high-dimensional geospatial data possible. Kevät will provide these statistical modeling products to interested companies.

In particular, Kevat will provide statistical modeling enhanced by generative AI (in an energy-efficient manner) to essentially fill in the gaps between any environmental sensors (for example, the ocean subsurface) , allowing interested companies a comprehensive view of remote environments. This technology, known as neural conditional simulation (NCS), is currently being developed at Carnegie Mellon. NCS is the cornerstone on which Kevat plans to build its statistical modeling services as NCS has already gained significant attention from carbon removal projects.

Julia Walchessen, Carnegie Mellon University, Graduate student

Kromha

Kromha is building modular Farm Pods, solar-powered mobile units that bring fresh, affordable food straight into underserved areas. The problem we’re solving is simple: people either can’t access fresh food or it goes to waste before it gets to them. That applies in U.S. food deserts and across African supply chains.

Most infrastructure is too permanent or too expensive. Farm Pods are flexible, low-cost, and built for mobility. They’re designed to serve local governments, NGOs, and community operators who need fast, reliable solutions for food access without waiting on big builds or long-term funding cycles.

Kojo Dokyi, Carnegie Mellon University, Graduate student

LOGOS

The product is an AI-powered early education learning management platform.

Helene Federici, Carnegie Mellon University, Graduate student
Arup Mukherjee, Carnegie Mellon University, Alumni

Neuraville

Neuraville’s technology simplifies the creation of artificial intelligence (AI) solutions in the robotics space. Our platform provides an interactive no-code environment for the AI developer to use pre-built AI models and provides intuitive methods to help assemble the AI models in building a solution to a robotic problem such as perception, sensory fusion, navigation, or object manipulation.

The novelty of our platform stems from how we can package AI models as building blocks akin to electronic circuit components that can be used to build a circuit board with a capability. Furthermore, our core technology which is inspired by the biological brain evolutionary and development journey, is capable of enabling real-time learning and unprecedented adaptability that is quintessential for service robots – a subset of robots that work autonomously alongside humans and perform useful tasks.

Mohammad Nadji-tehrani, Pittsburgh 

Nystag

Dizziness can range from a benign condition, like vertigo, to a critical warning sign of a stroke. In emergency rooms, patients experiencing dizziness often undergo costly and time-consuming MRI scans to differentiate between these causes. This not only places a financial burden on patients and hospitals but also contributes to overcrowding and delays in emergency care.

Jaime Romero, Carnegie Mellon University, Alumni

Qiubot

Autonomous mobile platforms for material transportation for retailing/logistics

Chao Cao, Carnegie Mellon University, Graduate student

S3AM

The Smart Sustainable Shellfish Aquaculture Management framework takes the guesswork out of on-bottom oyster aquaculture by providing detailed maps of the bottom substrate quality and oyster inventory (live/dead, size, density) to improve production output and provides this information with an optimized harvest map to oyster farmers through a GPS enabled smart phone app.

Allen Pattillo, University of Maryland, Faculty

Sensify

Sensify’s high Sensor suite is a small retrofit device that is placed on existing multi-stream waste receptacles and informs end-users where each waste item should be placed based on its material and form factor. Once an item is recognized the location it should be placed is audio-visually identified to assist users in correct disposal. This information is also used to generate reports for facility managers in support of federal compliance and sustainability reporting/certifications.

Crystal Hall, Carnegie Mellon University
Rishi Basdeo, Carnegie Mellon University
Shane Deng, Northwestern University

True AI

Whilst LLMs have remarkable generative capabilities, they are prone to hallucinations — outputs that are factually incorrect or fabricated — which has limited their use in critical sectors such as law and healthcare. Current hallucination detection and mitigation services predominantly rely on retrieval-augmented generation, which are notoriously difficult to integrate and are themselves prone to hallucinations. LLM evaluations only evaluate the models tendency to hallucinate as a whole, and don’t give the user information about a particular generation's factual accuracy. Medical diagnoses, defense decisions, and insurance approvals are all areas in which we believe LLMs can make a massive difference - but not if they can’t be trusted. True AI will enable generative AI use in these fields, while promoting accuracy and openness.

Firstly, we have built a plug-and-play solution to LLM hallucinations, detecting and mitigating hallucinations in real time during generation. Our solution is low-latency, based on state-of-the-art research and is easily integratable into any foundation or domain specific LLM. We are also looking to use our hallucination mitigation product in the training stage of fine-tuning foundational models for domain specific applications, to develop safer LLMs for these critical use cases. Early testing has shown this to be promising.

Cyprien Riboud-Seydoux, Carnegie Mellon University, Undergraduate student
Guillaume Atencia, Carnegie Mellon University, Undergraduate student
Aidan Zhang, Carnegie Mellon University, Undergraduate student 

Venera AI

We build small language model for mobile application to manage patient's health, make it more accessible for free:

• Centralize users' health records from multiple sources and compile those into a comprehensive format
• - Analyze health records and provide insights/track plans for user
• - Personalize AI assistant for general medical knowledge and activities planning (exercise, book appointment)

Minh Tran, Carnegie Mellon University, Graduate Student