Blog

Inside a giant autonomous warehouse, hundreds of robots dart down aisles as they collect and distribute items to fulfill a steady stream of customer orders. In this busy environment, even small traffic jams or minor collisions can snowball into massive slowdowns.

To avoid such an avalanche of inefficiencies, researchers from MIT and the tech firm Symbotic developed a new method that automatically keeps a fleet of robots moving smoothly. Their method learns which robots should go first at each moment, based on how congestion is forming, and adapts to prioritize robots that are about to get stuck. In this way, the system can reroute robots in advance to avoid bottlenecks.

The hybrid system utilizes deep reinforcement learning, a powerful artificial intelligence method for solving complex problems, to figure out which robots should be prioritized. Then, a fast and reliable planning algorithm feeds instructions to the robots, enabling them to respond rapidly in constantly changing conditions.

In simulations inspired by actual e-commerce warehouse layouts, this new approach achieved about a 25 percent gain in throughput over other methods. Importantly, the system can quickly adapt to new environments with different quantities of robots or varied warehouse layouts.

“There are a lot of decision-making problems in manufacturing and logistics where companies rely on algorithms designed by human experts. But we have shown that, with the power of deep reinforcement learning, we can achieve super-human performance. This is a very promising approach, because in these giant warehouses even a 2 or 3 percent increase in throughput can have a huge impact,” says Han Zheng, a graduate student in the Laboratory for Information and Decision Systems (LIDS) at MIT and lead author of a paper on this new approach.

Zheng is joined on the paper by Yining Ma, a LIDS postdoc; Brandon Araki and Jingkai Chen of Symbotic; and senior author Cathy Wu, the Class of 1954 Career Development Associate Professor in Civil and Environmental Engineering (CEE) and the Institute for Data, Systems, and Society (IDSS) at MIT, and a member of LIDS. The research appears today in the Journal of Artificial Intelligence Research.

Rerouting robots

Coordinating hundreds of robots in an e-commerce warehouse simultaneously is no easy task.

The problem is especially complicated because the warehouse is a dynamic environment, and robots continually receive new tasks after reaching their goals. They need to be rapidly redirected as they leave and enter the warehouse floor.

Companies often leverage algorithms written by human experts to determine where and when robots should move to maximize the number of packages they can handle.

But if there is congestion or a collision, a firm may have no choice but to shut down the entire warehouse for hours to manually sort the problem out.

“In this setting, we don’t have an exact prediction of the future. We only know what the future might hold, in terms of the packages that come in or the distribution of future orders. The planning system needs to be adaptive to these changes as the warehouse operations go on,” Zheng says.

The MIT researchers achieved this adaptability using machine learning. They began by designing a neural network model to take observations of the warehouse environment and decide how to prioritize the robots. They train this model using deep reinforcement learning, a trial-and-error method in which the model learns to control robots in simulations that mimic actual warehouses. The model is rewarded for making decisions that increase overall throughput while avoiding conflicts.

Over time, the neural network learns to coordinate many robots efficiently.

“By interacting with simulations inspired by real warehouse layouts, our system receives feedback that we use to make its decision-making more intelligent. The trained neural network can then adapt to warehouses with different layouts,” Zheng explains.

It is designed to capture the long-term constraints and obstacles in each robot’s path, while also considering dynamic interactions between robots as they move through the warehouse.

By predicting current and future robot interactions, the model plans to avoid congestion before it happens.

After the neural network decides which robots should receive priority, the system employs a tried-and-true planning algorithm to tell each robot how to move from one point to another. This efficient algorithm helps the robots react quickly in the changing warehouse environment.

This combination of methods is key.

“This hybrid approach builds on my group’s work on how to achieve the best of both worlds between machine learning and classical optimization methods. Pure machine-learning methods still struggle to solve complex optimization problems, and yet it is extremely time- and labor-intensive for human experts to design effective methods. But together, using expert-designed methods the right way can tremendously simplify the machine learning task,” says Wu.

Overcoming complexity

Once the researchers trained the neural network, they tested the system in simulated warehouses that were different than those it had seen during training. Since industrial simulations were too inefficient for this complex problem, the researchers designed their own environments to mimic what happens in actual warehouses.

On average, their hybrid learning-based approach achieved 25 percent greater throughput than traditional algorithms as well as a random search method, in terms of number of packages delivered per robot. Their approach could also generate feasible robot path plans that overcame congestion caused by traditional methods.

“Especially when the density of robots in the warehouse goes up, the complexity scales exponentially, and these traditional methods quickly start to break down. In these environments, our method is much more efficient,” Zheng says.

While their system is still far away from real-world deployment, these demonstrations highlight the feasibility and benefits of using a machine learning-guided approach in warehouse automation.

In the future, the researchers want to include task assignments in the problem formulation, since determining which robot will complete each task impacts congestion. They also plan to scale up their system to larger warehouses with thousands of robots.

This research was funded by Symbotic.

Each spring, river herring populations migrate from Massachusetts coastal waters to begin their annual journey up rivers and streams to freshwater spawning habitat. River herring have faced severe population declines over the past several decades, and their migration is extensively monitored across the region, primarily through traditional visual counting and volunteer-based programs. 

Monitoring fish movement and understanding population dynamics are essential for informing conservation efforts and supporting fisheries management. With the annual herring run getting underway this month, researchers and resource managers once again take on the challenge of counting and estimating the migrating fish population as accurately as possible. 

A team of researchers from the Woodwell Climate Research Center, MIT Sea Grant, the MIT Computer Science and Artificial Intelligence Lab (CSAIL), MIT Lincoln Laboratory, and Intuit explored a new monitoring method using underwater video and computer vision to supplement citizen science efforts. The researchers — Zhongqi Chen and Linda Deegan from the Woodwell Climate Research Center, Robert Vincent and Kevin Bennett from MIT Sea Grant, Sara Beery and Timm Haucke from MIT CSAIL, Austin Powell from Intuit, and Lydia Zuehsow from MIT Lincoln Laboratory — published a paper describing this work in the journal Remote Sensing in Ecology and Conservation this February. 

The open-access paper, “From snapshots to continuous estimates: Augmenting citizen science with computer vision for fish monitoring,” outlines how recent advancements in computer vision and deep learning, from object detection and tracking to species classification, offer promising real-world solutions for automating fish counting with improved efficiency and data quality. 

Traditional monitoring methods are constrained by time, environmental conditions, and labor intensity. Volunteer visual counts are limited to brief daytime sampling windows, missing nighttime movement and short migration pulses, when hundreds of fish pass by within the span of a few minutes. While technologies like passive acoustic monitoring and imaging sonar have advanced continuous fish monitoring under certain conditions, the most promising and low-cost option — manual review of underwater video — is still labor-intensive and time-consuming. With the growing demand for automated video processing solutions, this study presents a scalable, cost-effective, and efficient deep learning-based system for reliable automated fish monitoring. 

The team built an end-to-end pipeline — from in-field underwater cameras to video labeling and model training — to achieve automated, computer vision-powered fish counting. Videos were collected from three rivers in Massachusetts: the Coonamessett River in Falmouth, the Ipswich River (Ipswich), and the Santuit River in Mashpee. 

To prepare the training dataset, the team selected video clips with variations in lighting, water clarity, fish species and density, time of day, and season to ensure that the computer vision model would work reliably across diverse real-world scenarios. They used an open-source web platform to manually label the videos frame-by-frame with bounding boxes to track fish movement. In total, they labeled 1,435 video clips and annotated 59,850 frames. 

The researchers compared and validated the computer vision counts with human video reviews, stream-side visual counts, and data from passive integrated transponder (PIT) tagging. They concluded that models trained on diverse multi-site and multi-year data performed best and produced season-long, high-resolution counts consistent with traditionally established estimates. Going one step further, the system provided insights into migration behavior, timing, and movement patterns linked to environmental factors. Using video from the 2024 Coonamesset River migration, the system counted 42,510 river herring and revealed that upstream migration peaked at dawn, while downstream migration was largely nocturnal, with fish utilizing darker, quieter periods to avoid predators.

With this real-world application, the researchers aim to advance computer vision in fisheries management and provide a framework and best practices for integrating the technology into conservation efforts for a wide range of aquatic species. “MIT Sea Grant has been funding work on this topic for some time now, and this excellent work by Zhongqi Chen and colleagues will advance fisheries monitoring capabilities and improve fish population assessments for fisheries managers and conservation groups,” Vincent says. “It will also provide education and training for students, the public, and citizen science groups in support of the ecologically and culturally important river herring populations along our coasts.”

Still, continued traditional monitoring is essential for maintaining consistency in long-term datasets until fisheries management agencies fully implement automated counting systems. Even then, computer vision and citizen science should be seen as complementary. Volunteers will be necessary for camera maintenance and for contributing directly to the computer vision workflow, from video annotation to model verification. The researchers envision that integrating citizen observations and computer vision-generated data will help create a more comprehensive and holistic approach to environmental monitoring.

This work was funded by MIT Sea Grant, with additional support provided by the Northeast Climate Adaptation Science Center, an MIT Abdul Latif Jameel Water and Food Systems seed grant, the AI and Biodiversity Change Global Center (supported by the National Science Foundation and the Natural Sciences and Engineering Research Council of Canada), and the MIT Undergraduate Research Opportunities Program.

The next time you’re scrolling your phone, take a moment to appreciate the feat: The seemingly mundane act is possible thanks to the coordination of 34 muscles, 27 joints, and over 100 tendons and ligaments in your hand. Indeed, our hands are the most nimble parts of our bodies. Mimicking their many nuanced gestures has been a longstanding challenge in robotics and virtual reality.

Now, MIT engineers have designed an ultrasound wristband that precisely tracks a wearer’s hand movements in real-time. The wristband produces ultrasound images of the wrist’s muscles, tendons, and ligaments as the hand moves, and is paired with an artificial intelligence algorithm that continuously translates the images into the corresponding positions of the five fingers and palm.

The researchers can train the wristband to learn a wearer’s hand motions, which the device can communicate in real-time to a robot or a virtual environment.

In demonstrations, the team has shown that a person wearing the wristband can wirelessly control a robotic hand. As the person gestures or points, the robot does the same. In a sort of wireless marionette interaction, the wearer can manipulate the robot to play a simple tune on the piano and shoot a small basketball into a desktop hoop. With the same wristband, a wearer can also manipulate objects on a computer screen, for instance pinching their fingers together to enlarge and minimize a virtual object.

The team is using the wristband to gather hand motion data from many more users with different hand sizes, finger shapes, and gestures. They envision building a large dataset of hand motions that can be plumbed, for instance, to train humanoid robots in dexterity tasks, such as performing certain surgical procedures. The ultrasound band could also be used to grasp, manipulate, and interact with objects in video games, design applications, or other virtual settings.

“We think this work has immediate impact in potentially replacing hand tracking techniques with wearable ultrasound bands in virtual and augmented reality,” says Xuanhe Zhao, the Uncas and Helen Whitaker Professor of Mechanical Engineering at MIT. “It could also provide huge amounts of training data for dexterous humanoid robots.”

Zhao, Gengxi Lu, and their colleagues present the wristband’s new design in a paper appearing today in Nature Electronics. Their MIT co-authors are former postdocs Xiaoyu Chen, Shucong Li, and Bolei Deng; graduate students SeongHyeon Kim and Dian Li; postdocs Shu Wang and Runze Li; and Anantha Chandrakasan, MIT provost and the Vannevar Bush Professor of Electrical Engineering and Computer Science. Other co-authors are graduate students Yushun Zheng and Junhang Zhang, Baoqiang Liu, Chen Gong, and Professor Qifa Zhou from the University of Southern California.

Seeing strings

There are currently a number of approaches to capturing and mimicking human hand dexterity in robots. Some approaches use cameras to record a person’s hand movements as they manipulate objects or perform tasks. Others involve having a person wear a glove with sensors, which records the person’s hand movements and transmits the data to a receiving robot. But erecting a complex camera system for different applications is impractical and prone to visual obstacles. And sensor-laden gloves could limit a person’s natural hand motions and sensations.

A third approach uses the electrical signals from muscles in the wrist or forearm that scientists then correlate with specific hand movements. Researchers have made significant advances in this approach, however these signals are easily affected by noise in the environment. They are also not sensitive enough to distinguish subtle changes in movements. For instance, they may discern whether a thumb and index finger are pinched together or pulled apart, but not much of the in-between path.

Zhao’s team wondered whether ultrasound imaging might capture more dexterous and continuous hand movements. His group has been developing various forms of ultrasound stickers — miniaturized versions of the transducers used in doctor’s offices that are paired with hydrogel material that can safely stick to skin.

In their new study, the team incorporated the ultrasound sticker design into a wearable wristband to continuously image the muscles and tendons in the wrist.

“The tendons and muscles in your wrist are like strings pulling on puppets, which are your fingers,” Lu says. “So the idea is: Each time you take a picture of the state of the strings, you’ll know the state of the hand.”

Mapping manipulation

The team designed a wristband with an ultrasound sticker that is the size of a smartwatch, and added onboard electronics that are about as small as a cellphone. They attached the wristband to a volunteer’s wrist and confirmed that the device produced clear and continuous images of the wrist as the volunteer moved their fingers in various gestures.

The challenge then was to relate the black and white ultrasound images of the wrist to specific positions of the hand. As it turns out, the fingers and thumb are capable of 22 degrees of freedom, or different ways of extending or angling. The researchers found that they could identify specific regions in their ultrasound images of the wrist that correlate to each of these 22 degrees of freedom. For instance, changes in one region relate to thumb extension, while changes in another region correlate with movements of the index finger.

To establish these connections, a volunteer wearing the wristband would move their hand in various positions while the researchers recorded the gestures with multiple cameras surrounding the volunteer. By matching changes in certain regions of the ultrasound images with hand positions recorded by the cameras, the team could label wrist image regions with the corresponding degree of freedom in the hand. But to do this translation continuously, and in real-time, would be an impossible task for humans.

So, the team turned to artificial intelligence. They used an AI algorithm that can be trained to recognize image patterns and correlate them with specific labels and, in this case, the hand’s various degrees of freedom. The researchers trained the algorithm with ultrasound images that they meticulously labeled, annotating the image regions associated with a specific degree of freedom. They tested the algorithm on a new set of ultrasound images and found it correctly predicted the corresponding hand gestures.

Once the researchers successfully paired the AI algorithm with the wristband, they tested the device on more volunteers. For the new study, eight volunteers with different hand and wrist sizes wore the wristband while they formed various hand gestures and grasps, including making the signs for all 26 letters in American Sign Language. They also held objects such as a tennis ball, a plastic bottle, a pair of scissors, and a pencil. In each case, the wristband precisely tracked and predicted the position of the hand.

To demonstrate potential applications, the team developed a simple computer program that they wirelessly paired with the wristband. As a wearer went through the motions of pinching and grasping, the gestures corresponded to zooming in and out on an object on the computer screen, and virtually moving and manipulating it in a smooth and continuous fashion.

The researchers also tested the wristband as a wireless controller of a simple commercial robotic hand. While wearing the wristband, a volunteer went through the motions of playing a keyboard. The robot in turn mimicked the motions in real-time to play a simple tune on a piano. The same robot was also able to mimic a person’s finger taps to play a desktop basketball game.

Zhao is planning to further miniaturize the wristband’s hardware, as well as train the AI software on many more gestures and movements from volunteers with wider ranging hand sizes and shapes. Ultimately, the team is building toward a wearable hand tracker that can be worn by anyone, to wirelessly manipulate humanoid robots or virtual objects with high dexterity.

“We believe this is the most advanced way to track dexterous hand motion, through wearable imaging of the wrist,” Zhao says. “We think these wearable ultrasound bands can provide intuitive and versatile controls for virtual reality and robotic hands.”

This research was supported, in part, by MIT, the U.S. National Institutes of Health, the U.S. National Science Foundation, the U.S. Department of Defense, and Singapore National Research Foundation through the Singapore-MIT Alliance for Research and Technology.

Artificial intelligence holds promise for helping doctors diagnose patients and personalize treatment options. However, an international group of scientists led by MIT cautions that AI systems, as currently designed, carry the risk of steering doctors in the wrong direction because they may overconfidently make incorrect decisions.

One way to prevent these mistakes is to program AI systems to be more “humble,” according to the researchers. Such systems would reveal when they are not confident in their diagnoses or recommendations and would encourage users to gather additional information when the diagnosis is uncertain.

“We’re now using AI as an oracle, but we can use AI as a coach. We could use AI as a true co-pilot. That would not only increase our ability to retrieve information but increase our agency to be able to connect the dots,” says Leo Anthony Celi, a senior research scientist at MIT’s Institute for Medical Engineering and Science, a physician at Beth Israel Deaconess Medical Center, and an associate professor at Harvard Medical School.

Celi and his colleagues have created a framework that they say can guide AI developers in designing systems that display curiosity and humility. This new approach could allow doctors and AI systems to work as partners, the researchers say, and help prevent AI from exerting too much influence over doctors’ decisions.

Celi is the senior author of the study, which appears today in BMJ Health and Care Informatics. The paper’s lead author is Sebastián Andrés Cajas Ordoñez, a researcher at MIT Critical Data, a global consortium led by the Laboratory for Computational Physiology within the MIT Institute for Medical Engineering and Science.

Instilling human values

Overconfident AI systems can lead to errors in medical settings, according to the MIT team. Previous studies have found that ICU physicians defer to AI systems that they perceive as reliable even when their own intuition goes against the AI suggestion. Physicians and patients alike are more likely to accept incorrect AI recommendations when they are perceived as authoritative.

In place of systems that offer overconfident but potentially incorrect advice, health care facilities should have access to AI systems that work more collaboratively with clinicians, the researchers say.

“We are trying to include humans in these human-AI systems, so that we are facilitating humans to collectively reflect and reimagine, instead of having isolated AI agents that do everything. We want humans to become more creative through the usage of AI,” Cajas Ordoñez says.

To create such a system, the consortium designed a framework that includes several computational modules that can be incorporated into existing AI systems. The first of these modules requires an AI model to evaluate its own certainty when making diagnostic predictions. Developed by consortium members Janan Arslan and Kurt Benke of the University of Melbourne, the Epistemic Virtue Score acts as a self-awareness check, ensuring the system’s confidence is appropriately tempered by the inherent uncertainty and complexity of each clinical scenario.

With that self-awareness in place, the model can tailor its response to the situation. If the system detects that its confidence exceeds what the available evidence supports, it can pause and flag the mismatch, requesting specific tests or history that would resolve the uncertainty, or recommending specialist consultation. The goal is an AI that not only provides answers but also signals when those answers should be treated with caution.

“It’s like having a co-pilot that would tell you that you need to seek a fresh pair of eyes to be able to understand this complex patient better,” Celi says.

Celi and his colleagues have previously developed large-scale databases that can be used to train AI systems, including the Medical Information Mart for Intensive Care (MIMIC) database from Beth Israel Deaconess Medical Center. His team is now working on implementing the new framework into AI systems based on MIMIC and introducing it to clinicians in the Beth Israel Lahey Health system.

This approach could also be implemented in AI systems that are used to analyze X-ray images or to determine the best treatment options for patients in the emergency room, among others, the researchers say.

Toward more inclusive AI

This study is part of a larger effort by Celi and his colleagues to create AI systems that are designed by and for the people who are ultimately going to be most impacted by these tools. Many AI models, such as MIMIC, are trained on publicly available data from the United States, which can lead to the introduction of biases toward a certain way of thinking about medical issues, and exclusion of others.

Bringing in more viewpoints is critical to overcoming these potential biases, says Celi, emphasizing that each member of the global consortium brings a distinct perspective to a broader, collective understanding.

Another problem with existing AI systems used for diagnostics is that they are usually trained on electronic health records, which weren’t originally intended for that purpose. This means that the data lack much of the context that would be useful in making diagnoses and treatment recommendations. Additionally, many patients never get included in those datasets because of lack of access, such as people who live in rural areas.

At data workshops hosted by MIT Critical Data, groups of data scientists, health care professionals, social scientists, patients, and others work together on designing new AI systems. Before beginning, everyone is prompted to think about whether the data they’re using captures all the drivers of whatever they aim to predict, ensuring they don’t inadvertently encode existing structural inequities into their models.

“We make them question the dataset. Are they confident about their training data and validation data? Do they think that there are patients that were excluded, unintentionally or intentionally, and how will that affect the model itself?” he says. “Of course, we cannot stop or even delay the development of AI, not just in health care, but in every sector. But, we must be more deliberate and thoughtful in how we do this.”

The research was funded by the Boston-Korea Innovative Research Project through the Korea Health Industry Development Institute.

The sense of support and community was palpable when Sojun Park, a postdoc at the MIT Center for International Studies (CIS), delivered a recent presentation on The Global Diffusion of AI Technologies and Its Political Drivers. The event, part of the CIS Global Research and Policy Seminar, filled the venue with audience members from across MIT. 

“My work is directly connected to what CIS faculty have previously done on international trade and security,” Park said afterwards. “If I hadn’t received a postdoctoral fellowship and come to MIT, I wouldn’t have been able to think through the security implications of my intellectual property research. I’ve been tremendously motivated by these scholars.”

Park’s time at CIS has been both grounding and transformative, offering him a scholarly home that has shaped his research and helped broaden his intellectual horizons.

Pursuing interdisciplinary research and connections 

Before pursuing a tenure-track position, Park set his sights on conducting research at MIT. When he came across a public posting about the CIS Postdoctoral Associate Program, he took a chance and applied.

“My own research is interdisciplinary, and I knew that I could really benefit from the interdisciplinary environment at MIT, and specifically at CIS, where faculty are coming not only from political science, but also affiliated with the Department of Economics and MIT Sloan [School of Management],” he says.

Park was thrilled to receive the paid fellowship, which offers an academic year at MIT and dedicated office space at CIS. At MIT, he is free to use his time toward his own research, and has found value in pursuing topics that are of interest to the CIS community — whether it’s AI or global governance. He’s published prolifically along the way, including two articles in the Review of International Organizations and the Review of International Political Economy.

He’s also continued to work on his forthcoming book, “From Privilege to Prosperity: Knowledge Diffusion and the Global Governance of Intellectual Property,” which examines how technologies can be transferred legitimately across borders. “By ‘legitimately,’ I am asking under what circumstances would firms volunteer to share their technologies? I’m interested in institutions and institutional environments that allow large businesses to share their technologies with smaller businesses based in the development world that may not possess the ability to come up with their own technologies,” he explains.

During the spring 2026 semester, he is collaborating with the center’s Undergraduate Fellows Program. This program enables postdocs to work on their research projects with MIT undergraduates. Park is working with two CIS undergraduate fellows to develop a new dataset examining international trade in green technologies. This opportunity reconnects Park to his early academic experiences in South Korea that set him on the path to MIT.

Path to MIT

“Students in South Korea are trained to be problem-solvers,” explains Park, who was born and raised in Seoul. The country’s rigorous college entrance exams reward those who can answer the most questions quickly and accurately in a limited amount of time.

While taking a test in high school, Park stumbled over a question that he couldn’t answer, regardless of how much time he spent concentrating on it. He handed in the exam, but took the problem home and spent hours puzzling over it — he just couldn’t let it go. “In hindsight, I see this as the moment I decided that I wanted to become a scholar,” Park says.

While majoring in international studies and economics (statistics) at Korea University, he had the opportunity to participate in a semester-long exchange program at the University of Texas at Austin. There, Park enrolled in a political science course on game theory that explored how individual state actors’ decisions influenced one another’s choices and outcomes in trade, conflict, and diplomacy. The instructor used the ongoing war between North and South Korea as a case study, demonstrating the unique circumstances for escalation or de-escalation depending upon how the key actors made choices along the way.

“I saw for the first time how quantitative methods could be applied to international relations and political economy,” Park says — and he knew that his next step was going to be graduate work in the United States. He began a joint MA and PhD program in political science at Princeton University the following year, supported by a Fulbright Fellowship.

Park’s 2025 dissertation examined the global governance of intellectual property rights — and it was timely. He began his PhD program in 2018, “the point at which the U.S. and China trade war had just begun.” During the pandemic, he was moved by the ongoing debates regarding vaccine inequality. “I realized then that intellectual property was at the center of these global economic challenges.” With little political science research on the topic, he “set out to create a systemic framework” to study it.

Simultaneously, he served as a teaching assistant in undergraduate courses in statistical analysis and realized that he deeply enjoyed the experience of teaching and interacting with students. It was a very different experience from his own college years. 

“In South Korea, it’s common for the learning environment to be one in which the professor just delivers lectures, but I found that in the United States’ higher education system, the classroom is truly interactive. I learned something from each of my students.” Soon, Park was certain that he not only wanted to build a career in academic research, but also a future that heavily incorporated teaching and mentoring students.

Before graduating, he spent a year at Georgetown University as a predoctoral fellow affiliated with the Mortara Center for International Studies. This experience enabled him to explore the policy implications of his research and engage with policymakers in Washington — skills he will draw on in his new position.

Lasting lessons from CIS

Park recently accepted a position as assistant professor at the National University of Singapore. Beginning fall 2026, he will be teaching graduate students affiliated with the school of public policy — most of whom will have career experience as practitioners in the public or private sectors. 

He’ll take many lessons from MIT to his new academic home, he says. “Based on what I learned in the United States, I’ll make the learning environment in the graduate courses I teach much more interactive and collaborative.”

At CIS, Mihaela Papa, director of research and principal research scientist, and Evan Lieberman, the center’s director and professor of political science, connected Park to associated faculty whose research interests were related with his own. “Meeting with all of these scholars whose research relates in some way to intellectual property rights made me think about how my own interests can expand to other topics,” Park explains. 

But the biggest takeaway of all is that he learned how to share his own research with scholars who study unfamiliar topics, to exchange ideas and discover commonality. “I’ll never stop using the communication skills that I got here at MIT,” Park says.

From enhancing international business logistics to freeing up more hospital beds to helping farmers, MIT Professor Dimitris Bertsimas SM ’87, PhD ’88 summarized how his work in operations research has helped drive real-world improvements, while delivering the 54th annual James R. Killian Faculty Achievement Award Lecture at MIT on Thursday, March 19.

Bertsimas also described how artificial intelligence is now being used in some of his scholarly projects and as a tool in MIT Open Learning efforts, which he currently directs — another facet of a highly productive and lauded career over four decades at the Institute. The Killian Award is the highest prize MIT gives its faculty.

“I have tried to improve the human condition,” Bertsimas said, summarizing the breadth of his work and the many applications to everyday living that he has found for it.

At MIT, Bertsimas is the vice provost for open learning, associate dean for online education and artificial intelligence, Boeing Leaders for Global Operations Professor of Management, and professor of operations research in the MIT Sloan School of Management. He also served as the inaugural faculty director of the master of business analytics program at MIT Sloan, and has held the position of associate dean of business analytics.

Bertsimas’ remarks encompassed both his past insights and his ongoing studies, as well as his current efforts to add AI to his research. Describing the concept of “robust optimization,” a highly influential approach that Bertsimas helped develop in the early 2000s, he explained how it has enabled, for instance, more reliable shipping through the Panama Canal. Other approaches to optimization aimed at getting more vessels through the canal every day — up to 48 — but would encounter significant problems at times. Bertsimas’ approach identified that 45 vessels a day was better — a slightly lower number, but one that “was always feasible,” he noted.

Over time, Bertsimas’ work has helped structure all kinds of solutions in business logistics; it has even been used for the allocation of school buses in Boston.

More recently, as Bertsimas explained in the lecture, he and his collaborators have been working with Hartford HealthCare in Connecticut on a wide range of issues, and are increasingly incorporating AI into the development of tools for diagnostics, among other things. On the optimization front, their research has suggested ways to reduce the average stay of a hospital patient, from 5.38 days to 4.93 days. In the main Hartford hospital they have studied, given the number of existing beds, that reduction has enabled more than 5,000 additional patient stays per year.

“It’s a very different ballgame,” Bertsimas said.

Bertsimas delivered his lecture, titled “Algorithms for Life: AI and Operations Research Transforming Healthcare, Education, and Agriculture,” to an audience of over 300 MIT community members in Huntington Hall (Room 10-250) on campus.

The award was established in 1971 to honor James Killian, whose distinguished career included serving as MIT’s 10th president, from 1948 to 1959, and subsequently as chair of the MIT Corporation, from 1959 to 1971.

“Professor Bertsimas’ scholarly contributions are both extensive and groundbreaking,” said Roger Levy, chair of the MIT faculty and a professor in the Department of Brain and Cognitive Sciences, while making introductory remarks. “He’s one of the rare individuals who has made significant contributions to both intellectual threads in the field of operations research: one, optimization — combinatorial, linear, and nonlinear — and number two, stochastic processes.”

Indeed, Bertsimas’ work has helped develop both better tools for studying and conducting operations, while also having a wide range of applications. As Bertsimas noted in his lecture, the deaths of both of his parents in 2009 helped propel him to start looking at extensively at ways operations research could help health care.

Bertsimas received his BS in electrical engineering and computer science from the National Technical University of Athens in Greece. Moving to MIT for his graduate work, he then earned his MS in operations research and his PhD in applied mathematics and operations research. Bertsimas joined the MIT faculty after receiving his doctorate, and has remained at the Institute ever since.

Bertsimas is also known as an energetic teacher who has been the principal advisor to a remarkable number of PhD students — 106 and counting, at this point.

“It is far and away my favorite activity, to supervise my doctoral students,” Bertsimas said. “It is a privilege, in my opinion, to work with exceptional young people like the ones we have at MIT, in ability and character and aspiration. They actually make me a better scientist, and a better person.”

“MIT is part of my identity,” Bertsimas quipped while noting that he is the only faculty member on campus who has those three letters, in order, in his first name.

In the latter part of the lecture, Bertsimas highlighted work he has been doing as vice provost of open learning at MIT. He has personally developed an large online course based on his own material, “The Analytics Edge.” In his current role, Bertsimas said, he now aspires for MIT to reach a billion learners with online courses, part of his effort to “democratize access to education.”

Bertsimas also demonstrated for the audience some AI tools he and his colleagues are working to bring to online education, including ways of condensing material, and the translation of online material into other languages.

It is just one more chapter in a long and broad-ranging career dedicated to grasping phenomena and developing tools to help us navigate it.

Or as Berstimas noted while summarizing his scholarship at one point in the lecture, “I try to increase the human understanding of how the world works.” 

Who benefits from artificial intelligence? This basic question, which has been especially salient during the AI surge of the last few years, was front and center at a conference at MIT on Wednesday, as speakers and audience members grappled with the many dimensions of AI’s impact.

In one of the conferences’s keynote talks, journalist Karen Hao ’15 called for an altered trajectory of AI development, including a move away from the massive scale-up of data use, data centers, and models being used to develop tools under the rubric of “artificial general intelligence.”

“This scale is unnecessary,” said Hao, who has become a prominent voice in AI discussions. “You do not need this scale of AI and compute to realize the benefits.” Indeed, she added, “If we really want AI to be broadly beneficial, we urgently need to shift away from this approach.”

Hao is a former staff member at The Wall Street Journal and MIT Technology Review, and author of the 2025 book, “Empire of AI.” She has reported extensively on the growth of the AI industry.

In her remarks, Hao outlined the astonishing size of datasets now being used by the biggest AI firms to develop large language models. She also emphasized some of the tradeoffs in this scale-up, such as the massive energy consumption and emissions of hyper-scale data centers, which also consume large amounts of water. Drawing on her own reporting, Hao also noted the human toll from the input work that global gig-economy employees do, inputting data manually for the hyper-scale models.

By contrast, Hao offered, an alternate path for AI might exist in the example of AlphaFold, the Nobel Prize-winning tool used to identify protein structures. This represents the concept of the “small, task-specific AI model tackling a well-scoped problem that lends itself to the computational strengths of AI,” Hao said.

She added: “It’s trained on highly curated data sets that only have to do with the problem at hand: protein folding and amino acid sequences. … There’s no need for fast supercomputing because the datasets are small, the model is small, and it’s still unlocking enormous benefit.”

In a second keynote address, scholar Paola Ricaurte underscored the desirability of purpose-driven AI approaches, outlining a number of conceptual keys to evaluating the usefulness of AI.

“There is no sense in having technologies that are not going to respond to the communities that are going to use them,” said Ricaurte.

She is a professor at Tecnologico de Monterrey in Mexico and a faculty associate at Harvard University’s Berkman Klein Center for Internet and Society. Ricaurte has also served on expert committees such as the Global Partnership for AI, UNESCO’s AI Ethics Experts Without Borders, and the Women for Ethical AI project.

The event was hosted by the MIT Program in Women’s and Gender Studies. Manduhai Buyandelger, the program’s director and a professor of anthropology, provided introductory remarks.

Titled “Gender, Empire, and AI: Symposium and Design Workshop,” the event was held in the conference space at the MIT Schwarzman College of Computing, with over 300 people in attendance for the keynote talks. There was also a segment of the event devoted to discussion groups, and an afternoon session on design, in a half-dozen different subject areas.

In her talk, Hao decried the often-vague nature of AI discourse, suggesting it impedes a more thoughtful discussion about the industry’s direction.

“Part of the challenge in talking about AI is the complete lack of specificity in the term ‘artificial intelligence,’” Hao said. “It’s like the word ‘transportation.’ You could be referring to anything from a bicycle to a rocket.” As a result, she said, “when we talk about accessing its benefits, we actually have to be very specific. Which AI technologies are we talking about, and which ones do we want more of?”

In her view, the smaller-sized tools — more akin to the bicycle, by analogy — are more useful on an everyday basis. As another example, Hao mentioned the project Climate Change AI, focused on tools that can help improve the energy efficiency of buildings, track emissions, optimize supply chains, forecast extreme weather, and more.

“This is the vision of AI that we should be building towards,” Hao said.

In conclusion, Hao encouraged audience members to be active participants in AI-related discourse and projects, saying the trajectory of the technology was not yet fixed, and that public interventions matter.

Citing the writer Rebecca Solnit, Hao suggested to the audience that “Hope locates itself in the premise that we don’t know what will happen, and that in the spaciousness of uncertainty is room to act.” She also noted, “Each and every one of you has an active role to play in shaping technology development.”

Ricaurte, similarly, encouraged attendees to be proactive participants in AI matters, noting that technologies will work best when the pressing everyday needs of all citizens are addressed.

“We have the responsibility to make hope possible,” Ricaurte said.

The following is a joint announcement from the MIT School of Architecture and Planning, MIT Schwarzman College of Computing, Hasso Plattner Institute, and Hasso Plattner Foundation.

The MIT Morningside Academy for Design (MAD), MIT Schwarzman College of Computing, Hasso Plattner Institute (HPI), and Hasso Plattner Foundation celebrated the launch of the MIT and HPI AI and Creativity Hub (MHACH) at a signing ceremony this week. This 10-year initiative aims to deepen ties between computing and design as advances in artificial intelligence are reshaping how ideas are conceived and shared.

Funded by the Hasso Plattner Foundation, MIT and HPI will work together to foster collaborative interdisciplinary research and support a portfolio of educational programs, fellowships, and faculty engagement focused on AI and creativity, expanding scholarly inquiry into AI applications across disciplines, industries, and societal challenges. The collaboration begins with an inaugural two-day workshop March 19-20 at MIT, bringing together faculty, students, and researchers to set early priorities.

“As we hear from our faculty, as the Information Age gives way to an era of imagination, we expect a new emphasis on human creativity,” reflects MIT President Sally Kornbluth. “Through this collaboration, MIT and HPI are creating a shared space where students and faculty will come together across disciplines to explore new ideas, experiment with emerging tools, and invent new frontiers at the intersection of human creativity and AI.”

“The best minds need the right environment to do their most creative work,” says Rouven Westphal, from the Hasso Plattner Foundation. “When HPI and MIT come together across disciplines and borders, they create exactly that. The Hasso Plattner Foundation is committed to supporting this collaboration for the long term, building on Hasso Plattner’s vision of uniting technological excellence with human-centered design and creativity.”

Deepening collaboration at the intersection of technology, creativity, and societal impact

Building on the success of the Hasso Plattner Institute-MIT Research Program on Designing for Sustainability, established in 2022 between MIT MAD and HPI, the new MHACH hub represents a commitment to deepen collaboration at the intersection of technology, creativity, and societal impact.

“MIT and HPI share a common commitment to turning scientific excellence into real-world impact. Through this collaboration, we will create an environment where students and researchers from both sides of the Atlantic can work together, experiment across disciplines, and learn from one another — at a time when artificial intelligence is set to profoundly shape our lives. We are convinced that this collaboration will generate ideas with impact far beyond both institutions and inspire international cooperation and innovation,” says Professor Tobias Friedrich, dean and managing director of the Hasso Plattner Institute.

“HPI and MIT exist at the nexus of technology and creativity. Expanding this dynamic relationship will generate new paths for the infusion of AI, design, and creativity, enabling students, faculty, and researchers to dream and discover novel solutions, moving more quickly than ever from idea to implementation. MAD was established to connect thinkers across and beyond the Institute, and this new era of collaboration with HPI advances that mission on a global scale,” comments Hashim Sarkis, dean of the MIT School of Architecture and Planning and the Elizabeth and James Killian (1926) Professor.

Academic leadership from MIT and HPI will jointly shape the hub’s research and teaching agenda. Based in Potsdam, Germany, HPI is a center of excellence for digital engineering advancing research, education, and societal transfer in IT systems engineering, data engineering, cybersecurity, entrepreneurship, and digital health. Through its globally recognized HPI d-school and pioneering work in design thinking methodology, HPI brings a distinctive perspective on human-centered innovation to the collaboration, alongside a strong record in AI and data science research and technology transfer.

Expanding research and education on AI and creativity

The efforts of this multifaceted initiative are intended to foster a dynamic academic community spanning MIT and HPI, anchored by Hasso Plattner–named professorships and graduate fellowships whose recipients will be actively engaged in the hub. The long-term framework is designed to provide continuity for faculty appointments, doctoral training, and cross-campus research.

The agreement also includes the development of classes and educational programs in areas of shared AI focus, along with expanded experiential opportunities through AI-focused workshops, hackathons, and summer exchanges. A steering committee composed of representatives from the MIT School of Architecture and Planning, MIT Schwarzman College of Computing, and Hasso Plattner Institute will facilitate the shared governance of MHACH.

“Creativity has always been about extending human capability. At its core, this collaboration asks what it truly means to create something new. The question isn’t whether AI diminishes creativity, but how new forms of intelligence can deepen and enrich that process. Our goal is to explore that intersection with rigor and build a cross-disciplinary scholarly and research community that shapes how AI supports the creation of new ideas and knowledge,” says Dan Huttenlocher, dean of the MIT Schwarzman College of Computing and the Henry Ellis Warren (1894) Professor of Electrical Engineering and Computer Science.

This collaboration is made possible by the Hasso Plattner Foundation’s long-term philanthropic commitment to institutions that connect technological innovation with design thinking and education. The Hasso Plattner Foundation has played a central role in establishing and supporting institutions such as the Hasso Plattner Institute and international design thinking programs that bridge disciplines and geographies.

MIT researchers have spent more than a decade studying techniques that enable robots to find and manipulate hidden objects by “seeing” through obstacles. Their methods utilize surface-penetrating wireless signals that reflect off concealed items.

Now, the researchers are leveraging generative artificial intelligence models to overcome a longstanding bottleneck that limited the precision of prior approaches. The result is a new method that produces more accurate shape reconstructions, which could improve a robot’s ability to reliably grasp and manipulate objects that are blocked from view.

This new technique builds a partial reconstruction of a hidden object from reflected wireless signals and fills in the missing parts of its shape using a specially trained generative AI model.

The researchers also introduced an expanded system that uses generative AI to accurately reconstruct an entire room, including all the furniture. The system utilizes wireless signals sent from one stationary radar, which reflect off humans moving in the space.  

This overcomes one key challenge of many existing methods, which require a wireless sensor to be mounted on a mobile robot to scan the environment. And unlike some popular camera-based techniques, their method preserves the privacy of people in the environment.

These innovations could enable warehouse robots to verify packed items before shipping, eliminating waste from product returns. They could also allow smart home robots to understand someone’s location in a room, improving the safety and efficiency of human-robot interaction.

“What we’ve done now is develop generative AI models that help us understand wireless reflections. This opens up a lot of interesting new applications, but technically it is also a qualitative leap in capabilities, from being able to fill in gaps we were not able to see before to being able to interpret reflections and reconstruct entire scenes,” says Fadel Adib, associate professor in the Department of Electrical Engineering and Computer Science, director of the Signal Kinetics group in the MIT Media Lab, and senior author of two papers on these techniques. “We are using AI to finally unlock wireless vision.”

Adib is joined on the first paper by lead author and research assistant Laura Dodds; as well as research assistants Maisy Lam, Waleed Akbar, and Yibo Cheng; and on the second paper by lead author and former postdoc Kaichen Zhou; Dodds; and research assistant Sayed Saad Afzal. Both papers will be presented at the IEEE Conference on Computer Vision and Pattern Recognition.

Surmounting specularity

The Adib Group previously demonstrated the use of millimeter wave (mmWave) signals to create accurate reconstructions of 3D objects that are hidden from view, like a lost wallet buried under a pile.

These waves, which are the same type of signals used in Wi-Fi, can pass through common obstructions like drywall, plastic, and cardboard, and reflect off hidden objects.

But mmWaves usually reflect in a specular manner, which means a wave reflects in a single direction after striking a surface. So large portions of the surface will reflect signals away from the mmWave sensor, making those areas effectively invisible.

“When we want to reconstruct an object, we are only able to see the top surface and we can’t see any of the bottom or sides,” Dodds explains.

The researchers previously used principles from physics to interpret reflected signals, but this limits the accuracy of the reconstructed 3D shape.

In the new papers, they overcame that limitation by using a generative AI model to fill in parts that are missing from a partial reconstruction.

“But the challenge then becomes: How do you train these models to fill in these gaps?” Adib says.

Usually, researchers use extremely large datasets to train a generative AI model, which is one reason models like Claude and Llama exhibit such impressive performance. But no mmWave datasets are large enough for training.

Instead, the researchers adapted the images in large computer vision datasets to mimic the properties in mmWave reflections.

“We were simulating the property of specularity and the noise we get from these reflections so we can apply existing datasets to our domain. It would have taken years for us to collect enough new data to do this,” Lam says.

The researchers embed the physics of mmWave reflections directly into these adapted data, creating a synthetic dataset they use to teach a generative AI model to perform plausible shape reconstructions.

The complete system, called Wave-Former, proposes a set of potential object surfaces based on mmWave reflections, feeds them to the generative AI model to complete the shape, and then refines the surfaces until it achieves a full reconstruction.

Wave-Former was able to generate faithful reconstructions of about 70 everyday objects, such as cans, boxes, utensils, and fruit, boosting accuracy by nearly 20 percent over state-of-the-art baselines. The objects were hidden behind or under cardboard, wood, drywall, plastic, and fabric.

Seeing “ghosts”

The team used this same approach to build an expanded system that fully reconstructs entire indoor scenes by leveraging mmWave reflections off humans moving in a room.

Human motion generates multipath reflections. Some mmWaves reflect off the human, then reflect again off a wall or object, and then arrive back at the sensor, Dodds explains.

These secondary reflections create so-called “ghost signals,” which are reflected copies of the original signal that change location as a human moves. These ghost signals are usually discarded as noise, but they also hold information about the layout of the room.

“By analyzing how these reflections change over time, we can start to get a coarse understanding of the environment around us. But trying to directly interpret these signals is going to be limited in accuracy and resolution.” Dodds says.

They used a similar training method to teach a generative AI model to interpret those coarse scene reconstructions and understand the behavior of multipath mmWave reflections. This model fills in the gaps, refining the initial reconstruction until it completes the scene.

They tested their scene reconstruction system, called RISE, using more than 100 human trajectories captured by a single mmWave radar. On average, RISE generated reconstructions that were about twice as precise than existing techniques.

In the future, the researchers want to improve the granularity and detail in their reconstructions. They also want to build large foundation models for wireless signals, like the foundation models GPT, Claude, and Gemini for language and vision, which could open new applications.

This work is supported, in part, by the National Science Foundation (NSF), the MIT Media Lab, and Amazon.

Large language models (LLMs) can generate credible but inaccurate responses, so researchers have developed uncertainty quantification methods to check the reliability of predictions. One popular method involves submitting the same prompt multiple times to see if the model generates the same answer.

But this method measures self-confidence, and even the most impressive LLM might be confidently wrong. Overconfidence can mislead users about the accuracy of a prediction, which might result in devastating consequences in high-stakes settings like health care or finance.   

To address this shortcoming, MIT researchers introduced a new method for measuring a different type of uncertainty that more reliably identifies confident but incorrect LLM responses.

Their method involves comparing a target model’s response to responses from a group of similar LLMs. They found that measuring cross-model disagreement more accurately captures this type of uncertainty than traditional approaches.

They combined their approach with a measure of LLM self-consistency to create a total uncertainty metric, and evaluated it on 10 realistic tasks, such as question-answering and math reasoning. This total uncertainty metric consistently outperformed other measures and was better at identifying unreliable predictions.

“Self-consistency is being used in a lot of different approaches for uncertainty quantification, but if your estimate of uncertainty only relies on a single model’s outcome, it is not necessarily trustable. We went back to the beginning to understand the limitations of current approaches and used those as a starting point to design a complementary method that can empirically improve the results,” says Kimia Hamidieh, an electrical engineering and computer science (EECS) graduate student at MIT and lead author of a paper on this technique.

She is joined on the paper by Veronika Thost, a research scientist at the MIT-IBM Watson AI Lab; Walter Gerych, a former MIT postdoc who is now an assistant professor at Worcester Polytechnic Institute; Mikhail Yurochkin, a staff research scientist at the MIT-IBM Watson AI Lab; and senior author Marzyeh Ghassemi, an associate professor in EECS and a member of the Institute of Medical Engineering Sciences and the Laboratory for Information and Decision Systems.

Understanding overconfidence

Many popular methods for uncertainty quantification involve asking a model for a confidence score or testing the consistency of its responses to the same prompt. These methods estimate aleatoric uncertainty, or how internally confident a model is in its own prediction.

However, LLMs can be confident when they are completely wrong. Research has shown that epistemic uncertainty, or uncertainty about whether one is using the right model, can be a better way to assess true uncertainty when a model is overconfident.

The MIT researchers estimate epistemic uncertainty by measuring disagreement across a similar group of LLMs.    

“If I ask ChatGPT the same question multiple times and it gives me the same answer over and over again, that doesn’t mean the answer is necessarily correct. If I switch to Claude or Gemini and ask them the same question, and I get a different answer, that is going to give me a sense of the epistemic uncertainty,” Hamidieh explains.

Epistemic uncertainty attempts to capture how far a target model diverges from the ideal model for that task. But since it is impossible to build an ideal model, researchers use surrogates or approximations that often rely on faulty assumptions.

To improve uncertainty quantification, the MIT researchers needed a more accurate way to estimate epistemic uncertainty.

An ensemble approach

The method they developed involves measuring the divergence between the target model and a small ensemble of models with similar size and architecture. They found that comparing semantic similarity, or how closely the meanings of the responses match, could provide a better estimate of epistemic uncertainty.

To achieve the most accurate estimate, the researchers needed a set of LLMs that covered diverse responses, weren’t too similar to the target model, and were weighted based on credibility.

“We found that the easiest way to satisfy all these properties is to take models that are trained by different companies. We tried many different approaches that were more complex, but this very simple approach ended up working best,” Hamidieh says.

Once they had developed this method for estimating epistemic uncertainty, they combined it with a standard approach that measures aleatoric uncertainty. This total uncertainty metric (TU) offered the most accurate reflection of whether a model’s confidence level is trustworthy.

“Uncertainty depends on the uncertainty of the given prompt as well as how close our model is to the optimal model. This is why summing up these two uncertainty metrics is going to give us the best estimate,” Hamidieh says.

TU could more effectively identify situations where an LLM is hallucinating, since epistemic uncertainty can flag confidently wrong outputs that aleatoric uncertainty might miss. It could also enable researchers to reinforce an LLM’s confidently correct answers during training, which may improve performance.

They tested TU using multiple LLMs on 10 common tasks, such as question-answering, summarization, translation, and math reasoning. Their method more effectively identified unreliable predictions than either measure on its own.

Measuring total uncertainty often required fewer queries than calculating aleatoric uncertainty, which could reduce computational costs and save energy.

Their experiments also revealed that epistemic uncertainty is most effective on tasks with a unique correct answer, like factual question-answering, but may underperform on more open-ended tasks.

In the future, the researchers could adapt their technique to improve its performance on open-ended queries. They may also build on this work by exploring other forms of aleatoric uncertainty.

This work is funded, in part, by the MIT-IBM Watson AI Lab.