There are some jobs human bodies just weren’t meant to do. Unloading trucks and shipping containers is a repetitive, grueling task — and a big reason warehouse injury rates are more than twice the national average.

The Pickle Robot Company wants its machines to do the heavy lifting. The company’s one-armed robots autonomously unload trailers, picking up boxes weighing up to 50 pounds and placing them onto onboard conveyor belts for warehouses of all types.

The company name, an homage to The Apple Computer Company, hints at the ambitions of founders AJ Meyer ’09, Ariana Eisenstein ’15, SM ’16, and Dan Paluska ’97, SM ’00. The founders want to make the company the technology leader for supply chain automation.

The company’s unloading robots combine generative AI and machine-learning algorithms with sensors, cameras, and machine-vision software to navigate new environments on day one and improve performance over time. Much of the company’s hardware is adapted from industrial partners. You may recognize the arm, for instance, from car manufacturing lines — though you may not have seen it in bright pickle-green.

The company is already working with customers like UPS, Ryobi Tools, and Yusen Logistics to take a load off warehouse workers, freeing them to solve other supply chain bottlenecks in the process.

“Humans are really good edge-case problem solvers, and robots are not,” Paluska says. “How can the robot, which is really good at the brute force, repetitive tasks, interact with humans to solve more problems? Human bodies and minds are so adaptable, the way we sense and respond to the environment is so adaptable, and robots aren’t going to replace that anytime soon. But there’s so much drudgery we can get rid of.”

Finding problems for robots

Meyer and Eisenstein majored in computer science and electrical engineering at MIT, but they didn’t work together until after graduation, when Meyer started the technology consultancy Leaf Labs, which specializes in building embedded computer systems for things like robots, cars, and satellites.

“A bunch of friends from MIT ran that shop,” Meyer recalls, noting it’s still running today. “Ari worked there, Dan consulted there, and we worked on some big projects. We were the primary software and digital design team behind Project Ara, a smartphone for Google, and we worked on a bunch of interesting government projects. It was really a lifestyle company for MIT kids. But 10 years go by, and we thought, ‘We didn’t get into this to do consulting. We got into this to do robots.’”

When Meyer graduated in 2009, problems like robot dexterity seemed insurmountable. By 2018, the rise of algorithmic approaches like neural networks had brought huge advances to robotic manipulation and navigation.

To figure out what problem to solve with robots, the founders talked to people in industries as diverse as agriculture, food prep, and hospitality. At some point, they started visiting logistics warehouses, bringing a stopwatch to see how long it took workers to complete different tasks.

“In 2018, we went to a UPS warehouse and watched 15 guys unloading trucks during a winter night shift,” Meyer recalls. “We spoke to everyone, and not a single person had worked there for more than 90 days. We asked, ‘Why not?’ They laughed at us. They said, ‘Have you tried to do this job before?’”

It turns out warehouse turnover is one of the industry’s biggest problems, limiting productivity as managers constantly grapple with hiring, onboarding, and training.

The founders raised a seed funding round and built robots that could sort boxes because it was an easier problem that allowed them to work with technology like grippers and barcode scanners. Their robots eventually worked, but the company wasn’t growing fast enough to be profitable. Worse yet, the founders were having trouble raising money.

“We were desperately low on funds,” Meyer recalls. “So we thought, ‘Why spend our last dollar on a warm-up task?’”

With money dwindling, the founders built a proof-of-concept robot that could unload trucks reliably for about 20 seconds at a time and posted a video of it on YouTube. Hundreds of potential customers reached out. The interest was enough to get investors back on board to keep the company alive.

The company piloted its first unloading system for a year with a customer in the desert of California, sparing human workers from unloading shipping containers that can reach temperatures up to 130 degrees in the summer. It has since scaled deployments with multiple customers and gained traction among third-party logistics centers across the U.S.

The company’s robotic arm is made by the German industrial robotics giant KUKA. The robots are mounted on a custom mobile base with an onboard computing systems so they can navigate to docks and adjust their positions inside trailers autonomously while lifting. The end of each arm features a suction gripper that clings to packages and moves them to the onboard conveyor belt.

The company’s robots can pick up boxes ranging in size from 5-inch cubes to 24-by-30 inch boxes. The robots can unload anywhere from 400 to 1,500 cases per hour depending on size and weight. The company fine tunes pre-trained generative AI models and uses a number of smaller models to ensure the robot runs smoothly in every setting.

The company is also developing a software platform it can integrate with third-party hardware, from humanoid robots to autonomous forklifts.

“Our immediate product roadmap is load and unload,” Meyer says. “But we’re also hoping to connect these third-party platforms. Other companies are also trying to connect robots. What does it mean for the robot unloading a truck to talk to the robot palletizing, or for the forklift to talk to the inventory drone? Can they do the job faster? I think there’s a big network coming in which we need to orchestrate the robots and the automation across the entire supply chain, from the mines to the factories to your front door.”

“Why not us?”

The Pickle Robot Company employs about 130 people in its office in Charlestown, Massachusetts, where a standard — if green — office gives way to a warehouse where its robots can be seen loading boxes onto conveyor belts alongside human workers and manufacturing lines.

This summer, Pickle will be ramping up production of a new version of its system, with further plans to begin designing a two-armed robot sometime after that.

“My supervisor at Leaf Labs once told me ‘No one knows what they’re doing, so why not us?’” Eisenstein says. “I carry that with me all the time. I’ve been very lucky to be able to work with so many talented, experienced people in my career. They all bring their own skill sets and understanding. That’s a massive opportunity — and it’s the only way something as hard as what we’re doing is going to work.”

Moving forward, the company sees many other robot-shaped problems for its machines.

“We didn’t start out by saying, ‘Let’s load and unload a truck,’” Meyers says. “We said, ‘What does it take to make a great robot business?’ Unloading trucks is the first chapter. Now we’ve built a platform to make the next robot that helps with more jobs, starting in logistics but then ultimately in manufacturing, retail, and hopefully the entire supply chain.”

Around 25 million Americans have rare genetic diseases, and many of them struggle with not only a lack of effective treatments, but also a lack of good information about their disease. Clinicians may not know what causes a patient’s symptoms, know how their disease will progress, or even have a clear diagnosis. Researchers have looked to the human genome for answers, and many disease-causing genetic mutations have been identified, but as many as 70 percent of patients still lack a clear genetic explanation.

In a paper published in Molecular Cell on Nov. 7, Whitehead Institute for Biomedical Research member Iain Cheeseman, graduate student Jimmy Ly, and colleagues propose that researchers and clinicians may be able to get more information from patients’ genomes by looking at them in a different way.

The common wisdom is that each gene codes for one protein. Someone studying whether a patient has a mutation or version of a gene that contributes to their disease will therefore look for mutations that affect the “known” protein product of that gene. However, Cheeseman and others are finding that the majority of genes code for more than one protein. That means that a mutation that might seem insignificant because it does not appear to affect the known protein could nonetheless alter a different protein made by the same gene. Now, Cheeseman and Ly have shown that mutations affecting one or multiple proteins from the same gene can contribute differently to disease.

In their paper, the researchers first share what they have learned about how cells make use of the ability to generate different versions of proteins from the same gene. Then, they examine how mutations that affect these proteins contribute to disease. Through a collaboration with co-author Mark Fleming, the pathologist-in-chief at Boston Children’s Hospital, they provide two case studies of patients with atypical presentations of a rare anemia linked to mutations that selectively affect only one of two proteins produced by the gene implicated in the disease.

“We hope this work demonstrates the importance of considering whether a gene of interest makes multiple versions of a protein, and what the role of each version is in health and disease,” Ly says. “This information could lead to better understanding of the biology of disease, better diagnostics, and perhaps one day to tailored therapies to treat these diseases.”

Cells have several ways to make different versions of a protein, but the variation that Cheeseman and Ly study happens during protein production from genetic code. Cellular machines build each protein according to the instructions within a genetic sequence that begins at a “start codon” and ends at a “stop codon.” However, some genetic sequences contain more than one start codon, many of them hiding in plain sight. If the cellular machinery skips the first start codon and detects a second one, it may build a shorter version of the protein. In other cases, the machinery may detect a section that closely resembles a start codon at a point earlier in the sequence than its typical starting place, and build a longer version of the protein.

These events may sound like mistakes: the cell’s machinery accidentally creating the wrong version of the correct protein. To the contrary, protein production from these alternate starting places is an important feature of cell biology that exists across species. When Ly traced when certain genes evolved to produce multiple proteins, he found that this is a common, robust process that has been preserved throughout evolutionary history for millions of years.

Ly shows that one function this serves is to send versions of a protein to different parts of the cell. Many proteins contain ZIP code-like sequences that tell the cell’s machinery where to deliver them so the proteins can do their jobs. Ly found many examples in which longer and shorter versions of the same protein contained different ZIP codes and ended up in different places within the cell.

In particular, Ly found many cases in which one version of a protein ended up in mitochondria, structures that provide energy to cells, while another version ended up elsewhere. Because of the mitochondria’s role in the essential process of energy production, mutations to mitochondrial genes are often implicated in disease.

Ly wondered what would happen when a disease-causing mutation eliminates one version of a protein but leaves the other intact, causing the protein to only reach one of its two intended destinations. He looked through a database containing genetic information from people with rare diseases to see if such cases existed, and found that they did. In fact, there may be tens of thousands of such cases. However, without access to the people, Ly had no way of knowing what the consequences of this were in terms of symptoms and severity of disease.

Meanwhile, Cheeseman, who is also a professor of biology at MIT, had begun working with Boston Children’s Hospital to foster collaborations between Whitehead Institute and the hospital’s researchers and clinicians to accelerate the pathway from research discovery to clinical application. Through these efforts, Cheeseman and Ly met Fleming.

One group of Fleming’s patients have a type of anemia called SIFD — sideroblastic anemia with B-cell immunodeficiency, periodic fevers, and developmental delay — that is caused by mutations to the TRNT1 gene. TRNT1 is one of the genes Ly had identified as producing a mitochondrial version of its protein and another version that ends up elsewhere: in the nucleus.

Fleming shared anonymized patient data with Ly, and Ly found two cases of interest in the genetic data. Most of the patients had mutations that impaired both versions of the protein, but one patient had a mutation that eliminated only the mitochondrial version of the protein, while another patient had a mutation that eliminated only the nuclear version.

When Ly shared his results, Fleming revealed that both of those patients had very atypical presentations of SIFD, supporting Ly’s hypothesis that mutations affecting different versions of a protein would have different consequences. The patient who only had the mitochondrial version was anemic, but developmentally normal. The patient missing the mitochondrial version of the protein did not have developmental delays or chronic anemia, but did have other immune symptoms, and was not correctly diagnosed until his 50s. There are likely other factors contributing to each patient’s exact presentation of the disease, but Ly’s work begins to unravel the mystery of their atypical symptoms.

Cheeseman and Ly want to make more clinicians aware of the prevalence of genes coding for more than one protein, so they know to check for mutations affecting any of the protein versions that could contribute to disease. For example, several TRNT1 mutations that only eliminate the shorter version of the protein are not flagged as disease-causing by current assessment tools. Cheeseman lab researchers, including Ly and graduate student Matteo Di Bernardo, are now developing a new assessment tool for clinicians, called SwissIsoform, that will identify relevant mutations that affect specific protein versions, including mutations that would otherwise be missed.

“Jimmy and Iain’s work will globally support genetic disease variant interpretation and help with connecting genetic differences to variation in disease symptoms,” Fleming says. “In fact, we have recently identified two other patients with mutations affecting only the mitochondrial versions of two other proteins, who similarly have milder symptoms than patients with mutations that affect both versions.”

Long term, the researchers hope that their discoveries could aid in understanding the molecular basis of disease and in developing new gene therapies: Once researchers understand what has gone wrong within a cell to cause disease, they are better equipped to devise a solution. More immediately, the researchers hope that their work will make a difference by providing better information to clinicians and people with rare diseases.

“As a basic researcher who doesn’t typically interact with patients, there’s something very satisfying about knowing that the work you are doing is helping specific people,” Cheeseman says. “As my lab transitions to this new focus, I’ve heard many stories from people trying to navigate a rare disease and just get answers, and that has been really motivating to us, as we work to provide new insights into the disease biology.”

Each year, faculty and researchers across the MIT School of Engineering are recognized with prestigious awards for their contributions to research, technology, society, and education. To celebrate these achievements, the school periodically highlights select honors received by members of its departments, institutes, labs, and centers. The following individuals were recognized in summer 2025:

Iwnetim Abate, the Chipman Career Development Professor and assistant professor in the Department of Materials Science and Engineering, was honored as one of MIT Technology Review’s 2025 Innovators Under 35. He was recognized for his research on sodium-ion batteries and ammonia production.

Daniel G. Anderson, the Joseph R. Mares (1924) Professor in the Department of Chemical Engineering and the Institute of Medical Engineering and Science (IMES), received the 2025 AIChE James E. Bailey Award. The award honors outstanding contributions in biological engineering and commemorates the pioneering work of James Bailey.

Regina Barzilay, the School of Engineering Distinguished Professor for AI and Health in the Department of Electrical Engineering and Computer Science (EECS), was named to Time’s AI100 2025 list, recognizing her groundbreaking work in AI and health.

Richard D. Braatz, the Edwin R. Gilliland Professor in the Department of Chemical Engineering, received the 2025 AIChE CAST Distinguished Service Award. The award recognizes exceptional service and leadership within the Computing and Systems Technology Division of AIChE.

Rodney Brooks, the Panasonic Professor of Robotics, Emeritus in the Department of Electrical Engineering and Computer Science, was elected to the National Academy of Sciences, one of the highest honors in scientific research.

Arup K. Chakraborty, the John M. Deutch (1961) Institute Professor in the Department of Chemical Engineering and IMES, received the 2025 AIChE Alpha Chi Sigma Award. This award honors outstanding accomplishments in chemical engineering research over the past decade.

Connor W. Coley, the Class of 1957 Career Development Professor and associate professor in the departments of Chemical Engineering and EECS, received the 2025 AIChE CoMSEF Young Investigator Award for Modeling and Simulation. The award recognizes outstanding research in computational molecular science and engineering. Coley was also one of 74 highly accomplished, early-career engineers selected to participate in the Grainger Foundation Frontiers of Engineering Symposium, a signature activity of the National Academy of Engineering.

Henry Corrigan-Gibbs, the Douglas Ross (1954) Career Development Professor of Software Technology and associate professor in the Department of EECS, received the Google ML and Systems Junior Faculty Award, presented to assistant professors who are leading the analysis, design and implementation of efficient, scalable, secure, and trustworthy computing systems.

Christina Delimitrou, the KDD Career Development Professor in Communications and Technology and associate professor in the Department of EECS, received the Google ML and Systems Junior Faculty Award. The award supports assistant professors advancing scalable and trustworthy computing systems for machine learning and cloud computing. Delimitrou also received the Google ML and Systems Junior Faculty Award, presented to assistant professors who are leading the analysis, design, and implementation of efficient, scalable, secure, and trustworthy computing systems.

Priya Donti, the Silverman (1968) Family Career Development Professor and assistant professor in the Department of EECS, was named to Time’s AI100 2025 list, which honors innovators reshaping the world through artificial intelligence.

Joel Emer, a professor of the practice in the Department of EECS, received the Alan D. Berenbaum Distinguished Service Award from ACM SIGARCH. He was honored for decades of mentoring and leadership in the computer architecture community.

Roger Greenwood Mark, the Distinguished Professor of Health Sciences and Technology, Emeritus in IMES, received the IEEE Biomedical Engineering Award for leadership in ECG signal processing and global dissemination of curated biomedical and clinical databases, thereby accelerating biomedical research worldwide.

Ali Jadbabaie, the JR East Professor and head of the Department of Civil and Environmental Engineering, received the 2025 Multidisciplinary University Research Initiative (MURI) award for research projects in areas of critical importance to national defense.

Yoon Kim, associate professor in the Department of EECS, received the Google ML and Systems Junior Faculty Award, presented to assistant professors who are leading the analysis, design, and implementation of efficient, scalable, secure, and trustworthy computing systems.

Mathias Kolle, an associate professor in the Department of Mechanical Engineering, received the 2025 Multidisciplinary University Research Initiative (MURI) award for research projects in areas of critical importance to national defense.

Muriel Médard, the NEC Professor of Software Science and Engineering in the Department of EECS, was elected an International Fellow of the United Kingdom's Royal Academy of Engineering. The honor recognizes exceptional contributions to engineering and technology across sectors.

Pablo Parrilo, the Joseph F. and Nancy P. Keithley Professor in Electrical Engineering in the Department of EECS, received the 2025 INFORMS Computing Society Prize. The award honors outstanding contributions at the interface of computing and operations research. Parrilo was recognized for pioneering work on accelerating gradient descent through stepsize hedging, introducing concepts such as Silver Stepsizes and recursive gluing.

Nidhi Seethapathi, the Frederick A. (1971) and Carole J. Middleton Career Development Professor of Neuroscience and assistant professor in the Department of EECS, was named to MIT Technology Review’s “2025 Innovators Under 35” list. The honor celebrates early-career scientists and entrepreneurs driving real-world impact.

Justin Solomon, an associate professor in the Department of EECS, was named a 2025 Schmidt Science Polymath. The award supports novel, early-stage research across disciplines, including acoustics and climate simulation.

Martin Staadecker, a research assistant in the Sustainable Supply Chain Lab, received the MIT-GE Vernova Energy and Climate Alliance Technology and Policy Program Project Award. The award recognizes his work on Scope 3 emissions and sustainable supply chain practices.

Antonio Torralba, the Delta Electronics Professor and faculty head of AI+D in the Department of EECS, received the 2025 Multidisciplinary University Research Initiative (MURI) award for research projects in areas of critical importance to national defense.

Ryan Williams, a professor in the Department of EECS, received the Best Paper Award at STOC 2025 for his paper “Simulating Time With Square-Root Space,” recognized for its technical merit and originality. Williams was also selected as a Member of the Institute for Advanced Study for the 2025–26 academic year. This prestigious fellowship recognizes the significance of these scholars' work, and it is an opportunity to advance their research and exchange ideas with scholars from around the world.

Gioele Zardini, the Rudge (1948) and Nancy Allen Career Development Professor in the Department of Civil and Environmental Engineering, received the 2025 DARPA Young Faculty Award. The award supports rising stars among early-career faculty, helping them develop research ideas aligned with national security needs.

There are several narratives surrounding the American Revolution, a well-traveled and -documented series of events leading to the drafting and signing of the Declaration of Independence and the war that followed. 

MIT philosopher Brad Skow is taking a new approach to telling this story: a collection of 47 poems about the former American colonies’ journey from England’s imposition of the Stamp Act in 1765 to the war for America’s independence that began in 1775.

When asked why he chose poetry to retell the story, Skow, the Laurence S. Rockefeller Professor in the Department of Linguistics and Philosophy, said he “wanted to take just the great bits of these speeches and writings, while maintaining their intent and integrity.” Poetry, Skow argues, allows for that kind of nuance and specificity.

“American Independence in Verse,” published by Pentameter Press, traces a story of America’s origins through a collection of vignettes featuring some well-known characters, like politician and orator Patrick Henry, alongside some lesser-known but no less important ones, like royalist and former chief justice of North Carolina Martin Howard. Each is rendered in blank verse, a nursery-style rhyme, or free verse. 

The book is divided into three segments: “Taxation Without Representation,” “Occupation and Massacre,” and “War and Independence.” Themes like freedom, government, and authority, rendered in a style of writing and oratory seldom seen today, lent themselves to being reimagined as poems. “The options available with poetic license offer opportunities for readers that might prove more difficult with prose,” Skow reports.

Skow based each of the poems on actual speeches, letters, pamphlets, and other printed materials produced by people on both sides of the debate about independence. “While reviewing a variety of primary sources for the book, I began to see the poetry in them,” he says. 

In the poem “Everywhere, the spirit of equality prevails,” during an “Interlude” between the “Occupation and Massacre” and “War and Independence” sections of the book, British commissioner of customs Henry Hulton, writing to Robert Nicholson in Liverpool, England, describes the America he experienced during a trip with his wife:

The spirit of equality prevails.

Regarding social differences, they’ve no

Notion of rank, and will show more respect

To one another than to those above them.

They’ll ask a thousand strange impertinent 

Questions, sit down when they should wait at a table,

React with puzzlement when you do not

Invite your valet to come share your meal.

Here, Skow, using Hulton’s words, illustrates the tension between agreed-upon social conventions — remnants of the Old World — and the society being built in the New World that animates a portion of the disconnect leading both toward war. “These writings are really powerful, and poetry offers a way to convey that power,” Skow says.

The journey to the printed page 

Skow’s interest in exploring the American Revolution came, in part, from watching the Emmy Award-winning play “Hamilton.” The book ends where the play begins. “It led me to want to learn more,” he says of the play and his experience watching it. “Its focus on the Revolution made the era more exciting for me.”

While conducting research for another poetry project, Skow read an interview with American diplomat, inventor, and publisher Benjamin Franklin in the House of Commons conducted in 1766. “There were lots of amazing poetic moments in the interview,” he says. Skow began reading additional pamphlets, letters, and other writings, disconnecting his work as a philosopher from the research that would yield the book.

“I wanted to remove my philosopher hat with this project,” he says. “Poetry can encourage ambiguity and, unlike philosophy, can focus on emotional and non-rational connections between ideas.” 

Although eager to approach the work as a poet and author, rather than a philosopher, Skow discovered that more primary sources than he expected were themselves often philosophical treatises. “Early in the resistance movement there were sophisticated arguments, often printed in newspapers, that it was unjust to tax the colonies without granting them representation in Parliament,” he notes. 

A series of new perspectives and lessons

Skow made some discoveries that further enhanced his passion for the project. “Samuel Adams is an important figure who isn’t as well-known as he should be,” he says. “I wanted to raise his profile.”

Skow also notes that American separatists used strong-arm tactics to “encourage” support for independence, and that prevailing narratives regarding America and its eventual separation from England are more complex and layered than we might believe. “There were arguments underway about legitimate forms of government and which kind of government was right,” he says, “and many Americans wanted to retain the existing relationship with England.”

Skow says the American Revolution is a useful benchmark when considering subsequent political movements, a notion he hopes readers will take away from the book. “The book is meant to be fun and not just a collection of dry, abstract ideas,” he believes. 

“There’s a simple version of the independence story we tell when we’re in a hurry; and there is the more complex truth, printed in long history books,” he continues. “I wanted to write something that was both short and included a variety of perspectives.”

Skow believes the book and its subjects are a testament to ideas he’d like to see return to political and practical discourse. “The ideals around which this country rallied for its independence are still good ideals, and the courage the participants exhibited is still worth admiring,” he says.

Growing energy demand means the U.S. will almost certainly have to expand its electricity grid in coming years. What’s the best way to do this? A new study by MIT researchers examines legislation introduced in Congress and identifies relative tradeoffs involving reliability, cost, and emissions, depending on the proposed approach.

The researchers evaluated two policy approaches to expanding the U.S. electricity grid: One would concentrate on regions with more renewable energy sources, and the other would create more interconnections across the country. For instance, some of the best untapped wind-power resources in the U.S. lie in the center of the country, so one type of grid expansion would situate relatively more grid infrastructure in those regions. Alternatively, the other scenario involves building more infrastructure everywhere in roughly equal measure, which the researchers call the “prescriptive” approach. How does each pencil out?

After extensive modeling, the researchers found that a grid expansion could make improvements on all fronts, with each approach offering different advantages. A more geographically unbalanced grid buildout would be 1.13 percent less expensive, and would reduce carbon emissions by 3.65 percent compared to the prescriptive approach. And yet, the prescriptive approach, with more national interconnection, would significantly reduce power outages due to extreme weather, among other things.

“There’s a tradeoff between the two things that are most on policymakers’ minds: cost and reliability,” says Christopher Knittel, an economist at the MIT Sloan School of Management, who helped direct the research. “This study makes it more clear that the more prescriptive approach ends up being better in the face of extreme weather and outages.”

The paper, “Implications of Policy-Driven Transmission Expansion on Costs, Emissions and Reliability in the United States,” is published today in Nature Energy.

The authors are Juan Ramon L. Senga, a postdoc in the MIT Center for Energy and Environmental Policy Research; Audun Botterud, a principal research scientist in the MIT Laboratory for Information and Decision Systems; John E. Parson, the deputy director for research at MIT’s Center for Energy and Environmental Policy Research; Drew Story, the managing director at MIT’s Policy Lab; and Knittel, who is the George P. Schultz Professor at MIT Sloan, and associate dean for climate and sustainability at MIT.

The new study is a product of the MIT Climate Policy Center, housed within MIT Sloan and committed to bipartisan research on energy issues. The center is also part of the Climate Project at MIT, founded in 2024 as a high-level Institute effort to develop practical climate solutions.

In this case, the project was developed from work the researchers did with federal lawmakers who have introduced legislation aimed at bolstering and expanding the U.S. electric grid. One of these bills, the BIG WIRES Act, co-sponsored by Sen. John Hickenlooper of Colorado and Rep. Scott Peters of California, would require each transmission region in the U.S. to be able to send at least 30 percent of its peak load to other regions by 2035.

That would represent a substantial change for a national transmission scenario where grids have largely been developed regionally, without an enormous amount of national oversight.

“The U.S. grid is aging and it needs an upgrade,” Senga says. “Implementing these kinds of policies is an important step for us to get to that future where we improve the grid, lower costs, lower emissions, and improve reliability. Some progress is better than none, and in this case, it would be important.”

To conduct the study, the researchers looked at how policies like the BIG WIRES Act would affect energy distribution. The scholars used a model of energy generation developed at the MIT Energy Initiative — the model is called “Gen X” — and examined the changes proposed by the legislation.

With a 30 percent level of interregional connectivity, the study estimates, the number of outages due to extreme cold would drop by 39 percent, for instance, a substantial increase in reliability. That would help avoid scenarios such as the one Texas experienced in 2021, when winter storms damaged distribution capacity.

“Reliability is what we find to be most salient to policymakers,” Senga says.

On the other hand, as the paper details, a future grid that is “optimized” with more transmission capacity near geographic spots of new energy generation would be less expensive.

“On the cost side, this kind of optimized system looks better,” Senga says.

A more geographically imbalanced grid would also have a greater impact on reducing emissions. Globally, the levelized cost of wind and solar dropped by 89 percent and 69 percent, respectively, from 2010 to 2022, meaning that incorporating less-expensive renewables into the grid would help with both cost and emissions.

“On the emissions side, a priori it’s not clear the optimized system would do better, but it does,” Knittel says. “That’s probably tied to cost, in the sense that it’s building more transmission links to where the good, cheap renewable resources are, because they’re cheap. Emissions fall when you let the optimizing action take place.”

To be sure, these two differing approaches to grid expansion are not the only paths forward. The study also examines a hybrid approach, which involves both national interconnectivity requirements and local buildouts based around new power sources on top of that. Still, the model does show that there may be some tradeoffs lawmakers will want to consider when developing and considering future grid legislation.

“You can find a balance between these factors, where you’re still going to still have an increase in reliability while also getting the cost and emission reductions,” Senga observes.

For his part, Knittel emphasizes that working with legislation as the basis for academic studies, while not generally common, can be productive for everyone involved. Scholars get to apply their research tools and models to real-world scenarios, and policymakers get a sophisticated evaluation of how their proposals would work.

“Compared to the typical academic path to publication, this is different, but at the Climate Policy Center, we’re already doing this kind of research,” Knittel says. 

To make large language models (LLMs) more accurate when answering harder questions, researchers can let the model spend more time thinking about potential solutions.

But common approaches that give LLMs this capability set a fixed computational budget for every problem, regardless of how complex it is. This means the LLM might waste computational resources on simpler questions or be unable to tackle intricate problems that require more reasoning.

To address this, MIT researchers developed a smarter way to allocate computational effort as the LLM solves a problem. Their method enables the model to dynamically adjust its computational budget based on the difficulty of the question and the likelihood that each partial solution will lead to the correct answer.

The researchers found that their new approach enabled LLMs to use as little as one-half the computation as existing methods, while achieving comparable accuracy on a range of questions with varying difficulties. In addition, their method allows smaller, less resource-intensive LLMs to perform as well as or even better than larger models on complex problems.

By improving the reliability and efficiency of LLMs, especially when they tackle complex reasoning tasks, this technique could reduce the energy consumption of generative AI systems and enable the use of LLMs in more high-stakes and time-sensitive applications.

“The computational cost of inference has quickly become a major bottleneck for frontier model providers, and they are actively trying to find ways to improve computational efficiency per user queries. For instance, the recent GPT-5.1 release highlights the efficacy of the ‘adaptive reasoning’ approach our paper proposes. By endowing the models with the ability to know what they don’t know, we can enable them to spend more compute on the hardest problems and most promising solution paths, and use far fewer tokens on easy ones. That makes reasoning both more reliable and far more efficient,” says Navid Azizan, the Alfred H. and Jean M. Hayes Career Development Assistant Professor in the Department of Mechanical Engineering and the Institute for Data, Systems, and Society (IDSS), a principal investigator of the Laboratory for Information and Decision Systems (LIDS), and the senior author of a paper on this technique.

Azizan is joined on the paper by lead author Young-Jin Park, a LIDS/MechE graduate student; Kristjan Greenewald, a research scientist in the MIT-IBM Watson AI Lab; Kaveh Alim, an IDSS graduate student; and Hao Wang, a research scientist at the MIT-IBM Watson AI Lab and the Red Hat AI Innovation Team. The research is being presented this week at the Conference on Neural Information Processing Systems.

Computation for contemplation

A recent approach called inference-time scaling lets a large language model take more time to reason about difficult problems.

Using inference-time scaling, the LLM might generate multiple solution attempts at once or explore different reasoning paths, then choose the best ones to pursue from those candidates.

A separate model, known as a process reward model (PRM), scores each potential solution or reasoning path. The LLM uses these scores to identify the most promising ones.     

Typical inference-time scaling approaches assign a fixed amount of computation for the LLM to break the problem down and reason about the steps.

Instead, the researchers’ method, known as instance-adaptive scaling, dynamically adjusts the number of potential solutions or reasoning steps based on how likely they are to succeed, as the model wrestles with the problem.

“This is how humans solve problems. We come up with some partial solutions and then decide, should I go further with any of these, or stop and revise, or even go back to my previous step and continue solving the problem from there?” Wang explains.

To do this, the framework uses the PRM to estimate the difficulty of the question, helping the LLM assess how much computational budget to utilize for generating and reasoning about potential solutions.

At every step in the model’s reasoning process, the PRM looks at the question and partial answers and evaluates how promising each one is for getting to the right solution. If the LLM is more confident, it can reduce the number of potential solutions or reasoning trajectories to pursue, saving computational resources.

But the researchers found that existing PRMs often overestimate the model’s probability of success.

Overcoming overconfidence

“If we were to just trust current PRMs, which often overestimate the chance of success, our system would reduce the computational budget too aggressively. So we first had to find a way to better calibrate PRMs to make inference-time scaling more efficient and reliable,” Park says.

The researchers introduced a calibration method that enables PRMs to generate a range of probability scores rather than a single value. In this way, the PRM creates more reliable uncertainty estimates that better reflect the true probability of success.

With a well-calibrated PRM, their instance-adaptive scaling framework can use the probability scores to effectively reduce computation while maintaining the accuracy of the model’s outputs.

When they compared their method to standard inference-time scaling approaches on a series of mathematical reasoning tasks, it utilized less computation to solve each problem while achieving similar accuracy.

“The beauty of our approach is that this adaptation happens on the fly, as the problem is being solved, rather than happening all at once at the beginning of the process,” says Greenewald.

In the future, the researchers are interested in applying this technique to other applications, such as code generation and AI agents. They are also planning to explore additional uses for their PRM calibration method, like for reinforcement learning and fine-tuning.

“Human employees learn on the job — some CEOs even started as interns — but today’s agents remain largely static pieces of probabilistic software. Work like this paper is an important step toward changing that: helping agents understand what they don’t know and building mechanisms for continual self-improvement. These capabilities are essential if we want agents that can operate safely, adapt to new situations, and deliver consistent results at scale,” says Akash Srivastava, director and chief architect of Core AI at IBM Software, who was not involved with this work.

This work was funded, in part, by the MIT-IBM Watson AI Lab, the MIT-Amazon Science Hub, the MIT-Google Program for Computing Innovation, and MathWorks. 

In the future, tiny flying robots could be deployed to aid in the search for survivors trapped beneath the rubble after a devastating earthquake. Like real insects, these robots could flit through tight spaces larger robots can’t reach, while simultaneously dodging stationary obstacles and pieces of falling rubble.

So far, aerial microrobots have only been able to fly slowly along smooth trajectories, far from the swift, agile flight of real insects — until now.

MIT researchers have demonstrated aerial microrobots that can fly with speed and agility that is comparable to their biological counterparts. A collaborative team designed a new AI-based controller for the robotic bug that enabled it to follow gymnastic flight paths, such as executing continuous body flips.

With a two-part control scheme that combines high performance with computational efficiency, the robot’s speed and acceleration increased by about 450 percent and 250 percent, respectively, compared to the researchers’ best previous demonstrations.

The speedy robot was agile enough to complete 10 consecutive somersaults in 11 seconds, even when wind disturbances threatened to push it off course.

“We want to be able to use these robots in scenarios that more traditional quad copter robots would have trouble flying into, but that insects could navigate. Now, with our bioinspired control framework, the flight performance of our robot is comparable to insects in terms of speed, acceleration, and the pitching angle. This is quite an exciting step toward that future goal,” says Kevin Chen, an associate professor in the Department of Electrical Engineering and Computer Science (EECS), head of the Soft and Micro Robotics Laboratory within the Research Laboratory of Electronics (RLE), and co-senior author of a paper on the robot.

Chen is joined on the paper by co-lead authors Yi-Hsuan Hsiao, an EECS MIT graduate student; Andrea Tagliabue PhD ’24; and Owen Matteson, a graduate student in the Department of Aeronautics and Astronautics (AeroAstro); as well as EECS graduate student Suhan Kim; Tong Zhao MEng ’23; and co-senior author Jonathan P. How, the Ford Professor of Engineering in the Department of Aeronautics and Astronautics and a principal investigator in the Laboratory for Information and Decision Systems (LIDS). The research appears today in Science Advances.

An AI controller

Chen’s group has been building robotic insects for more than five years.

They recently developed a more durable version of their tiny robot, a microcassette-sized device that weighs less than a paperclip. The new version utilizes larger, flapping wings that enable more agile movements. They are powered by a set of squishy artificial muscles that flap the wings at an extremely fast rate.

But the controller — the “brain” of the robot that determines its position and tells it where to fly — was hand-tuned by a human, limiting the robot’s performance.

For the robot to fly quickly and aggressively like a real insect, it needed a more robust controller that could account for uncertainty and perform complex optimizations quickly.

Such a controller would be too computationally intensive to be deployed in real time, especially with the complicated aerodynamics of the lightweight robot.

To overcome this challenge, Chen’s group joined forces with How’s team and, together, they crafted a two-step, AI-driven control scheme that provides the robustness necessary for complex, rapid maneuvers, and the computational efficiency needed for real-time deployment.

“The hardware advances pushed the controller so there was more we could do on the software side, but at the same time, as the controller developed, there was more they could do with the hardware. As Kevin’s team demonstrates new capabilities, we demonstrate that we can utilize them,” How says.

For the first step, the team built what is known as a model-predictive controller. This type of powerful controller uses a dynamic, mathematical model to predict the behavior of the robot and plan the optimal series of actions to safely follow a trajectory.

While computationally intensive, it can plan challenging maneuvers like aerial somersaults, rapid turns, and aggressive body tilting. This high-performance planner is also designed to consider constraints on the force and torque the robot could apply, which is essential for avoiding collisions.

For instance, to perform multiple flips in a row, the robot would need to decelerate in such a way that its initial conditions are exactly right for doing the flip again.

“If small errors creep in, and you try to repeat that flip 10 times with those small errors, the robot will just crash. We need to have robust flight control,” How says.

They use this expert planner to train a “policy” based on a deep-learning model, to control the robot in real time, through a process called imitation learning. A policy is the robot’s decision-making engine, which tells the robot where and how to fly.

Essentially, the imitation-learning process compresses the powerful controller into a computationally efficient AI model that can run very fast.

The key was having a smart way to create just enough training data, which would teach the policy everything it needs to know for aggressive maneuvers.

“The robust training method is the secret sauce of this technique,” How explains.

The AI-driven policy takes robot positions as inputs and outputs control commands in real time, such as thrust force and torques.

Insect-like performance

In their experiments, this two-step approach enabled the insect-scale robot to fly 447 percent faster while exhibiting a 255 percent increase in acceleration. The robot was able to complete 10 somersaults in 11 seconds, and the tiny robot never strayed more than 4 or 5 centimeters off its planned trajectory.

“This work demonstrates that soft and microrobots, traditionally limited in speed, can now leverage advanced control algorithms to achieve agility approaching that of natural insects and larger robots, opening up new opportunities for multimodal locomotion,” says Hsiao.

The researchers were also able to demonstrate saccade movement, which occurs when insects pitch very aggressively, fly rapidly to a certain position, and then pitch the other way to stop. This rapid acceleration and deceleration help insects localize themselves and see clearly.

“This bio-mimicking flight behavior could help us in the future when we start putting cameras and sensors on board the robot,” Chen says.

Adding sensors and cameras so the microrobots can fly outdoors, without being attached to a complex motion capture system, will be a major area of future work.

The researchers also want to study how onboard sensors could help the robots avoid colliding with one another or coordinate navigation.

“For the micro-robotics community, I hope this paper signals a paradigm shift by showing that we can develop a new control architecture that is high-performing and efficient at the same time,” says Chen.

“This work is especially impressive because these robots still perform precise flips and fast turns despite the large uncertainties that come from relatively large fabrication tolerances in small-scale manufacturing, wind gusts of more than 1 meter per second, and even its power tether wrapping around the robot as it performs repeated flips,” says Sarah Bergbreiter, a professor of mechanical engineering at Carnegie Mellon University, who was not involved with this work.

“Although the controller currently runs on an external computer rather than onboard the robot, the authors demonstrate that similar, but less precise, control policies may be feasible even with the more limited computation available on an insect-scale robot. This is exciting because it points toward future insect-scale robots with agility approaching that of their biological counterparts,” she adds.

This research is funded, in part, by the National Science Foundation (NSF), the Office of Naval Research, Air Force Office of Scientific Research, MathWorks, and the Zakhartchenko Fellowship.

With every step we take, our brains are already thinking about the next one. If a bump in the terrain or a minor misstep has thrown us off balance, our stride may need to be altered to prevent a fall. Our two-legged posture makes maintaining stability particularly complex, which our brains solve in part by continually monitoring our bodies and adjusting where we place our feet.

Now, scientists at MIT have determined that animals with very different bodies likely use a shared strategy to balance themselves when they walk.

Nidhi Seethapathi, the Frederick A. and Carole J. Middleton Career Development Assistant Professor in Brain and Cognitive Sciences and Electrical Engineering and Computer Science at MIT, and K. Lisa Yang ICoN Center Fellow Antoine De Comite found that humans, mice, and fruit flies all use an error-correction process to guide foot placement and maintain stability while walking. Their findings, published Oct. 21 in the journal PNAS, could inform future studies exploring how the brain achieves stability during locomotion — bridging the gap between animal models and human balance.

Corrective action

Information must be integrated by the brain to keep us upright when we walk or run. Our steps must be continually adjusted according to the terrain, our desired speed, and our body’s current velocity and position in space.

“We rely on a combination of vestibular, proprioceptive, and visual information to build an estimate of our body’s state, determining if we are about to fall. Once we know the body’s state, we can decide which corrective actions to take,” explains Seethapathi, who is also an associate investigator at the McGovern Institute for Brain Research.

While humans are known to adjust where they place their feet to correct for errors, it is not known whether animals whose bodies are more stable do this, too.

To find out, Seethapathi and De Comite, who is a postdoc in Seethapathi’s and Guoping Feng's lab at the McGovern Institute, turned to locomotion data from mice, fruit flies, and humans shared by other labs, enabling an analysis across species that is otherwise challenging. Importantly, Seethapathi notes, all the animals they studied were walking in everyday natural environments, such as around a room — not on a treadmill or over unusual terrain.

Even in these ordinary circumstances, missteps and minor imbalances are common, and the team’s analysis showed that these errors predicted where all of the animals placed their feet in subsequent steps, regardless of whether they had two, four, or six legs.

One foot in front of another

By tracking the animals’ bodies and the step-by-step placement of their feet, Seethapathi and De Comite were able to find a measure of error that informs each animal’s next step. “By taking this comparative approach, we’ve forced ourselves to come up with a definition of error that generalizes across species,” Seethapathi says. “An animal moves with an expected body state for a particular speed. If it deviates from that ideal state, that deviation — at any given moment — is the error.”

“It was surprising to find similarities across these three species, which, at first sight, look very different,” says DeComite. “The methods themselves are surprising because we now have a pipeline to analyze foot placement and locomotion stability in any legged species,” explains DeComite, “which could lead similar analyses in even more species in the future.”

The team’s data suggest that in all of the species in the study, placement of the feet is guided both by an error-correction process and the speed at which an animal is traveling. Steps tend to lengthen and feet spend less time on the ground as animals pick up their pace, while the width of each step seems to change largely to compensate for body-state errors.

Now, Seethapathi says, we can look forward to future studies to explore how the dual control systems might be generated and integrated in the brain to keep moving bodies stable.

Studying how brains help animals move stably may also guide the development of more-targeted strategies to help people improve their balance and, ultimately, prevent falls.

“In elderly individuals and individuals with sensorimotor disorders, minimizing fall risk is one of the major functional targets of rehabilitation,” says Seethapathi. “A fundamental understanding of the error correction process that helps us remain stable will provide insight into why this process falls short in populations with neural deficits,” she says. 

Peripheral nerves — the network connecting the brain, spinal cord, and central nervous system to the rest of the body — transmit sensory information, control muscle movements, and regulate automatic bodily functions. Bioelectronic devices implanted on these nerves offer remarkable potential for the treatment and rehabilitation of neurological and systemic diseases. However, because the body perceives these implants as foreign objects, they often trigger the formation of dense fibrotic tissue at bioelectronic device–tissue interfaces, which can significantly compromise device performance and longevity.

New research published in the journal Science Advances presents a robust bioadhesive strategy that establishes non-fibrotic bioelectronic interfaces on diverse peripheral nerves — including the occipital, vagus, deep peroneal, sciatic, tibial, and common peroneal nerves — for up to 12 weeks.

“We discovered that adhering the bioelectrodes to peripheral nerves can fully prevent the formation of fibrosis on the interfaces,” says Xuanhe Zhao, the Uncas and Helen Whitaker Professor, and professor of mechanical engineering and civil engineering at MIT. “We further demonstrated long-term, drug-free hypertension mitigation using non-fibrotic bioelectronics over four weeks, and ongoing.”

The approach inhibits immune cell infiltration at the device-tissue interface, thereby preventing the formation of fibrous capsules within the inflammatory microenvironment. In preclinical rodent models, the team demonstrated that the non-fibrotic, adhesive bioelectronic device maintained stable, long-term regulation of blood pressure.

“Our long-term blood pressure regulation approach was inspired by traditional acupuncture,” says Hyunmin Moon, lead author of the study and a postdoc in the Department of Mechanical Engineering. “The lower leg has long been used in hypertension treatment, and the deep peroneal nerve lies precisely at an acupuncture point. We were thrilled to see that stimulating this nerve achieved blood pressure regulation for the first time. The convergence of our non-fibrotic, adhesive bioelectronic device with this long-term regulation capability holds exciting promise for translational medicine.”

Importantly, after 12 weeks of implantation with continuous nerve stimulation, only minimal macrophage activity and limited deposition of smooth muscle actin and collagen were detected, underscoring the device’s potential to deliver long-term neuromodulation without triggering fibrosis. “The contrast between the immune response of the adhered device and that of the non-adhered control is striking,” says Bastien Aymon, a study co-author and a PhD candidate in mechanical engineering. “The fact that we can observe immunologically pristine interfaces after three months of adhesive implantation is extremely encouraging for future clinical translation.”

This work offers a broadly applicable strategy for all implantable bioelectronic systems by preventing fibrosis at the device interface, paving the way for more effective and long-lasting therapies such as hypertension mitigation.

Hypertension is a major contributor to cardiovascular diseases, the leading cause of death worldwide. Although medications are effective in many cases, more than 50 percent of patients remain hypertensive despite treatment — a condition known as resistant hypertension. Traditional carotid sinus or vagus nerve stimulation methods are often accompanied by side effects including apnea, bradycardia, cough, and paresthesia.

“In contrast, our non-fibrotic, adhesive bioelectronic device targeting the deep peroneal nerve enables long-term blood pressure regulation in resistant hypertensive patients without metabolic side effects,” says Moon.

A noninvasive method for measuring blood glucose levels, developed at MIT, could save diabetes patients from having to prick their fingers several times a day.

The MIT team used Raman spectroscopy — a technique that reveals the chemical composition of tissues by shining near-infrared or visible light on them — to develop a shoebox-sized device that can measure blood glucose levels without any needles.

In tests in a healthy volunteer, the researchers found that the measurements from their device were similar to those obtained by commercial continuous glucose monitoring sensors that require a wire to be implanted under the skin. While the device presented in this study is too large to be used as a wearable sensor, the researchers have since developed a wearable version that they are now testing in a small clinical study.

“For a long time, the finger stick has been the standard method for measuring blood sugar, but nobody wants to prick their finger every day, multiple times a day. Naturally, many diabetic patients are under-testing their blood glucose levels, which can cause serious complications,” says Jeon Woong Kang, an MIT research scientist and the senior author of the study. “If we can make a noninvasive glucose monitor with high accuracy, then almost everyone with diabetes will benefit from this new technology.”

MIT postdoc Arianna Bresci is the lead author of the new study, which appears today in the journal Analytical Chemistry. Other authors include Peter So, director of the MIT Laser Biomedical Research Center (LBRC) and an MIT professor of biological engineering and mechanical engineering; and Youngkyu Kim and Miyeon Jue of Apollon Inc., a biotechnology company based in South Korea.

Noninvasive glucose measurement

While most diabetes patients measure their blood glucose levels by drawing blood and testing it with a glucometer, some use wearable monitors, which have a sensor that is inserted just under the skin. These sensors provide continuous glucose measurements from the interstitial fluid, but they can cause skin irritation and they need to be replaced every 10 to 15 days.

In hopes of creating wearable glucose monitors that would be more comfortable for patients, researchers in MIT’s LBRC have been pursuing noninvasive sensors based on Raman spectroscopy. This type of spectroscopy reveals the chemical composition of tissue or cells by analyzing how near-infrared light is scattered, or deflected, as it encounters different kinds of molecules.

In 2010, researchers at the LBRC showed that they could indirectly calculate glucose levels based on a comparison between Raman signals from the interstitial fluid that bathes skin cells and a reference measurement of blood glucose levels. While this approach produced reliable measurements, it wasn’t practical for translating to a glucose monitor.

More recently, the researchers reported a breakthrough that allowed them to directly measure glucose Raman signals from the skin. Normally, this glucose signal is too small to pick out from all of the other signals generated by molecules in tissue. The MIT team found a way to filter out much of the unwanted signal by shining near-infrared light onto the skin at a different angle from which they collected the resulting Raman signal.

The researchers obtained those measurements using equipment that was around the size of a desktop printer, and since then, they have been working on further shrinking the footprint of the device.

In their new study, they were able to create a smaller device by analyzing just three bands — spectral regions that correspond to specific molecular features — in the Raman spectrum.

Typically, a Raman spectrum may contain about 1,000 bands. However, the MIT team found that they could determine blood glucose levels by measuring just three bands — one from the glucose plus two background measurements. This approach allowed the researchers to reduce the amount and cost of equipment needed, allowing them to perform the measurement with a cost-effective device about the size of a shoebox.

“By refraining from acquiring the whole spectrum, which has a lot of redundant information, we go down to three bands selected from about 1,000,” Bresci says. “With this new approach, we can change the components commonly used in Raman-based devices, and save space, time, and cost.”

Toward a wearable sensor

In a clinical study performed at the MIT Center for Clinical Translation Research (CCTR), the researchers used the new device to take readings from a healthy volunteer over a four-hour period. As the subject rested their arm on top of the device, a near-infrared beam shone through a small glass window onto the skin to perform the measurement.

Each measurement takes a little more than 30 seconds, and the researchers took a new reading every five minutes.

During the study, the subject consumed two 75-gram glucose drinks, allowing the researchers to monitor significant changes in blood glucose concentration. They found that the Raman-based device showed accuracy levels similar to those of two commercially available, invasive glucose monitors worn by the subject.

Since finishing that study, the researchers have developed a smaller prototype, about the size of a cellphone, that they’re currently testing at the MIT CCTR as a wearable monitor in healthy and prediabetic volunteers. Next year, they plan to run a larger study working with a local hospital, which will include people with diabetes.

The researchers are also working on making the device even smaller, about the size of a watch. Additionally, they are exploring ways to ensure that the device can obtain accurate readings from people with different skin tones.

The research was funded by the National Institutes of Health, the Korean Technology and Information Promotion Agency for SMEs, and Apollon Inc.

For the first time, MIT chemists have synthesized a fungal compound known as verticillin A, which was discovered more than 50 years ago and has shown potential as an anticancer agent.

The compound has a complex structure that made it more difficult to synthesize than related compounds, even though it differed by only a couple of atoms.

“We have a much better appreciation for how those subtle structural changes can significantly increase the synthetic challenge,” says Mohammad Movassaghi, an MIT professor of chemistry. “Now we have the technology where we can not only access them for the first time, more than 50 years after they were isolated, but also we can make many designed variants, which can enable further detailed studies.”

In tests in human cancer cells, a derivative of verticillin A showed particular promise against a type of pediatric brain cancer called diffuse midline glioma. More tests will be needed to evaluate its potential for clinical use, the researchers say.

Movassaghi and Jun Qi, an associate professor of medicine at Dana-Farber Cancer Institute/Boston Children’s Cancer and Blood Disorders Center and Harvard Medical School, are the senior authors of the study, which appears today in the Journal of the American Chemical Society. Walker Knauss PhD ’24 is the lead author of the paper. Xiuqi Wang, a medicinal chemist and chemical biologist at Dana-Farber, and Mariella Filbin, research director in the Pediatric Neurology-Oncology Program at Dana-Farber/Boston Children’s Cancer and Blood Disorders Center, are also authors of the study.

A complex synthesis

Researchers first reported the isolation of verticillin A from fungi, which use it for protection against pathogens, in 1970. Verticillin A and related fungal compounds have drawn interest for their potential anticancer and antimicrobial activity, but their complexity has made them difficult to synthesize.

In 2009, Movassaghi’s lab reported the synthesis of (+)-11,11'-dideoxyverticillin A, a fungal compound similar to verticillin A. That molecule has 10 rings and eight stereogenic centers, or carbon atoms that have four different chemical groups attached to them. These groups have to be attached in a way that ensures they have the correct orientation, or stereochemistry, with respect to the rest of the molecule.

Once that synthesis was achieved, however, synthesis of verticillin A remained challenging, even though the only difference between verticillin A and (+)-11,11'-dideoxyverticillin A is the presence of two oxygen atoms.

“Those two oxygens greatly limit the window of opportunity that you have in terms of doing chemical transformations,” Movassaghi says. “It makes the compound so much more fragile, so much more sensitive, so that even though we had had years of methodological advances, the compound continued to pose a challenge for us.”

Both of the verticillin A compounds consist of two identical fragments that must be joined together to form a molecule called a dimer. To create (+)-11,11'-dideoxyverticillin A, the researchers had performed the dimerization reaction near the end of the synthesis, then added four critical carbon-sulfur bonds.

Yet when trying to synthesize verticillin A, the researchers found that waiting to add those carbon-sulfur bonds at the end did not result in the correct stereochemistry. As a result, the researchers had to rethink their approach and ended up creating a very different synthetic sequence.

“What we learned was the timing of the events is absolutely critical. We had to significantly change the order of the bond-forming events,” Movassaghi says.

The verticillin A synthesis begins with an amino acid derivative known as beta-hydroxytryptophan, and then step-by-step, the researchers add a variety of chemical functional groups, including alcohols, ketones, and amides, in a way that ensures the correct stereochemistry.

A functional group containing two carbon-sulfur bonds and a disulfide bond were introduced early on, to help control the stereochemistry of the molecule, but the sensitive disulfides had to be “masked” and protected as a pair of sulfides to prevent them from breakdown under subsequent chemical reactions. The disulfide-containing groups were then regenerated after the dimerization reaction.

“This particular dimerization really stands out in terms of the complexity of the substrates that we’re bringing together, which have such a dense array of functional groups and stereochemistry,” Movassaghi says.

The overall synthesis requires 16 steps from the beta-hydroxytryptophan starting material to verticillin A.

Killing cancer cells

Once the researchers had successfully completed the synthesis, they were also able to tweak it to generate derivates of verticillin A. Researchers at Dana-Farber then tested these compounds against several types of diffuse midline glioma (DMG), a rare brain tumor that has few treatment options.

The researchers found that the DMG cell lines most susceptible to these compounds were those that have high levels of a protein called EZHIP. This protein, which plays a role in the methylation of DNA, has been previously identified as a potential drug target for DMG.

“Identifying the potential targets of these compounds will play a critical role in further understanding their mechanism of action, and more importantly, will help optimize the compounds from the Movassaghi lab to be more target specific for novel therapy development,” Qi says.

The verticillin derivatives appear to interact with EZHIP in a way that increases DNA methylation, which induces the cancer cells to under programmed cell death. The compounds that were most successful at killing these cells were N-sulfonylated (+)-11,11'-dideoxyverticillin A and N-sulfonylated verticillin A. N-sulfonylation — the addition of a functional group containing sulfur and oxygen — makes the molecules more stable.

“The natural product itself is not the most potent, but it’s the natural product synthesis that brought us to a point where we can make these derivatives and study them,” Movassaghi says.

The Dana-Farber team is now working on further validating the mechanism of action of the verticillin derivatives, and they also hope to begin testing the compounds in animal models of pediatric brain cancers.

“Natural compounds have been valuable resources for drug discovery, and we will fully evaluate the therapeutic potential of these molecules by integrating our expertise in chemistry, chemical biology, cancer biology, and patient care. We have also profiled our lead molecules in more than 800 cancer cell lines, and will be able to understand their functions more broadly in other cancers,” Qi says.

The research was funded by the National Institute of General Medical Sciences, the Ependymoma Research Foundation, and the Curing Kids Cancer Foundation.

More than 600 undergraduate students crowded into the Stratton Student Center on Oct. 28, for MIT’s first-ever Institute-wide Undergraduate Research Opportunities Program (UROP) mixer.

“At MIT, we believe in the transformative power of learning by doing, and there’s no better example than UROP,” says MIT President Sally Kornbluth, who attended the mixer with Provost Anantha Chandrakasan and Chancellor Melissa Nobles. “The energy at the inaugural UROP mixer was exhilarating, and I’m delighted that students now have this easy way to explore different paths to the frontiers of research.”

The event gave students the chance to explore internships and undergraduate research opportunities — in fields ranging from artificial intelligence to the life sciences to the arts, and beyond — all in one place, with approximately 150 researchers from labs available to discuss the projects and answer questions in real time. The offices of the Chancellor and Provost co-hosted the event, which the UROP office helped coordinate. 

First-year student Isabell Luo recently began a UROP project in the Living Matter lab led by Professor Rafael Gómez-Bombarelli, where she is benchmarking machine-learned interatomic potentials that simulate chemical reactions at the molecular level and exploring fine-tuning strategies to improve their accuracy. She’s passionate about AI and machine learning, eco-friendly design, and entrepreneurship, and was attending the UROP mixer to find more “real-world” projects to work on.

“I’m trying to dip my toes into different areas, which is why I’m at the mixer,” said Luo. “On the internet it would be so hard to find the right opportunities. It’s nice to have a physical space and speak to people from so many disciplines.”

More than nine out of every 10 members of MIT’s class of 2025 took part in a UROP before graduating. In recent years, approximately 3,200 undergraduates have participated in a UROP project each year. To meet the strong demand for UROPs, the Institute will commit up to $1 million in funding this year to create more of them. The funding will come from MIT’s schools and Office of the Provost. 

“UROPs have become an indispensable part of the MIT undergraduate education, providing hands-on experience that really helps students learn new ways to problem-solve and innovate,” says Chandrakasan. “I was thrilled to see so many students at the mixer — it was a testament to their willingness to roll up their sleeves and get to work on really tough challenges.”

Arielle Berman, a postdoc in the Raman Lab, was looking to recruit an undergraduate researcher for a project on sensor integration for muscle actuators for biohybrid robots — robots that include living parts. She spoke about how her own research experience as an undergraduate had shaped her career.

“It’s a really important event because we’re able to expose undergraduates to research,” says Berman. “I’m the first PhD in my family, so I wasn’t aware that research existed, or could be a career. Working in a research lab as an undergraduate student changed my life trajectory, and I’m happy to pass it forward and help students have experiences they wouldn’t have otherwise.”

The event drew students with interests as varied as the projects available. For first-year Nate Black, who plans to major in mechanical engineering, “I just wanted something to develop my interest in 3D printing and additive manufacturing.” First-year Akpandu Ekezie, who expects to major in Course 6-5 (Electrical Engineering with Computing), was interested in photonic circuits. “I’m looking mainly for EE-related things that are more hands-on,” he explained. “I want to get more physical experience.”

Nobles has a message for students considering a UROP project: Just go for it. “There’s a UROP for every student, regardless of experience,” she says. “Find something that excites you and give it a try.” She encourages students who weren’t able to attend the mixer, as well as those who did attend but still have questions, to get in touch with the UROP office.

First-year students Ruby Mykkanen and Aditi Deshpande attended the mixer together. Both were searching for UROP projects they could work on during Independent Activities Period in January. Deshpande also noted that the mixer was helpful for understanding “what research is being done at MIT.”

Said Mykkanen, “It’s fun to have it all in one place!”

Imagine having a continuum soft robotic arm bend around a bunch of grapes or broccoli, adjusting its grip in real time as it lifts the object. Unlike traditional rigid robots that generally aim to avoid contact with the environment as much as possible and stay far away from humans for safety reasons, this arm senses subtle forces, stretching and flexing in ways that mimic more of the compliance of a human hand. Its every motion is calculated to avoid excessive force while achieving the task efficiently. In MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and Laboratory for Information and Decisions Systems (LIDS) labs, these seemingly simple movements are the culmination of complex mathematics, careful engineering, and a vision for robots that can safely interact with humans and delicate objects.

Soft robots, with their deformable bodies, promise a future where machines move more seamlessly alongside people, assist in caregiving, or handle delicate items in industrial settings. Yet that very flexibility makes them difficult to control. Small bends or twists can produce unpredictable forces, raising the risk of damage or injury. This motivates the need for safe control strategies for soft robots. 

“Inspired by advances in safe control and formal methods for rigid robots, we aim to adapt these ideas to soft robotics — modeling their complex behavior and embracing, rather than avoiding, contact — to enable higher-performance designs (e.g., greater payload and precision) without sacrificing safety or embodied intelligence,” says lead senior author and MIT Assistant Professor Gioele Zardini, who is a principal investigator in LIDS and the Department of Civil and Environmental Engineering, and an affiliate faculty with the Institute for Data, Systems, and Society (IDSS). “This vision is shared by recent and parallel work from other groups.”

Safety first

The team developed a new framework that blends nonlinear control theory (controlling systems that involve highly complex dynamics) with advanced physical modeling techniques and efficient real-time optimization to produce what they call “contact-aware safety.” At the heart of the approach are high-order control barrier functions (HOCBFs) and high-order control Lyapunov functions (HOCLFs). HOCBFs define safe operating boundaries, ensuring the robot doesn’t exert unsafe forces. HOCLFs guide the robot efficiently toward its task objectives, balancing safety with performance.

“Essentially, we’re teaching the robot to know its own limits when interacting with the environment while still achieving its goals,” says MIT Department of Mechanical Engineering PhD student Kiwan Wong, the lead author of a new paper describing the framework. “The approach involves some complex derivation of soft robot dynamics, contact models, and control constraints, but the specification of control objectives and safety barriers is rather straightforward for the practitioner, and the outcomes are very tangible, as you see the robot moving smoothly, reacting to contact, and never causing unsafe situations.”

“Compared with traditional kinematic CBFs — where forward-invariant safe sets are hard to specify — the HOCBF framework simplifies barrier design, and its optimization formulation accounts for system dynamics (e.g., inertia), ensuring the soft robot stops early enough to avoid unsafe contact forces,” says Worcester Polytechnic Institute Assistant Professor and former CSAIL postdoc Wei Xiao.

“Since soft robots emerged, the field has highlighted their embodied intelligence and greater inherent safety relative to rigid robots, thanks to passive material and structural compliance. Yet their “cognitive” intelligence — especially safety systems — has lagged behind that of rigid serial-link manipulators,” says co-lead author Maximilian Stölzle, a research intern at Disney Research and formerly a Delft University of Technology PhD student and visiting researcher at MIT LIDS and CSAIL. “This work helps close that gap by adapting proven algorithms to soft robots and tailoring them for safe contact and soft-continuum dynamics.”

The LIDS and CSAIL team tested the system on a series of experiments designed to challenge the robot’s safety and adaptability. In one test, the arm pressed gently against a compliant surface, maintaining a precise force without overshooting. In another, it traced the contours of a curved object, adjusting its grip to avoid slippage. In yet another demonstration, the robot manipulated fragile items alongside a human operator, reacting in real time to unexpected nudges or shifts. “These experiments show that our framework is able to generalize to diverse tasks and objectives, and the robot can sense, adapt, and act in complex scenarios while always respecting clearly defined safety limits,” says Zardini.

Soft robots with contact-aware safety could be a real value-add in high-stakes places, of course. In health care, they could assist in surgeries, providing precise manipulation while reducing risk to patients. In industry, they might handle fragile goods without constant supervision. In domestic settings, robots could help with chores or caregiving tasks, interacting safely with children or the elderly — a key step toward making soft robots reliable partners in real-world environments. 

“Soft robots have incredible potential,” says co-lead senior author Daniela Rus, director of CSAIL and a professor in the Department of Electrical Engineering and Computer Science. “But ensuring safety and encoding motion tasks via relatively simple objectives has always been a central challenge. We wanted to create a system where the robot can remain flexible and responsive while mathematically guaranteeing it won’t exceed safe force limits.”

Combining soft robot models, differentiable simulation, and control theory

Underlying the control strategy is a differentiable implementation of something called the Piecewise Cosserat-Segment (PCS) dynamics model, which predicts how a soft robot deforms and where forces accumulate. This model allows the system to anticipate how the robot’s body will respond to actuation and complex interactions with the environment. “The aspect that I most like about this work is the blend of integration of new and old tools coming from different fields like advanced soft robot models, differentiable simulation, Lyapunov theory, convex optimization, and injury-severity–based safety constraints. All of this is nicely blended into a real-time controller fully grounded in first principles,” says co-author Cosimo Della Santina, who is an associate professor at Delft University of Technology. 

Complementing this is the Differentiable Conservative Separating Axis Theorem (DCSAT), which estimates distances between the soft robot and obstacles in the environment that can be approximated with a chain of convex polygons in a differentiable manner. “Earlier differentiable distance metrics for convex polygons either couldn’t compute penetration depth — essential for estimating contact forces — or yielded non-conservative estimates that could compromise safety,” says Wong. “Instead, the DCSAT metric returns strictly conservative, and therefore safe, estimates while simultaneously allowing for fast and differentiable computation.” Together, PCS and DCSAT give the robot a predictive sense of its environment for more proactive, safe interactions.

Looking ahead, the team plans to extend their methods to three-dimensional soft robots and explore integration with learning-based strategies. By combining contact-aware safety with adaptive learning, soft robots could handle even more complex, unpredictable environments. 

“This is what makes our work exciting,” says Rus. “You can see the robot behaving in a human-like, careful manner, but behind that grace is a rigorous control framework ensuring it never oversteps its bounds.”

“Soft robots are generally safer to interact with than rigid-bodied robots by design, due to the compliance and energy-absorbing properties of their bodies,” says University of Michigan Assistant Professor Daniel Bruder, who wasn’t involved in the research. “However, as soft robots become faster, stronger, and more capable, that may no longer be enough to ensure safety. This work takes a crucial step towards ensuring soft robots can operate safely by offering a method to limit contact forces across their entire bodies.”

The team’s work was supported, in part, by The Hong Kong Jockey Club Scholarships, the European Union’s Horizon Europe Program, Cultuurfonds Wetenschapsbeurzen, and the Rudge (1948) and Nancy Allen Chair. Their work was published earlier this month in the Institute of Electrical and Electronics Engineers’ Robotics and Automation Letters.

Researchers at MIT have demonstrated that wedge-shaped vortex generators attached to a ship’s hull can reduce drag by up to 7.5 percent, which reduces overall ship emissions and fuel expenses. The paper, “Net Drag Reduction in High Block Coefficient Ships and Vehicles Using Vortex Generators,” was presented at the Society of Naval Architects and Marine Engineers 2025 Maritime Convention in Norfolk, Virginia.

The work offers a promising path toward decarbonization, addressing the pressing need to meet the International Maritime Organization (IMO) goal to reduce carbon intensity of international shipping by at least 40 percent by 2030, compared to 2008 levels. Achieving such ambitious emissions reduction will require a coordinated approach, employing multiple methods, from redesigning ship hulls, propellers, and engines to using novel fuels and operational methods.

The researchers — José del Águila Ferrandis, Jack Kimmeth, and Michael Triantafyllou of MIT Sea Grant and the Department of Mechanical Engineering, along with Alfonso Parra Rubio and Neil Gershenfeld of the Center for Bits and Atoms — determined the optimized vortex generator shape and size using a combination of computational fluid dynamics (CFD) and experimental methods guided by AI optimization methods. 

The team first established parametric trends through extensive CFD analysis, and then tested multiple hulls through rapid prototyping to validate the results experimentally. Scale models of an axisymmetric hull with a bare tail, a tail with delta wing vortex generators, and a tail with wedge vortex generators were produced and tested. The team identified wedge-like vortex generators as the key shape that could achieve this level of drag reduction. 

Through flow visualization, the researchers could see that drag was reduced by delaying turbulent flow separation, helping water flow more smoothly along the ship’s hull, shrinking the wake behind the vessel. This also allows the propeller and rudder to work more efficiently in a uniform flow. “We document for the first time experimentally a reduction in fuel required by ships using vortex generators, relatively small structures in the shape of a wedge attached at a specific point of the ship’s hull,” explains Michael Triantafyllou, professor of mechanical engineering and director of MIT Sea Grant. 

Vortex generators have long been used in aircraft-wing design to maintain lift and delay stalling. This study is the first to show that the vortex generators can be applied for drag reduction in commercial ships.

The modular adaptability of the wedge vortex generators would allow integration into a broad range of hull forms, including bulk carriers and tankers, and the devices can synergize with, or even replace, existing technologies like pre-swirl stators (fixed fins mounted in front of propellers), improving overall system performance. As an example case, the researchers estimate that installing the vortex generators on a 300-meter Newcastlemax bulk carrier operating at 14.5 knots over a cross-Pacific route would result in significantly reduced emissions and approximately $750,000 in fuel savings per year.

The findings offer a practical, cost-effective solution that could be implemented efficiently across existing fleets. This study was supported through the CBA Consortium, working with Oldendorff Carriers, which operates about 700 bulk carriers around the world. An extension of this research is supported by the MIT Maritime Consortium, led by MIT professors Themis Sapsis and Fotini Christia. The Maritime Consortium was formed in 2025 to address critical gaps in the modernization of the commercial fleet through interdisciplinary research and collaboration across academia, industry, and regulatory agencies.

Norway is the world’s largest producer of farmed Atlantic salmon and a top exporter of seafood, while the United States remains the largest importer of these products, according to the Food and Agriculture Organization. Two MIT students recently traveled to Trondheim, Norway to explore the cutting-edge technologies being developed and deployed in offshore aquaculture. 

Beckett Devoe, a senior in artificial intelligence and decision-making, and Tony Tang, a junior in mechanical engineering, first worked with MIT Sea Grant through the Undergraduate Research Opportunities Program (UROP). They contributed to projects focusing on wave generator design and machine learning applications for analyzing oyster larvae health in hatcheries. While near-shore aquaculture is a well-established industry across Massachusetts and the United States, open-ocean farming is still a nascent field here, facing unique and complex challenges. 

To help better understand this emerging industry, MIT Sea Grant created a collaborative initiative, AquaCulture Shock, with funding from an Aquaculture Technologies and Education Travel Grant through the National Sea Grant College Program. Collaborating with the MIT-Scandinavia MISTI (MIT International Science and Technology Initiatives) program, MIT Sea Grant matched Devoe and Tang with aquaculture-related summer internships at SINTEF Ocean, one of the largest research institutes in Europe. 

“The opportunity to work on this hands-on aquaculture project, under a world-renowned research institution, in an area of the world known for its innovation in marine technology — this is what MISTI is all about,” says Madeline Smith, managing director for MIT-Scandinavia. “Not only are students gaining valuable experience in their fields of study, but they’re developing cultural understanding and skills that equip them to be future global leaders.” Both students worked within SINTEF Ocean’s Aquaculture Robotics and Autonomous Systems Laboratory (ACE-Robotic Lab), a facility designed to develop and test new aquaculture technologies. 

“Norway has this unique geography where it has all of these fjords,” says Sveinung Ohrem, research manager for the Aquaculture Robotics and Automation Group at SINTEF Ocean. “So you have a lot of sheltered waters, which makes it ideal to do sea-based aquaculture.” He estimates that there are about a thousand fish farms along Norway’s coast, and walks through some of the tools being used in the industry: decision-making systems to gather and visualize data for the farmers and operators; robots for inspection and cleaning; environmental sensors to measure oxygen, temperature, and currents; echosounders that send out acoustic signals to track where the fish are; and cameras to help estimate biomass and fine-tune feeding. “Feeding is a huge challenge,” he notes. “Feed is the largest cost, by far, so optimizing feeding leads to a very significant decrease in your cost.”

During the internship, Devoe focused on a project that uses AI for fish feeding optimization. “I try to look at the different features of the farm — so maybe how big the fish are, or how cold the water is ... and use that to try to give the farmers an optimal feeding amount for the best outcomes, while also saving money on feed,” he explains. “It was good to learn some more machine learning techniques and just get better at that on a real-world project.” 

In the same lab, Tang worked on the simulation of an underwater vehicle-manipulator system to navigate farms and repair damage on cage nets with a robotic arm. Ohrem says there are thousands of aquaculture robots operating in Norway today. “The scale is huge,” he says. “You can’t have 8,000 people controlling 8,000 robots — that’s not economically or practically feasible. So the level of autonomy in all of these robots needs to be increased.”

The collaboration between MIT and SINTEF Ocean began in 2023 when MIT Sea Grant hosted Eleni Kelasidi, a visiting research scientist from the ACE-Robotic Lab. Kelasidi collaborated with MIT Sea Grant director Michael Triantafyllou and professor of mechanical engineering Themistoklis Sapsis developing controllers, models, and underwater vehicles for aquaculture, while also investigating fish-machine interactions. 

“We have had a long and fruitful collaboration with the Norwegian University of Science and Technology (NTNU) and SINTEF, which continues with important efforts such as the aquaculture project with Dr. Kelasidi,” Triantafyllou says. “Norway is at the forefront of offshore aquaculture and MIT Sea Grant is investing in this field, so we anticipate great results from the collaboration.”

Kelasidi, who is now a professor at NTNU, also leads the Field Robotics Lab, focusing on developing resilient robotic systems to operate in very complex and harsh environments. “Aquaculture is one of the most challenging field domains we can demonstrate any autonomous solutions, because everything is moving,” she says. Kelasidi describes aquaculture as a deeply interdisciplinary field, requiring more students with backgrounds both in biology and technology. “We cannot develop technologies that are applied for industries where we don’t have biological components,” she explains, “and then apply them somewhere where we have a live fish or other live organisms.” 

Ohrem affirms that maintaining fish welfare is the primary driver for researchers and companies operating in aquaculture, especially as the industry continues to grow. “So the big question is,” he says, “how can you ensure that?” SINTEF Ocean has four research licenses for farming fish, which they operate through a collaboration with SalMar, the second-largest salmon farmer in the world. The students had the opportunity to visit one of the industrial-scale farms, Singsholmen, on the island of Hitra. The farm has 10 large, round net pens about 50 meters across that extend deep below the surface, each holding up to 200,000 salmon. “I got to physically touch the nets and see how the [robotic] arm might be able to fix the net,” says Tang. 

Kelasidi emphasizes that the information gained in the field cannot be learned from the office or lab. “That opens up and makes you realize, what is the scale of the challenges, or the scale of the facilities,” she says. She also highlights the importance of international and institutional collaboration to advance this field of research and develop more resilient robotic systems. “We need to try to target that problem, and let’s solve it together.”

MIT Sea Grant and the MIT-Scandinavia MISTI program are currently recruiting a new cohort of four MIT students to intern in Norway this summer with institutes advancing offshore farming technologies, including NTNU’s Field Robotics Lab in Trondheim. Students interested in autonomy, deep learning, simulation modeling, underwater robotic systems, and other aquaculture-related areas are encouraged to reach out to Lily Keyes at MIT Sea Grant.

Advancements in battery innovation are transforming both mobility and energy systems alike, according to Kurt Kelty, vice president of battery, propulsion, and sustainability at General Motors (GM). At the MIT Energy Initiative (MITEI) Fall Colloquium, Kelty explored how GM is bringing next-generation battery technologies from lab to commercialization, driving American battery innovation forward. The colloquium is part of the ongoing MITEI Presents: Advancing the Energy Transition speaker series.

At GM, Kelty’s team is primarily focused on three things: first, improving affordability to get more electric vehicles (EVs) on the road. “How do you drive down the cost?” Kelty asked the audience. “It's the batteries. The batteries make up about 30 percent of the cost of the vehicle.” Second, his team strives to improve battery performance, including charging speed and energy density. Third, they are working on localizing the supply chain. “We've got to build up our resilience and our independence here in North America, so we're not relying on materials coming from China,” Kelty explained.

To aid their efforts, resources are being poured into the virtualization space, significantly cutting down on time dedicated to research and development. Now, Kelty’s team can do modeling up front using artificial intelligence, reducing what previously would have taken months to a couple of days.

“If you want to modify … the nickel content ever so slightly, we can very quickly model: ‘OK, how’s that going to affect the energy density? The safety? How’s that going to affect the charge capability?’” said Kelty. “We can look at that at the cell level, then the pack level, then the vehicle level.”

Kelty revealed that they have found a solution that addresses affordability, accessibility, and commercialization: lithium manganese-rich (LMR) batteries. Previously, the industry looked to reduce costs by lowering the amount of cobalt in batteries by adding greater amounts of nickel. These high-nickel batteries are in most cars on the road in the United States due to their high range. LMR batteries, though, take things a step further by reducing the amount of nickel and adding more manganese, which drives the cost of batteries down even further while maintaining range.

Lithium-iron-phosphate (LFP) batteries are the chemistry of choice in China, known for low cost, high cycle life, and high safety. With LMR batteries, the cost is comparable to LFP with a range that is closer to high-nickel. “That’s what’s really a breakthrough,” said Kelty.

LMR batteries are not new, but there have been challenges to adopting them, according to Kelty. “People knew about it, but they didn’t know how to commercialize it. They didn’t know how to make it work in an EV,” he explained. Now that GM has figured out commercialization, they will be the first to market these batteries in their EVs in 2028.

Kelty also expressed excitement over the use of vehicle-to-grid technologies in the future. Using a bidirectional charger with a two-way flow of energy, EVs could charge, but also send power from their batteries back to the electrical grid. This would allow customers to charge “their vehicles at night when the electricity prices are really low, and they can discharge it during the day when electricity rates are really high,” he said.

In addition to working in the transportation sector, GM is exploring ways to extend their battery expertise into applications in grid-scale energy storage. “It’s a big market right now, but it’s growing very quickly because of the data center growth,” said Kelty.

When looking to the future of battery manufacturing and EVs in the United States, Kelty remains optimistic: “we’ve got the technology here to make it happen. We’ve always had the innovation here. Now, we’re getting more and more of the manufacturing. We’re getting that all together. We’ve got just tremendous opportunity here that I’m hopeful we’re going to be able to take advantage of and really build a massive battery industry here.”

This speaker series highlights energy experts and leaders at the forefront of the scientific, technological, and policy solutions needed to transform our energy systems. Visit MITEI’s Events page for more information on this and additional events.

“MIT hasn’t just prepared me for the future of work — it’s pushed me to study it. As AI systems become more capable, more of our online activity will be carried out by artificial agents. That raises big questions: How should we design these systems to understand our preferences? What happens when AI begins making many of our decisions?”

These are some of the questions MIT Sloan School of Management PhD candidate Benjamin Manning is researching. Part of his work investigates how to design and evaluate artificial intelligence agents that act on behalf of people, and how their behavior shapes markets and institutions. 

Previously, he received a master’s degree in public policy from the Harvard Kennedy School and a bachelor’s in mathematics from Washington University in St. Louis. After working as a research assistant, Manning knew he wanted to pursue an academic career.

“There’s no better place in the world to study economics and computer science than MIT,” he says. “Nobel and Turing award winners are everywhere, and the IT group lets me explore both fields freely. It was my top choice — when I was accepted, the decision was clear.” 

After receiving his PhD, Manning hopes to secure a faculty position at a business school and do the same type of work that MIT Sloan professors — his mentors — do every day.

“Even in my fourth year, it still feels surreal to be an MIT student. I don’t think that feeling will ever fade. My mom definitely won’t ever get over telling people about it.”

Of his MIT Sloan experience, Manning says he didn’t know it was possible to learn so much so quickly. “It’s no exaggeration to say I learned more in my first year as a PhD candidate than in all four years of undergrad. While the pace can be intense, wrestling with so many new ideas has been incredibly rewarding. It’s given me the tools to do novel research in economics and AI — something I never imagined I’d be capable of.”

As an economist studying AI simulations of humans, for Manning, the future of work not only means understanding how AI acts on our behalf, but also radically improving and accelerating social scientific discovery.

“Another part of my research agenda explores how well AI systems can simulate human responses. I envision a future where researchers test millions of behavioral simulations in minutes, rapidly prototyping experimental designs, and identifying promising research directions before investing in costly human studies. This isn’t about replacing human insight, but amplifying it: Scientists can focus on asking better questions, developing theory, and interpreting results while AI handles the computational heavy lifting.”

He’s excited by the prospect: “We are possibly moving toward a world where the pace of understanding may get much closer to the speed of economic change.”

Our muscles are nature’s actuators. The sinewy tissue is what generates the forces that make our bodies move. In recent years, engineers have used real muscle tissue to actuate “biohybrid robots” made from both living tissue and synthetic parts. By pairing lab-grown muscles with synthetic skeletons, researchers are engineering a menagerie of muscle-powered crawlers, walkers, swimmers, and grippers.

But for the most part, these designs are limited in the amount of motion and power they can produce. Now, MIT engineers are aiming to give bio-bots a power lift with artificial tendons.

In a study appearing today in the journal Advanced Science, the researchers developed artificial tendons made from tough and flexible hydrogel. They attached the rubber band-like tendons to either end of a small piece of lab-grown muscle, forming a “muscle-tendon unit.” Then they connected the ends of each artificial tendon to the fingers of a robotic gripper.

When they stimulated the central muscle to contract, the tendons pulled the gripper’s fingers together. The robot pinched its fingers together three times faster, and with 30 times greater force, compared with the same design without the connecting tendons.

The researchers envision the new muscle-tendon unit can be fit to a wide range of biohybrid robot designs, much like a universal engineering element.

“We are introducing artificial tendons as interchangeable connectors between muscle actuators and robotic skeletons,” says lead author Ritu Raman, an assistant professor of mechanical engineering (MechE) at MIT. “Such modularity could make it easier to design a wide range of robotic applications, from microscale surgical tools to adaptive, autonomous exploratory machines.”

The study’s MIT co-authors include graduate students Nicolas Castro, Maheera Bawa, Bastien Aymon, Sonika Kohli, and Angel Bu; undergraduate Annika Marschner; postdoc Ronald Heisser; alumni Sarah J. Wu ’19, SM ’21, PhD ’24 and Laura Rosado ’22, SM ’25; and MechE professors Martin Culpepper and Xuanhe Zhao.

Muscle’s gains

Raman and her colleagues at MIT are at the forefront of biohybrid robotics, a relatively new field that has emerged in the last decade. They focus on combining synthetic, structural robotic parts with living muscle tissue as natural actuators.

“Most actuators that engineers typically work with are really hard to make small,” Raman says. “Past a certain size, the basic physics doesn’t work. The nice thing about muscle is, each cell is an independent actuator that generates force and produces motion. So you could, in principle, make robots that are really small.”

Muscle actuators also come with other advantages, which Raman’s team has already demonstrated: The tissue can grow stronger as it works out, and can naturally heal when injured. For these reasons, Raman and others envision that muscly droids could one day be sent out to explore environments that are too remote or dangerous for humans. Such muscle-bound bots could build up their strength for unforeseen traverses or heal themselves when help is unavailable. Biohybrid bots could also serve as small, surgical assistants that perform delicate, microscale procedures inside the body.

All these future scenarios are motivating Raman and others to find ways to pair living muscles with synthetic skeletons. Designs to date have involved growing a band of muscle and attaching either end to a synthetic skeleton, similar to looping a rubber band around two posts. When the muscle is stimulated to contract, it can pull the parts of a skeleton together to generate a desired motion.

But Raman says this method produces a lot of wasted muscle that is used to attach the tissue to the skeleton rather than to make it move. And that connection isn’t always secure. Muscle is quite soft compared with skeletal structures, and the difference can cause muscle to tear or detach. What’s more, it is often only the contractions in the central part of the muscle that end up doing any work — an amount that’s relatively small and generates little force.

“We thought, how do we stop wasting muscle material, make it more modular so it can attach to anything, and make it work more efficiently?” Raman says. “The solution the body has come up with is to have tendons that are halfway in stiffness between muscle and bone, that allow you to bridge this mechanical mismatch between soft muscle and rigid skeleton. They’re like thin cables that wrap around joints efficiently.”

“Smartly connected”

In their new work, Raman and her colleagues designed artificial tendons to connect natural muscle tissue with a synthetic gripper skeleton. Their material of choice was hydrogel — a squishy yet sturdy polymer-based gel. Raman obtained hydrogel samples from her colleague and co-author Xuanhe Zhao, who has pioneered the development of hydrogels at MIT. Zhao’s group has derived recipes for hydrogels of varying toughness and stretch that can stick to many surfaces, including synthetic and biological materials.

To figure out how tough and stretchy artificial tendons should be in order to work in their gripper design, Raman’s team first modeled the design as a simple system of three types of springs, each representing the central muscle, the two connecting tendons, and the gripper skeleton. They assigned a certain stiffness to the muscle and skeleton, which were previously known, and used this to calculate the stiffness of the connecting tendons that would be required in order to move the gripper by a desired amount.

From this modeling, the team derived a recipe for hydrogel of a certain stiffness. Once the gel was made, the researchers carefully etched the gel into thin cables to form artificial tendons. They attached two tendons to either end of a small sample of muscle tissue, which they grew using lab-standard techniques. They then wrapped each tendon around a small post at the end of each finger of the robotic gripper — a skeleton design that was developed by MechE professor Martin Culpepper, an expert in designing and building precision machines.

When the team stimulated the muscle to contract, the tendons in turn pulled on the gripper to pinch its fingers together. Over multiple experiments, the researchers found that the muscle-tendon gripper worked three times faster and produced 30 times more force compared to when the gripper is actuated just with a band of muscle tissue (and without any artificial tendons). The new tendon-based design also was able to keep up this performance over 7,000 cycles, or muscle contractions.

Overall, Raman saw that the addition of artificial tendons increased the robot’s power-to-weight ratio by 11 times, meaning that the system required far less muscle to do just as much work.

“You just need a small piece of actuator that’s smartly connected to the skeleton,” Raman says. “Normally, if a muscle is really soft and attached to something with high resistance, it will just tear itself before moving anything. But if you attach it to something like a tendon that can resist tearing, it can really transmit its force through the tendon, and it can move a skeleton that it wouldn’t have been able to move otherwise.”

The team’s new muscle-tendon design successfully merges biology with robotics, says biomedical engineer Simone Schürle-Finke, associate professor of health sciences and technology at ETH Zürich.

“The tough-hydrogel tendons create a more physiological muscle–tendon–bone architecture, which greatly improves force transmission, durability, and modularity,” says Schürle-Finke, who was not involved with the study. “This moves the field toward biohybrid systems that can operate repeatably and eventually function outside the lab.”

With the new artificial tendons in place, Raman’s group is moving forward to develop other elements, such as skin-like protective casings, to enable muscle-powered robots in practical, real-world settings.

This research was supported, in part, by the U.S. Department of Defense Army Research Office, the MIT Research Support Committee, and the National Science Foundation.

Large language models (LLMs) sometimes learn the wrong lessons, according to an MIT study.

Rather than answering a query based on domain knowledge, an LLM could respond by leveraging grammatical patterns it learned during training. This can cause a model to fail unexpectedly when deployed on new tasks.

The researchers found that models can mistakenly link certain sentence patterns to specific topics, so an LLM might give a convincing answer by recognizing familiar phrasing instead of understanding the question.

Their experiments showed that even the most powerful LLMs can make this mistake.

This shortcoming could reduce the reliability of LLMs that perform tasks like handling customer inquiries, summarizing clinical notes, and generating financial reports.

It could also have safety risks. A nefarious actor could exploit this to trick LLMs into producing harmful content, even when the models have safeguards to prevent such responses.

After identifying this phenomenon and exploring its implications, the researchers developed a benchmarking procedure to evaluate a model’s reliance on these incorrect correlations. The procedure could help developers mitigate the problem before deploying LLMs.

“This is a byproduct of how we train models, but models are now used in practice in safety-critical domains far beyond the tasks that created these syntactic failure modes. If you’re not familiar with model training as an end-user, this is likely to be unexpected,” says Marzyeh Ghassemi, an associate professor in the MIT Department of Electrical Engineering and Computer Science (EECS), a member of the MIT Institute of Medical Engineering Sciences and the Laboratory for Information and Decision Systems, and the senior author of the study.

Ghassemi is joined by co-lead authors Chantal Shaib, a graduate student at Northeastern University and visiting student at MIT; and Vinith Suriyakumar, an MIT graduate student; as well as Levent Sagun, a research scientist at Meta; and Byron Wallace, the Sy and Laurie Sternberg Interdisciplinary Associate Professor and associate dean of research at Northeastern University’s Khoury College of Computer Sciences. A paper describing the work will be presented at the Conference on Neural Information Processing Systems.

Stuck on syntax

LLMs are trained on a massive amount of text from the internet. During this training process, the model learns to understand the relationships between words and phrases — knowledge it uses later when responding to queries.

In prior work, the researchers found that LLMs pick up patterns in the parts of speech that frequently appear together in training data. They call these part-of-speech patterns “syntactic templates.”

LLMs need this understanding of syntax, along with semantic knowledge, to answer questions in a particular domain.

“In the news domain, for instance, there is a particular style of writing. So, not only is the model learning the semantics, it is also learning the underlying structure of how sentences should be put together to follow a specific style for that domain,” Shaib explains.   

But in this research, they determined that LLMs learn to associate these syntactic templates with specific domains. The model may incorrectly rely solely on this learned association when answering questions, rather than on an understanding of the query and subject matter.

For instance, an LLM might learn that a question like “Where is Paris located?” is structured as adverb/verb/proper noun/verb. If there are many examples of sentence construction in the model’s training data, the LLM may associate that syntactic template with questions about countries.

So, if the model is given a new question with the same grammatical structure but nonsense words, like “Quickly sit Paris clouded?” it might answer “France” even though that answer makes no sense.

“This is an overlooked type of association that the model learns in order to answer questions correctly. We should be paying closer attention to not only the semantics but the syntax of the data we use to train our models,” Shaib says.

Missing the meaning

The researchers tested this phenomenon by designing synthetic experiments in which only one syntactic template appeared in the model’s training data for each domain. They tested the models by substituting words with synonyms, antonyms, or random words, but kept the underlying syntax the same.

In each instance, they found that LLMs often still responded with the correct answer, even when the question was complete nonsense.

When they restructured the same question using a new part-of-speech pattern, the LLMs often failed to give the correct response, even though the underlying meaning of the question remained the same.

They used this approach to test pre-trained LLMs like GPT-4 and Llama, and found that this same learned behavior significantly lowered their performance.

Curious about the broader implications of these findings, the researchers studied whether someone could exploit this phenomenon to elicit harmful responses from an LLM that has been deliberately trained to refuse such requests.

They found that, by phrasing the question using a syntactic template the model associates with a “safe” dataset (one that doesn’t contain harmful information), they could trick the model into overriding its refusal policy and generating harmful content.

“From this work, it is clear to me that we need more robust defenses to address security vulnerabilities in LLMs. In this paper, we identified a new vulnerability that arises due to the way LLMs learn. So, we need to figure out new defenses based on how LLMs learn language, rather than just ad hoc solutions to different vulnerabilities,” Suriyakumar says.

While the researchers didn’t explore mitigation strategies in this work, they developed an automatic benchmarking technique one could use to evaluate an LLM’s reliance on this incorrect syntax-domain correlation. This new test could help developers proactively address this shortcoming in their models, reducing safety risks and improving performance.

In the future, the researchers want to study potential mitigation strategies, which could involve augmenting training data to provide a wider variety of syntactic templates. They are also interested in exploring this phenomenon in reasoning models, special types of LLMs designed to tackle multi-step tasks.

“I think this is a really creative angle to study failure modes of LLMs. This work highlights the importance of linguistic knowledge and analysis in LLM safety research, an aspect that hasn’t been at the center stage but clearly should be,” says Jessy Li, an associate professor at the University of Texas at Austin, who was not involved with this work.

This work is funded, in part, by a Bridgewater AIA Labs Fellowship, the National Science Foundation, the Gordon and Betty Moore Foundation, a Google Research Award, and Schmidt Sciences.

More than 300 people across academia and industry spilled into an auditorium to attend a BoltzGen seminar on Thursday, Oct. 30, hosted by the Abdul Latif Jameel Clinic for Machine Learning in Health (MIT Jameel Clinic). Headlining the event was MIT PhD student and BoltzGen’s first author Hannes Stärk, who had announced BoltzGen just a few days prior.

Building upon Boltz-2, an open-source biomolecular structure prediction model predicting protein binding affinity that made waves over the summer, BoltzGen (officially released on Sunday, Oct. 26.) is the first model of its kind to go a step further by generating novel protein binders that are ready to enter the drug discovery pipeline.

Three key innovations make this possible: first, BoltzGen’s ability to carry out a variety of tasks, unifying protein design and structure prediction while maintaining state-of-the-art performance. Next, BoltzGen’s built-in constraints are designed with feedback from wetlab collaborators to ensure the model creates functional proteins that don’t defy the laws of physics or chemistry. Lastly, a rigorous evaluation process tests the model on “undruggable” disease targets, pushing the limits of BoltzGen’s binder generation capabilities.

Most models used in industry or academia are capable of either structure prediction or protein design. Moreover, they’re limited to generating certain types of proteins that bind successfully to easy “targets.” Much like students responding to a test question that looks like their homework, as long as the training data looks similar to the target during binder design, the models often work. But existing methods are nearly always evaluated on targets for which structures with binders already exist, and end up faltering in performance when used on more challenging targets.

“There have been models trying to tackle binder design, but the problem is that these models are modality-specific,” Stärk points out. “A general model does not only mean that we can address more tasks. Additionally, we obtain a better model for the individual task since emulating physics is learned by example, and with a more general training scheme, we provide more such examples containing generalizable physical patterns.”

The BoltzGen researchers went out of their way to test BoltzGen on 26 targets, ranging from therapeutically relevant cases to ones explicitly chosen for their dissimilarity to the training data. 

This comprehensive validation process, which took place in eight wetlabs across academia and industry, demonstrates the model’s breadth and potential for breakthrough drug development.

Parabilis Medicines, one of the industry collaborators that tested BoltzGen in a wetlab setting, praised BoltzGen’s potential: “we feel that adopting BoltzGen into our existing Helicon peptide computational platform capabilities promises to accelerate our progress to deliver transformational drugs against major human diseases.”

While the open-source releases of Boltz-1, Boltz-2, and now BoltzGen (which was previewed at the 7th Molecular Machine Learning Conference on Oct. 22) bring new opportunities and transparency in drug development, they also signal that biotech and pharmaceutical industries may need to reevaluate their offerings. 

Amid the buzz for BoltzGen on the social media platform X, Justin Grace, a principal machine learning scientist at LabGenius, raised a question. “The private-to-open performance time lag for chat AI systems is [seven] months and falling,” Grace wrote in a post. “It looks to be even shorter in the protein space. How will binder-as-a-service co’s be able to [recoup] investment when we can just wait a few months for the free version?” 

For those in academia, BoltzGen represents an expansion and acceleration of scientific possibility. “A question that my students often ask me is, ‘where can AI change the therapeutics game?’” says senior co-author and MIT Professor Regina Barzilay, AI faculty lead for the Jameel Clinic and an affiliate of the Computer Science and Artificial Intelligence Laboratory (CSAIL). “Unless we identify undruggable targets and propose a solution, we won’t be changing the game,” she adds. “The emphasis here is on unsolved problems, which distinguishes Hannes’ work from others in the field.” 

Senior co-author Tommi Jaakkola, the Thomas Siebel Professor of Electrical Engineering and Computer Science who is affiliated with the Jameel Clinic and CSAIL, notes that "models such as BoltzGen that are released fully open-source enable broader community-wide efforts to accelerate drug design capabilities.”

Looking ahead, Stärk believes that the future of biomolecular design will be upended by AI models. “I want to build tools that help us manipulate biology to solve disease, or perform tasks with molecular machines that we have not even imagined yet,” he says. “I want to provide these tools and enable biologists to imagine things that they have not even thought of before.”