Teaching theory of mind to robots to enhance collaboration

 BE CAREFUL OF WHAT YOU WISH FOR

Duke roboticists present HUMAC, a new framework that enables robots to collaborate like humans with just 40 minutes of simple coaching

Duke University

LinkedInWeChatBlueskyMessageWhatsAppEmail

Nature is brimming with animals that collaborate in large numbers. Bees stake out the best feeding spots and let others know where they are. Ants construct complex hierarchical homes built for defense. Flocks of starlings move across the sky in beautiful formations as if they were a single entity.

None of these animals, however, collaborate in the way that humans do. Hive mind behaviors arise from simple rules followed by many individuals. Humans, however, have the ability to empathize with one another and predict each other’s actions—a trait known as Theory of Mind.

Now, a group of researchers from Duke University and Columbia University have figured out how to use this uniquely human trait to quickly train groups of robots to complete complex tasks. Whereas other control algorithms direct robots through mechanisms more similar to hive mind behaviors, this newly demonstrated framework called HUMAC teaches groups of robots how to collaborate through insights provided by a single human coach.

The research, accepted to the IEEE International Conference on Robotics and Automation (ICRA 2025), which will be held May 19-23, 2025, in Atlanta, Georgia, demonstrates how robots can learn to anticipate teammates’ actions, adapt strategies in real time and solve challenges that require human-like coordinated, collective intelligence.

The work could be a boon to applications such as wildfire response and wild survival tasks where robots need to cooperate and collaborate under constraints, with hierarchical team structures, uncertainty of the environment and communication bandwidth limits.

“Humans start to develop the skill of Theory of Mind around age four,” explained Boyuan Chen, the Dickinson Family Assistant Professor of Mechanical Engineering and Materials Science, Electrical and Computer Engineering, and Computer Science at Duke University. “It allows us to interpret and predict others’ intentions, allowing collaboration to emerge. This is an essential capability that our current robots are missing to allow them to work as a team with other robots and humans. We designed HUMAC to help robots learn from how humans think and coordinate in an efficient way.”

There have been other approaches to teaching robots to collaborate in meaningful tasks. One is to use reinforcement learning, where robots learn by interacting in the same environment with millions to billions of trials and errors, which is inefficient with no guarantee of success. Another method involves imitation learning from large groups of collaborative human experts, which is costly and impractical.

HUMAC takes a radically different approach. During training, the framework allows a single human operator to briefly take control of different robots within a team when necessary, guiding them at key strategic moments, much like a coach giving targeted advice during a soccer game. These interactions show the groups how to conduct sophisticated collaborative tactics like ambushing and encircling.

Following these brief demonstrations, the system embeds the human interventions into the robots’ algorithms. The key idea is that, for the robots to be able to learn to collaborate, they must learn to form a mental representation to simultaneously predict what their teammates’ plans are and what their opponent players will do. In other words, implicitly embedding all players’ decisions into their own plans—Theory of Mind.

“Our framework imagines the future of human-AI teaming where humans are leaders,” Chen said. “In this case, one human is guiding a larger number of agents in a fast and adaptable way, which has not been done before.”

The team tested HUMAC in a dynamic hide and seek game, where a team of three seeker robots try to catch a team of three faster-moving hider robots within a bounded square-shaped arena filled with random obstacles, relying solely on partial visual observations. This setting is challenging as non-collaborative seekers who keep chasing the closest hiders can only achieve a 36% success rate.

With HUMAC, a human coach selectively takes control of individual robots when necessary. After just 40 minutes of guidance, the robot team exhibits strong collaborative behaviors such as ambushing and encircling. In simulations, the success rate jumped to 84%, and even in physical ground vehicle tests, the success rate held strong at 80%.

“We observed robots starting to behave like genuine teammates,” said Zhengran Ji, the lead student author and graduate student in Chen’s lab. “They predicted each other’s movements and coordinated naturally, without explicit commands.”

“It was truly exciting to watch, and we believe it opens up many opportunities for future collaborative robot teams and human-robot teams in various applications,” Chen added.

Imagine a swarm of drones coordinating in real time to locate survivors after a natural disaster, efficiently sweeping through debris-covered areas without overlapping paths. Any application where a small number of humans need to teach a large number of robots how to collaborate could use this approach. Researchers are already working on expanding HUMAC to larger robot teams and more complex tasks while exploring richer interaction methods to streamline and enhance human-robot teaming.

“AI is not just a tool for humans, it’s a teammate. The final form of super-intelligence will not be AI alone nor humans alone, it’s the collective intelligence from both humans and AI,” Chen said. “Just as humans evolved to collaborate, AI will become more adaptive to work alongside with each other and with us. HUMAC is a step toward that future.”

This work is supported in part by ARL STRONG program under awards W911NF2320182 and W911NF2220113.

“Enabling Multi-Robot Collaboration from Single-Human Guidance.” Zhengran Ji, Lingyu Zhang, Paul Sajda, Boyuan Chen. IEEE International Conference on Robotics and Automation (ICRA 2025).

Project Website: HUMAC: Enabling Multi-Robot Collaboration from Single-Human Guidance - Research Blog

General Robotics Lab Website: http://generalroboticslab.com

---

Method of Research

Experimental study

Subject of Research

Not applicable

Article Title

Enabling Multi-Robot Collaboration from Single-Human Guidance

Article Publication Date

19-May-2025

Team develops digital lab for data- and robot-driven materials science

The dLab fully automates processes from material synthesis to analysis

University of Tokyo

Facebook

X
BlueskyMessageWhatsAppEmail

 

image: 

Professor Taro Hitosugi stands by the autonomous experiment system developed by his team.

view more 

Credit: Junichi Kaizuka

Researchers at the University of Tokyo and their collaborators have developed a digital laboratory system that fully automates the material synthesis and the structural and physical property evaluation of thin-film samples. With the digital laboratory, or dLab, the team can autonomously synthesize thin-film samples and measure their material properties. The system demonstrates advanced automatic and autonomous material synthesis for data- and robot-driven materials science.

The current research is published in the journal Digital Discovery.

Machine learning, robotics and data are deemed vital to the discovery of new materials. However, although data collection is an essential component, there is a bottleneck in that part of the experimental process.

So, researchers constructed a digital laboratory with interconnected apparatuses for solid materials research. They used robots to collect experimental data, such as synthesis processes, and measured physical properties, including measurement conditions. Their dLab consists of a variety of modular experimental instruments that are physically interconnected. This allows researchers to fully automate processes from material synthesis to a wide range of measurements for surface microstructures, X-ray diffraction patterns, Raman spectra (a chemical analysis technique using scattered light), electrical conductivity and optical transmittance.

The dLab consists of two systems. One system integrates experimental instruments to perform automated materials synthesis and measurements, while the other handles data collection and analysis. Each measurement instrument provides data outputted in an XML format data storage called MaiML, which is collected in a cloud-based database. Then, the data is analyzed by software and utilized on the cloud.

“We demonstrated that the system can autonomously synthesize a thin-film material specified by a researcher,” said Professor Taro Hitosugi of the University of Tokyo’s Graduate School of Science. Using dLab, his team demonstrated the autonomous synthesis of lithium-ion positive-electrode thin films and their structural evaluation via X-ray diffraction pattern measurements.

In recent years, machine learning and robotics have provided researchers with new ways to conduct automatic and autonomous experiments. “Today, laboratories are not merely the places to house experimental instruments, but rather the factories for producing materials and data, where experimental equipment operates as a system,” said Hitosugi.

By assigning repetitive experimental tasks to robot systems controlled by machine learning, researchers can synthesize, measure and analyze a large number of samples, thereby generating extensive data. This data- and robot-driven science, along with the standardization of materials synthesis and measurement instruments, and the automation of data collection, will significantly impact how research is conducted.

“Our current work addresses the challenges of accelerating research in materials science,” said Hitosugi. “Our approach enhances the use of data in research. We aim to create a research environment where researchers can focus on creativity. Introducing machine learning and robotics will further advance materials science, deepening theory and finding new materials.”

However, even with the recent progress, the modularization and standardization in solid materials research is still rudimentary. One factor contributing to this is the lack of established standards for the shapes and sizes of samples and sample holders. Solid materials are available in a variety of physical shapes, including powder and bulk forms. Researchers need standardized sample shapes and sample holders. A unified format for measurement data is also lacking, complicating data collection. The Japan Analytical Instruments Manufacturers Association (JAIMA) has collaborated with member companies and the Ministry of Economy, Trade and Industry, to establish a data format called the Measurement Analysis Instrument Markup Language (MaiML). MaiML was registered as the Japanese Industrial Standard in 2024. This standardized format provides a unified format for data collection and use.

Looking ahead, the team hopes to improve the system by standardizing the orchestration software and scheduling. This would allow the researchers to expand the materials explorations and to manage tasks for multiple samples more efficiently. Their goal is to leverage dLab to accelerate material development. “We aim to digitalize the research and development environment, foster researchers who can utilize these technologies, and facilitate data sharing and utilization” said collaborating researcher and lead author Kazunori Nishio, a specially appointed associate professor at the Institute of Science Tokyo. “This environment will fully leverage the creativity of researchers.”

###

Research paper:

Kazunori Nishio, Akira Aiba, Kei Takihara, Yota Suzuki, Ryo Nakayama, Shigeru Kobayashi, Akira Abe, Haruki Baba, Shinichi Katagiri, Kazuki Omoto, Kazuki Ito, Ryota Shimizu, and Taro Hitosugi, “Digital laboratory with modular measurement system and standardized data format,” Digital Discovery: May 14, 2025, DOI: 10.1039/D4DD00326H

Link: https://doi.org/10.1039/D4DD00326H

Funding:

Japan Science and Technology Agency (JST) through Future Society Creation Project (MIRAI) and Strategic Innovation Program (CREST); Japan Society for the Promotion of Science (JSPS); Tokyo Institute of Technology Open Facility Center, Ministry of Education, Culture, Sports, Science and Technology (MEXT) Data-driven Materials Research and Development Project at the University of Tokyo “Center for Research and Development of Electrochemical Materials toward Maximum Introduction of Renewable Energy”; MEXT Grant-in-Aid for Transformative Research Areas “ion jamology.”

Related links:

Graduate School of Science

https://www.s.u-tokyo.ac.jp/en/

 

Hitosugi Lab

https://solid-state-chemistry.jp/english/index.html

 

About the University of Tokyo

The University of Tokyo is Japan's leading university and one of the world's top research universities. The vast research output of some 6,000 researchers is published in the world's top journals across the arts and sciences. Our vibrant student body of around 15,000 undergraduate and 15,000 graduate students includes over 4,000 international students. Find out more at https://www.u-tokyo.ac.jp/en/ or follow us on X (formerly Twitter) at @UTokyo_News_en

---

Journal

Digital Discovery

DOI

10.1039/D4DD00326H 

Article Title

Digital laboratory with modular measurement system and standardized data format

Article Publication Date

14-May-2025

The diagram shows the overall modular synthesis system (sputtering deposition system), which is connected to various measurement and analysis devices. Details of the shape of the sample holder and the communication protocols used when connecting each module have been made publicly available by the research team.

Credit

Kazunori Nishio, Science Tokyo

  

The system consists of a collection of microservices, with the layers below the orchestration software forming the physical layer directly connected to the hardware.

Credit

Taro Hitosugi, Akira Aiba, The University of Tokyo, Science Tokyo

The automated analysis program operates by smoothing data, then detecting, classifying and indexing peaks.

The automated analysis program operates by smoothing data, then detecting, classifying and indexing peaks.

Credit

Kei Takihara, Akira Aiba, Kazunori Nishio, Science Tokyo

 

Special report highlights LLM cybersecurity threats in radiology

Radiological Society of North America

LinkedInWeChatBlueskyMessageWhatsAppEmail

 

image: 

Summary of the cybersecurity threats posed by LLMs in health care. LLM = large language model, AI = artificial intelligence.

view more 

Credit: Radiological Society of North America (RSNA)

OAK BROOK, Ill. – In a new special report, researchers address the cybersecurity challenges of large language models (LLMs) and the importance of implementing security measures to prevent LLMs from being used maliciously in the health care system. The special report was published today in Radiology: Artificial Intelligence, a journal of the Radiological Society of North America (RSNA).

LLMs, such as OpenAI’s GPT-4 and Google’s Gemini, are a type of artificial intelligence (AI) that can understand and generate human language. LLMs have rapidly emerged as powerful tools across various health care domains, revolutionizing both research and clinical practice. These models are being employed for diverse tasks such as clinical decision support, patient data analysis, drug discovery and enhancing communication between health care providers and patients by simplifying medical jargon. An increasing number of health care providers are exploring ways to integrate advanced language models into their daily workflows.

“While integration of LLMs in health care is still in its early stages, their use is expected to expand rapidly,” said lead author Tugba Akinci D’Antonoli, M.D., neuroradiology fellow in the Department of Diagnostic and Interventional Neuroradiology, University Hospital Basell, Switzerland. “This is a topic that is becoming increasingly relevant and makes it crucial to start understanding the potential vulnerabilities now.”

LLM integration into medical practice offers significant opportunities to improve patient care, but these opportunities are not without risk. LLMs are susceptible to security threats and can be exploited by malicious actors to extract sensitive patient data, manipulate information or alter outcomes using techniques such as data poisoning or inference attacks.

AI-inherent vulnerabilities and threats can range from adding intentionally wrong or malicious information into the AI model’s training data to bypassing a model’s internal security protocol designed to prevent restricted output, resulting in harmful or unethical responses.

Non-AI-inherent vulnerabilities extend beyond the model and typically involve the ecosystem in which LLMs are deployed. Attacks can lead to severe data breaches, data manipulation or loss and service disruptions. In radiology, an attacker could manipulate image analysis results, access sensitive patient data or even install arbitrary software.

The authors caution that cybersecurity risks associated with LLMs must be carefully assessed before their deployment in health care, particularly in radiology, and radiologists should enact protective measures when dealing with LLMs.

“Radiologists can take several measures to protect themselves from cyberattacks,” Dr. D’Antonoli said.  “There are of course well-known strategies, like using strong passwords, enabling multi-factor authentication, and making sure all software is kept up to date with security patches. But because we are dealing with sensitive patient data, the stakes (as well as security requirements) are higher in health care.”

To safely integrate LLMs into healthcare, institutions must ensure secure deployment environments, strong encryption and continuous monitoring of model interactions. By implementing robust security measures and adhering to best practices during the development, training and deployment stages, stakeholders can help minimize risk and protect patient privacy.

Dr. D’Antonoli notes that it is also important to use only the tools that have been vetted and approved by an institution’s IT department, and any sensitive information used as input for these tools should be anonymized.

“Moreover, ongoing training about cybersecurity is important,” she said. “Just like we undergo regular radiation protection training in radiology, hospitals should implement routine cybersecurity training to keep everyone informed and prepared.”

According to Dr. D’Antonoli, patients should be aware of the risks but not overly worried.

“The landscape is changing, and the potential for vulnerability might grow when LLMs are integrated into hospital systems,” she said. “That said, we are not standing still. There is increasing awareness, stronger regulations and active investment in cybersecurity infrastructure. So, while patients should stay informed, they can also be reassured that these risks are being taken seriously, and steps are being taken to protect their data.”

###

“Cybersecurity Threats and Mitigation Strategies for Large Language Models in Healthcare.” Collaborating with Dr. D’Antonoli were Ali S. Tejani, M.D., Bardia Khosravi, M.D., M.P.H., Christian Bluethgen, M.D., M.Sc., Felix Busch, M.D., Keno K. Bressem, M.D., Lisa Adams, M.D., Ph.D., Mana Moassefi, M.D., Shahriar Faghani, M.D., and Judy Wawira Gichoya, M.B.Ch.B., M.S.

Radiology: Artificial Intelligence is edited by Charles E. Kahn Jr., M.D., M.S., Perelman School of Medicine at the University of Pennsylvania, and owned and published by the Radiological Society of North America, Inc. (https://pubs.rsna.org/journal/ai)

RSNA is an association of radiologists, radiation oncologists, medical physicists and related scientists promoting excellence in patient care and health care delivery through education, research and technologic innovation. The Society is based in Oak Brook, Illinois. (RSNA.org)

---

Journal

Radiology Artificial Intelligence

Subject of Research

Not applicable

Article Title

Cybersecurity Threats and Mitigation Strategies for Large Language Models in Healthcare

Article Publication Date

14-May-2025

 

UCF students’ AI system assists Orlando Health robotic surgeries

Developed through the Senior Design capstone course, the innovation increases efficient use of medical supplies during procedures

University of Central Florida

Facebook

XLinkedInWeChatBluesky\MessageWhatsAppEmail

 

image: 

Orlando Health’s Alexis Sanchez (far right) says medical-engineering partnerships, like this one with UCF’s Laura Brattain (far left) and her students, are the best way to bring together “the best minds” to solve medical challenges.

view more 

Credit: UCF

A University of Central Florida (UCF) student engineering project that began with using artificial intelligence (AI) to track cafeteria forks transformed into a system that will help Orlando Health surgeons perform robotic surgeries more efficiently.

Laura Brattain, a UCF biomedical engineer, mentored six College of Engineering and Computer Science seniors, who developed the AIMS (AI for Medical Surgery) system that keeps track of surgical staples, enabling surgical teams to operate more efficiently, reduce waste and improve sustainability. The new technology was developed as part of the college’s Senior Design capstone course that encourages students to create a usable product before they graduate.

Students built an end-to-end application and tested it in an operating room at Orlando Health several times to improve the application. Alexis Sanchez, robotic surgery program director at Orlando Health Orlando Regional Medical Center (ORMC), participated in the project and is now using the system in his surgeries.

As Florida’s Premier Engineering and Technology University, UCF is focused on leveraging technology to strengthen the health of communities. That is Brittain’s research focus — integrating biomedical AI, medical ultrasound and surgical robotics to create healthcare innovations that improve care. An associate professor at UCF’s College of Medicine and a faculty member of the UCF AI Initiative, she holds secondary positions in the College of Engineering and Computer Science.

Sanchez says the technology can be applied to many other processes in the future, such as keeping track of instrument usage during non-robotic surgeries.

“We work in a very fast-paced environment, so having this to be able to detect waste has incredible potential to improve both efficiency and sustainability,” he says. “This is just the beginning. And this collaboration underscores both Orlando Health and UCF’s commitment to innovation to improve healthcare for our community.”

How the System Works

Many of the new medical tools developed for the operating room are disposable. Once they have been removed from their sterile containers and placed on the operating room table, they must be discarded — even if unused — because they are no longer considered sterile. During robotic surgeries, the robot cuts and staples tissues at the same time to reduce bleeding. But no one is sure how many staples a particular surgery will take.

During a visit to the hospital, Brattain joined Sanchez in observing the entire process of performing a robotic surgery, from preparation to completion. After evaluating potential areas for improvement, they suggested that students develop an AI system to track how many staples are placed on the surgical table versus how many are actually used.

“I wanted the students to know that while they can all create a computer program, they can also make an impact in healthcare,” she says. “To avoid developing technologies that end up collecting dust on the shelf, we should work with clinical experts to solve problems that can ultimately improve the care of patients.”

AIMS has a camera feed linked to a computer in the operating room. During surgery, their AI software directs the camera to record each staple that comes into the operating room and track its use. That data can then be analyzed to determine exactly how many staples are used to avoid opening unnecessary staples for surgery.

Life in the real world of surgery offered unique challenges to the young scientists. They went through multiple iterations with Sanchez and his team at Orlando Health. Initially they didn’t account for the low light conditions in operating rooms, so they had to change the camera’s angles and settings to better capture photos of the staples. They had to address other issues: What happened if someone placed a tool in front of the staples during surgery? What happened if someone moved the staples or stepped in front of the camera?

“We are thankful that Dr. Sanchez and his team provided the students with the opportunities to test AIMS in real-world scenarios where a regular robotic procedure is happening in the operating room and the medical team is moving around as usual,” Brattain says. “You can’t imagine these things in a classroom. Students need to see their science through a medical provider’s eyes.”

Creating Real-World Technologies

The goal of the Senior Design capstone is to “give students the opportunity to say, ‘I actually made something,’” during their education, says Matthew Gerber, a faculty member who helps lead the module. “We love it when the project turns into an application that’s being used in the real world. We wish it would happen more often.”

To further engineering-medicine partnerships, Brattain offered an Introduction to Medical Robotics course to engineering majors this semester — the first time the course has been offered in a decade. Students learned about how medical robots are designed and manipulated. As part of the class, they visited Sanchez’s team and saw the hospital’s Da Vinci robot in action. Students had a unique chance to interact with robotic surgeons, who have been generous with sharing their knowledge and answering questions. All these visits were coordinated by Lillian Aguirre, a clinical nurse specialist on Sanchez’s team and a UCF College of Nursing alum. The students and Brattain are grateful for her dedicated assistance.

Sanchez was at UCF when the students presented their Senior Design project and says he was proud of what they had accomplished together.

“One of the students came up to me and there were tears in his eyes,” Sanchez says. “He said, ‘I always hoped my skills would help humanity one day, and now I have.’”

Gerber sees plenty of future opportunities to create UCF engineering-medicine systems that impact patient care.

“Doctors are faced with so much information,” he says. “With AI, we can quickly and objectively analyze all that information to help give doctors better, cleaner information. AI can say to doctors, ‘Don’t worry about that. Focus on this.’”

Orlando Health is a UCF Pegasus Partner, a program that offers opportunities for select partners to engage across the university in ways that create meaningful value for both organizations. That engagement includes talent development and recruitment, shared research projects, joint ventures and collaborations, and strategic philanthropy.

Working together on projects like this creates synergy and provides the potential for advances in the science of medicine in Central Florida.

Rachel Leiner ’25, who graduated from UCF this spring, was the student leader for the AI project.

“Coming into a project that’s for a grade and seeing that we made something that can help improve the hospital workflow makes me very proud,” she says. “We started this project by developing AI to track cafeteria forks. We had nothing, and in eight months we had a working software app and a usable AI model to track surgical staples in an operating room.”

 

Jean Zay supercomputer: France has increased its AI dedicated resources fourfold

CNRS

Facebook

XLinkedInWeChatBlueskyMessageWhatsAppEmail

 

image: 

Detail of the computing racks in the fourth extension of the Jean Zay supercomputer. Acquired in 2019 by the Grand équipement national de calcul intensif (Genci), the supercomputer is named in honor of Jean Zay, politician and architect of the creation of the CNRS. The 2024 extension designed by Eviden, comprising 14 BullSequana XH3000 computing racks, increases its computing power to 125.9 petaflops (over 125 million billion operations per second). It enables national academic and industrial research communities to carry out numerical simulations using high-performance computing, and to train or specialize generative artificial intelligence models. It is one of the most eco-efficient machines in Europe, thanks to accelerated technologies (GPU) and core cooling of the computing servers. Waste heat is also reused to help heat housing on the Saclay plateau. The Jean Zay supercomputer is hosted and operated by CNRS, via the Institut du développement et des ressources en informatique scientifique (Idris).

Translated with DeepL.com (free version)

view more 

Credit: © Cyril FRESILLON / IDRIS / CNRS Images

• 
  
• The fourth extension of the Jean Zay supercomputer has increased its computing power by 4, making it one of the most powerful, and the most used, supercomputers in France and Europe in the field of artificial intelligence.
• These new capacities will be made available for strategic scientific and industrial applications.
• Countless research and innovation fields will be able to take advantage of this expanded power free of charge for open research, including biomedical research, astronomical data analysis, self-driving, new material design, new energy, agriculture, decision support, and culture, among others.

The Jean Zay supercomputer was put into service in 2019. It replaced its predecessor Turing, whose computing power it increased tenfold. To meet the growing needs of the artificial intelligence community, it subsequently integrated successive extensions that improved its performance and functionalities. The fourth of these was announced in June 2023 by French President Emmanuel Macron during the VivaTech trade show.

Acquired by GENCI from the European manufacturer Eviden after a competitive dialogue procedure, Jean Zay 4 now has 125.9 petaflops of 64-bit computing power, corresponding to 125.9 million billion operations per second. If all of humanity counted at the rate of one operation per second, it would take 182 days to calculate what the Jean Zay supercomputer can do in one second. The supercomputer’s storage capacity also increased, reaching the order of 100 petabytes.
Thousands of research projects have already used Jean Zay for free, hence the importance of Jean Zay 4’s resources for academic research teams, start-ups, and large companies. Some projects are highly emblematic, and enjoy international renown. To illustrate this enthusiasm, three scientific projects were presented during the inauguration:

• Polymathic, the project by François Lanusse, a CNRS researcher at the Astrophysique, Instrumentation, Modélisation Laboratory (CNRS/CEA/Université Paris-Cité), applies deep learning and generative AI technologies to astronomical observations of the Universe on a large scale;
• The Owkin start-up project led by research engineer Jean-Baptiste Schiratti develops more solid and generalisable AI models for detecting pathologies, notably for untangling semantic content and texture in images from microscopes;
• The project by the start-up Pleias is led by its co-founder Pierre-Carl Langlais. It offers improved application of large language models, doing so offline and with traceability of information sources that is compatible with the new EU AI Act.

This version of the Jean Zay supercomputer with expanded capacities was produced by the French and European manufacturer Eviden (Atos Group). In addition to its converged HPC/AI1 computing power, Jean Zay is among Europe’s most eco-efficient supercomputers thanks to its graphic processors (GPU) provided by Nvidia, and next generation warm-water cooling proposed by Eviden. In addition, the installation’s residual heat is gathered to heat the equivalent of 1,500 households on the Plateau de Saclay.

Thousands of projects will be able to harness this new tool free of charge for open research in language processing, multimodal computing, biomedical research, fundamental physics, climatology, new materials and energies, autonomous vehicles, decision support, agriculture, and culture, among others. All of these fields will take advantage of the possibilities offered by the supercomputer, especially for the training, specialisation, and inference of AI foundation models.

A lever for accelerating French and European power in artificial intelligence, the Jean Zay supercomputer saw a 20-fold increase–thanks to the constant and exceptional support of its technical teams–in the annual number of AI projects selected over the last five years, rising from 72 in 2019 to more than 1,400 in 2024. It is among the machines with the most success in Europe in the field of AI.

Jean Zay will also be central to the future European project AI Factory France, which will also be celebrated soon, and brings together GENCI, the CNRS, Inria, the CEA, AMIAD, CINES, France Universités, the 9 AI Clusters, French Tech, the HubFranceIA, and Station-F. This project offers infrastructure services for computing, support, training, teaching, and expertise for the French and European AI community.

Finally, in a nod or mise en abyme, Jean Zay 4’s innovative design was completed by Obvious, an internationally recognized French collective that produces works of art using AI technology.  

The inauguration took place in the presence of Antoine Petit, the Chairman & CEO of the CNRS, Philippe Lavocat, the CEO of GENCI, Bruno Bonnell, General Secretary for Investments, Jean-Luc Moullet, the Director General for Research and Innovation at the French Ministry of Higher Education and Research, and Hélène Mouchard-Zay, Jean Zay’s daughter. Antoine Petit marked the occasion, declaring:

“I am proud to inaugurate the new extension of Jean Zay, the jewel of French supercomputers. France needs this innovative and cutting edge machine with the latest advances in AI to respond to major scientific challenges. It is with determination that the CNRS has mobilised its teams since 2018 to offer computing capacities and power to the entire national community.”

Philippe Lavocat also emphasized:

“The extension of Jean Zay 4 will provide new energy to all academic and industrial researchers seeking to expand their skills in developing artificial intelligence, with a view to meeting major scientific and societal challenges, notably with the help of generative AI models. We made a dual wager: to possess a powerful and technologically advanced machine, and to invest in human resources in support of users. Success in this effort offers a major advantage in networking with European machines. Jean Zay 4 addresses national sovereignty considerations, as joining forces in AI across Europe can still make a difference thanks to the enormous potential of our experts in this field.”

Notes

1 – Computing power benefiting from the convergence of traditional computing techniques as well as IA’s (CNRS/Obs. de Paris/Sorbonne Université) and the Centre de recherche astrophysique de Lyon (CNRS/ENS de Lyon/Université Claude Bernard).

About Jean Zay

This supercomputer was named after Jean Zay, French Minister of National Education and Fine Arts between 1936 and 1939. He was murdered by the French Milice on June 20th 1944. Along with Irène Joliot-Curie and Jean Perrin, he took the first steps that led to the creation of the CNRS on October 19th 1939. Jean Zay was added to the French Panthéon on May 27th 2015.