Self-driving labs are changing how chemists work

On a frigid afternoon in early February, Atinary Technologies, an artificial intelligence research and development firm, opened a chemistry laboratory in Boston’s Seaport neighborhood. Among the first visitors was a team from the drugmaker AstraZeneca that wanted to see the lab in action. The special thing about this lab is that it relies less on humans and more on AI agents and robots to perform chemistry experiments. It’s a step toward what is known as a self-driving lab.

Atinary has raised at least $10 million in outside investment to generate data from hundreds of chemical reactions each day. This process could take a human chemist weeks.

Last year, Radical AI, which has raised $65 million in funding, launched a similar facility at the Alexandria Center for Life Science, a campus in New York City that houses nearly 50 biotech and pharma companies. Inside the facility, a massive chest of drawers is piled with dime-sized pucks covered with metal alloy. The alloys, produced by the company’s AI-centric lab, are candidate materials for building machines such as hypersonic vehicles and spacecraft that operate in the harshest environments.

In both new labs, a human scientist queries the company’s AI agent about the best route to make a molecule or material. The agent, referred to as the “brain” of the lab, combs the scientific literature and chemistry databases. It generates a recipe and communicates it to robots that perform the experiments, analyze the results, and relay the data back to the agent, which learns from the findings and suggests the next iteration of the molecule or material.

Since this experimental loop happens under the supervision of or in collaboration with humans, the process is known as “human in the loop” in academic circles. Proponents of such labs say the method can quickly and efficiently explore the chemical space to find molecules and materials that, for example, help cure illness or stem climate change.

Atinary and Radical are just two of a growing number of start-ups that are raising massive amounts of venture capital to develop labs where AI and bots run the show. Such labs are not new, but earlier generations tended to be good at performing only specific tasks. The rise of AI agents, fueled by large language models (LLMs), plus plummeting hardware prices and advances in computer vision, has made these labs more autonomous, requiring less human intervention than before.

Self-driving setups remain costly, despite efforts by some research groups to bring prices down. And currently, bots can by no means perform every type of chemistry that a human chemist can. Some skeptics don’t expect fully self-driven labs to arrive anytime soon. Still, many researchers are confident that this day, long the subject of science fiction, could become a reality.

Humans prompt, agents plan, bots hustle in self-driving labs

At Atinary’s Boston launch, a huge monitor stood outside the lab. On the screen was a chatbot for communicating with the company’s AI agent. Justin Potnick, a senior principal lab scientist at the company, asked the agent to optimize the conditions for the Buchwald-Hartwig reaction to maximize yield while minimizing catalyst use. The reaction, catalyzed by the expensive metal palladium, forms carbon-nitrogen bonds. It is commonly used in industry to make the core structures of pharmaceuticals and agrochemicals.

In a few minutes, the AI agent scanned all publicly available databases and generated a detailed, step-by-step protocol for the reaction, including reaction conditions such as temperature and pressure. That blueprint—which is usually reviewed by a scientist–was transmitted to the lab’s robots.

“It can explain all that it is thinking and the reasoning behind it, which is great,” Potnick said in an interview during the tour. “The software accelerates the process of creating an experiment down to minutes. Scientists doing this could have spent half a day or 2 days building the experimental setup.”

Inside the lab, a robotic arm fixed on a workstation moved around confidently, sometimes rotating 360°, as if performing a pop-and-lock dance. It picked up test tubes and pipetted reagents from miniature containers prefilled by a human chemist. It then dispensed the reagents into tiny wells, where the reactants combined to form amide bonds.

Another part of the automated lab analyzed the products with mass spectrometry and nuclear magnetic resonance to determine the structure and quality of the resulting molecules. The results were then relayed back to the agent, which learns from the data and suggests a new set of reaction conditions for the next iteration.

Traditionally, reaction optimization—tweaking conditions, reagents, and catalysts to achieve the desired results—has been performed by human chemists, who often find the work repetitive and mundane. The beauty of automation is that it can do these experiments quickly with minimal errors, Potnick said.

Loïc Roch, Atinary’s cofounder and chief technology officer, said that every week, the lab produces roughly as much data as a student can produce in the course of a PhD program. And getting more data is crucial to improving the performance of AI models. “The scientific literature is filled with what has worked but has very little about what has failed,” Roch said. “We don’t publish things that do not work. And that’s a problem because then they don’t know where not to go” in the chemical space.

In New York City, where Radical is solving a similar problem in the materials world, the company’s AI agent, called Antimatter, is the lab’s mastermind. When queried, it reads thousands of publications and makes hypotheses about effective alloy recipes, says Joseph F. Krause, the company’s cofounder and CEO.

Human scientists review Antimatter’s suggestions and decide which are the best candidates for making, say, a turbine blade. Brown bottles automatically dispense small metal pellets, which a stationary robot mixes and heats according to the agent’s recipe.

The resulting materials are tested in a chamber slightly larger than a hotel-room safe; these chambers approximate extreme environments such as the heat generated by rockets or the cold and pressure of high altitudes. The results are relayed back to Antimatter, which learns and suggests revisions.

Krause says that AI can clearly work much faster than human chemists and that Radical’s lab can do in a day what would take human scientists several weeks. But the promise of AI is not just data generation. AI is often unafraid to explore the unexplored, he says. “The search space is so big, and one of the problems with the human scientist is that we have bias. We are used to things that we have previously made. And that’s not because we are not smart. It’s just natural to go to where you feel comfortable and want to explore,” he says.

In Radical’s present setup, the lab produces about 70 alloys a week, but by the end of summer, when it moves to a new facility across the East River in the Brooklyn Navy Yard, the company will be able to produce 100 alloys per day, Krause says. “And then we can look at all the results, all in real time, and be constantly updating that loop, very, very fast.”

Efforts to automate lab workflows and experiments date back to the late 19th century. “The earliest mention of automation in the chemical literature of the United States was in 1875, announcing a device to wash filtrates unattended,” Kevin Olsen, a biochemist, says in a 2012 SLAS Technology review. Even though a modern lab would be almost unrecognizable to a 19th-century scientist, Olsen writes, the reasons for automating experiments remain essentially unchanged: to save time and improve performance by eliminating human error.

But labs like those of Radical and Atinary do more than automate. They are also constantly learning from the new data they generate. Their machine learning methods enable the agents to make specific decisions and provide a degree of autonomy. “A self-driving lab is like a Waymo,” Krause says. “You can get in and tell it where you want to go. You get in, close your eyes, and you end up at that destination. That’s true self-driving.”

Krause acknowledges that right now, his lab still needs a human in the loop to make decisions. “Our AI is young. Like it’s still getting its PhD.” In the future, he says, he wants his lab’s AI agent to be as good as a human chemist with a lifetime of experience.

An evolution of both software and hardware in self-driving labs

Researchers began building autonomous labs around 2010. Many such labs emerged, but they typically could perform only very specific tasks. And sometimes they just weren’t efficient.

In 2017 Eli Lilly and Company spent about $90 million to expand its San Diego campus. A big part of that project was the Lilly Life Sciences Studio lab, a facility where researchers working remotely could control their experiments via a web-based interface in a fully automated, closed-loop lab.

The pharma giant gave up on that dream in 2024, when it disassociated itself from Strateos, the company that operated the interface. The lab’s equipment was eventually sold to two contract research companies: Arctoris and SpiroChem.

Since the Lilly lab was cloud based, the idea was that scientists could sit on their couches and request the synthesis of a molecule, but it never really worked out, says Thomas Fessard, cofounder of SpiroChem.

But Lilly’s foray into the automated lab largely predated the emergence of LLMs such as Claude, Gemini, and ChatGPT, which understand and generate text in human languages. LLMs, researchers say, are the fuel powering today’s self-driving chemistry labs.

“I think things really changed when large language models came out,” says Andrew White, cofounder of FutureHouse and Edison Scientific, firms that build AI agents for scientific discovery. “They are so flexible that they can do both the analysis and protocol design, and then they can even write robot instruction code to execute.”

In 2024 White and a team of researchers launched ChemCrow, an AI chemistry agent. To create it, the researchers plugged 18 well-established chemistry tools, such as RXN for Chemistry and AiZynthFinder, into the OpenAI LLM called GPT-4. The agent autonomously plans experiments and relays the workflows to an automated lab. The researchers demonstrated the synthesis of an insect repellent, three organocatalysts, and a novel chromophore, all using ChemCrow.

While ChemCrow was among the first of its kind, many such chemistry agents have sprung up in the last few years, powered by LLMs.

Alexander Hammer, cofounder and CEO of Dunia Innovations, a start-up that raised $11.5 million to set up a self-driving lab for materials discovery, says AI agents already exhibit superhuman capabilities. “Nobody can read every paper that has ever been published on a given topic. It is very difficult to hold it in our brains,” he says. Hammer advises thinking of agents as AI coscientists with whom human scientists can jointly discover “white spaces” in research where they can hypothesize new solutions.

Jonathan Grob, vice president of small molecules at the autonomous lab operator Ginkgo Bioworks, says thinking that AI is just spitting out an answer would be a mistake. Rather, it’s giving suggestions and having a conversation with a human chemist. “As a medicinal chemist, there is so much more than deciding whether to put a methyl group here or a methyl group there, and then the AI gives you some ideas and direction,” he says. 

Just as the emergence of LLMs has radically changed the landscape for lab automation, so too have hardware developments.

Up until 2021, one of the easiest things to automate was liquid handling, and that task was what most instrument companies offered in terms of automation, says Animesh Garg, a computer science professor at the Georgia Institute of Technology. “These were what you would call pick-and-place robots,” Garg says.

In the past 3 years, the field has leaped forward, Garg says. Once researchers saw what advanced LLMs could do, they became interested in robots that could be general purpose, performing multiple tasks and navigating the world in real time.

Today, research groups, including Garg’s, use machine learning to teach robots from hundreds of thousands of photos and videos of human scientists working in the lab. The images depict practical tasks such as picking up a slide, opening reagent bottles, wiping a surface, and preparing a sample.

Garg offers the analogy of asking a child to bake a cake. “No amount of [instructions] in English will make it happen. But if you show it step by step, like say, ‘Open the drawer, pull out a bowl, pick up three eggs,’ so on and so forth, maybe the child will learn.”

Manufacturing improvements have also made the robots cheaper. The price of a robotic arm for an automated lab has dropped from $50,000 in the early 2000s to under $2,000, Garg says, calling the figures a rough estimate.

And although Lilly abandoned its San Diego studio, it has not given up on autonomous labs. In January, Lilly and Nvidia, the AI and computing behemoth, announced a $1 billion coinvestment over the next 5 years in robotics and AI to ramp up the discovery of new drug candidates.

Making bots versatile, humanlike, and affordable

While many self-driving setups, like those at Atinary and Radical, remain anchored to the benchtop, some academic research groups are trying to build humanoid bots to work as chemists, and others are trying to cut the costs of such robots even more.

At the forefront of the endeavor to program mobile, humanoid bots with chemist-like capabilities is Andy Cooper’s lab at the University of Liverpool.

In 2020 the lab welcomed a new member, an AI-driven mobile robot about as tall as the humans who work there. The bot, built by the robotics company Kuka, rolled around the lab from one workstation to another, picking up vials, operating instruments, and dispensing liquids and solids. Called the mobile robotic chemist, it could work tirelessly for up to 21 h before it needed a recharge.

“Our strategy uses a dexterous free-roaming robot, automating the researcher rather than the instruments,” Cooper and colleagues write in the Nature publication in which they first reported their work on the mobile bot.

That mobile chemist was programmed to work in darkness, a talent that came in handy for discovering photocatalysts for producing hydrogen from water. The robot autonomously worked 24/7 for 8 days, performed 688 experiments, and identified photocatalyst mixtures that were six times more active than the initial formulations, the authors note.

Then, in 2024, Cooper’s lab announced that the mobile chemist had been programmed for new abilities. The bot could operate a chemical-synthesis platform, a liquid chromatography/mass spectrometry (LC/MS) instrument, and a benchtop NMR machine. It was able to analyze the data with an accuracy comparable to that of a trained synthetic chemist, says Sriram Vijayakrishnan, a senior postdoctoral researcher in the lab.

And this year, the group revealed that the bot has become even more versatile. It synthesized acetaminophen in an automated reactor, used an ultra-high-performance LC/MS instrument to analyze the product, and even cleaned the reactor between runs. “Reaction yields and purity match human chemist performance,” the team notes in a paper published in the journal Digital Discovery.

Vijayakrishnan says the fact that the mobile chemist can work alongside human scientists is a major advantage over highly engineered, immobile setups, which are hard to repurpose for different uses.

But flexibility can come at a cost. Timothy Noël, a professor at the Van’t Hoff Institute for Molecular Sciences at the University of Amsterdam, is on a quest to lower the cost of robotics. Noël completed his postdoctoral work at the Massachusetts Institute of Technology, where he didn’t see people losing sleep over funding. But the reality in Amsterdam is different.

Noël says his goal is to build a self-driving lab that is financially accessible. His team’s first iteration, RoboChem, is an automated flow-chemistry system. The latest iteration, RoboChem-Flex, was announced earlier this year. It combines readily available components with components that can be made through inexpensive in-house 3D printing, offering flexibility and affordability. Noël says these features have reduced the system’s cost to $5,000.

Can AI and robots really do it all in the chemistry lab?

Even with major advances in robotics and AI, almost nobody in the field says today’s generation of self-driving labs can do everything.

Cooper suggests that no fully autonomous labs exist in the truest sense—at least not yet. “Because that would imply that it can do everything and order the chemicals and fix itself. There’s a spectrum between automated, which is just programmed, and some level of autonomy,” he says.

Other researchers question whether the current generation of AI agents built on LLMs can generate novel hypotheses. “I think they [LLMs] can produce something that is plausible and that fits into the space of other hypotheses that have been proposed before,” says Heather J. Kulik, a chemical engineering professor at MIT. “But some truly breakthrough hypotheses would truly surprise me.”

Kulik recognizes AI companies’ claims that AI agents can reason scientifically, but she says it is not the same as a human scientist’s thinking. “The challenge is that we also don’t know, when we come up with a hypothesis in the lab, how much of that is intuition versus how much of that is built rigorously on our cognition.”

And while AI agents are not ready for Nobel Prize–quality science, Kulik says, they offer a lot of value for optimizing reactions and predicting properties based on well-characterized literature, or even in the field of combinatorial exploration.

Using AI to design high-entropy alloys like the ones Radical designs is a good example of using the tool, since a lot of data are already available, Kulik says. “These are just the low-hanging fruits.” But in the most cutting-edge areas of materials discovery, little to no data are available, making these areas ill suited for AI, at least for now.

Vijayakrishnan holds a similar view about small-molecule synthesis. Using AI and automation is useful mostly when a researcher knows something about the nature of the product of a chemical reaction but not when, say, the product is a molecule not represented in the literature. “That’s one of the hardest challenges, where you don’t have a lot of existing data that you can refer to,” he says.

Meanwhile, robotic chemists are only beginning to mature and aren’t adept at every task, Cooper says. “I think it’s important to realize that these robot-lab-chemist systems are not like an iPhone, which has gone through many iterations.” At the moment, he says, making any molecule that requires more than three synthetic steps is too hard for them.

Fessard of SpiroChem, the contract research organization that bought some of the automated equipment in the Lilly lab, says the machines were inept at precisely picking up and weighing solids, a task that is essential to a chemistry lab. “It works okay when it’s a nice crystalline powder like table salt, but if you have something sticky, like a gum, something hygroscopic, the robots cannot efficiently weigh them.”

Still, Fessard agrees that some repetitive tasks, when performed by the bots, are actually time-saving. He gives the example of peptide synthesis, which has a straightforward protocol of stringing amino acids with the same kinds of bonds. But he is not convinced that self-driving labs can outperform humans in troubleshooting or in discovering new molecules.

“Anyone who says self-driving labs are foolproof and can do everything that humans can—I will not buy that,” Fessard says.

Some researchers say the field is progressing so rapidly that many of these limitations will be overcome in the future. “It’s crazy how quickly things are happening,” says Alán Aspuru-Guzik, a chemist and computer scientist at the University of Toronto and a researcher at Nvidia. He is among the people who have been dreaming of and devising such labs and founded companies long before other start-ups began fiddling with the idea.

One of the agents in his lab is Organa, an AI and robotic assistant that came into being in 2024. The bot, which Aspuru-Guzik refers to as she, can interact with human scientists in English, help design and plan chemistry experiments, and conduct tasks such as titrations and electrochemical measurements.

For now, Organa sits fixed on a workbench, but Aspuru-Guzik says the self-driving lab of tomorrow will be a humanoid robot with two arms that walks around, picks up glassware, performs experiments, and even cleans up. And he holds those aspirations for Organa. “She would be this robot that’s working around the lab, doing things with you. I think that’s what’s going to happen, because robotics is developing so fast.”

Another central figure in Aspuru-Guzik’s lab is El Agente, Spanish for “the agent.” El Agente comprises multiple subagents, each of which can perform distinct computational tasks, such as quantum simulation or modeling of solid-state materials.

El Agente has agentic AI capabilities like developing hypotheses and desgining experiments but does not perform physical experiments like Organa. Aspuru-Guzik says he didn’t plug El Agente into the wet lab right away. Chemistry labs are home to potential hazards and risks due to the presence of chemicals that can be toxic, corrosive, and flammable. “What’s going to be the agent that is going to make sure nothing explodes in the lab?” Aspuru-Guzik says.

To address that, his lab has developed El Agente Seguro, a safety agent that ensures that El Agente’s recipes account for safety. “Whenever we execute the robot, we want to make sure that it doesn’t explode or doesn’t kill a person,” Aspuru-Guzik says. El Agente is now into the lab’s farm of 50 self-driving labs.

Many researchers still see a place for the human chemist in a world of self-driving labs, though the convergence of agents, robotics, and chemistry will change how people do science, Aspuru-Guzik says. He fantasizes about the day when the AI agents from his lab come up with a new coupling reaction, “say a new Suzuki-type coupling.” And who will get the recognition for such a reaction? “We will call it Toronto One,” he suggests, “and say it was discovered by an AI agent.”