Infinite Loop · Emerging Technologies
The Laboratory That Designs Its Own Next Question
A self-driving laboratory does not merely run experiments faster. It decides which experiment to run next — and that single shift turns automation into something closer to a thinking instrument. Here is what that actually means.
What You’ll Learn
01 · The Scene
A room where the work continues after everyone goes home
Walk into a certain kind of laboratory at two in the morning and you will find no one there, and yet the work is not paused. A robotic arm lifts a vial, mixes a precise dose of one powder into another, slides a sample into a furnace, and waits. Hours later an instrument measures what came out. Somewhere in a server, a program reads that measurement, compares it against everything the machine has tried so far, and decides — entirely on its own — what the next experiment should be.
No human chose that next step. That is the detail worth pausing on, because it is the detail that separates this room from every automated laboratory that came before it. The robots are impressive, but robots have been lifting vials for decades. What is new is the small, quiet act of judgment happening between experiments: the machine is not executing a list someone handed it. It is writing the list as it goes.
This is a self-driving laboratory, sometimes abbreviated SDL. The name borrows deliberately from self-driving cars, and the analogy is more exact than it first appears. A cruise-control car holds a fixed speed; a self-driving car perceives its surroundings, predicts what will happen, and chooses where to steer. The difference between conventional lab automation and a self-driving laboratory is the same difference — between holding a course someone set, and choosing the course.
To understand why scientists describe this as a reorganization of how research is done, and not just a faster version of the old way, we have to be precise about what is actually looping inside that empty room. So let us build the idea carefully, starting with where it came from.
02 · Origins
Sixty years of teaching machines to ask questions
The dream of a laboratory that reasons is older than most of the technology that finally made it possible. Its lineage runs through artificial intelligence, laboratory robotics, and statistics — three fields that spent decades growing toward each other before they met.
The earliest thread begins in the 1960s at Stanford, with a project called DENDRAL. It was among the first pieces of software built to do something scientists had assumed only humans could: form a hypothesis. DENDRAL took readings from a mass spectrometer — an instrument that breaks a molecule into fragments — and proposed which chemical structures could have produced those fragments. In the 1970s, a successor named Meta-DENDRAL went further, adding a form of closed-loop learning in which the system refined its own rules from the data it was given. The vocabulary of “self-driving laboratories” did not exist yet, but the essential move — let the machine propose, then let the results teach it — was already on the table.
For the next thirty years the two halves of the problem advanced separately. Laboratory robotics matured in the pharmaceutical industry, where high-throughput screening machines could run thousands of identical tests a day. Meanwhile, machine learning grew more capable. The pieces existed; what was missing was the wiring that would let the thinking part steer the doing part in a continuous loop.
We hope to have teams of human and robot scientists working together in laboratories.
— Prof. Ross King, on the Robot Scientist project, Aberystwyth University (2009)That wiring arrived, visibly, in 2009. A system called Adam, built by a team led by Ross King at Aberystwyth University and the University of Cambridge — a project begun in 1999 and physically commissioned in 2005 — became, by its creators’ account, the first machine to independently discover new scientific knowledge. Adam studied baker’s yeast. It generated hypotheses about which genes coded for which enzymes, designed experiments to test them, ran those experiments with its own robotics, interpreted the results, and revised its ideas. It identified roles for a set of genes in the yeast’s metabolism, and human researchers later confirmed the findings by hand. The work appeared in the journal Science.
As a co-author put it, the striking thing was not the data-handling but the reasoning: that Adam could formulate hypotheses on its own and test them. The machine had crossed a line from doing chemistry to investigating it.
From there the field accelerated. In 2018 a chemistry robot in Glasgow searched for new chemical reactions guided by machine learning. In 2020 two landmarks appeared together: Ada, a modular self-driving lab in Canada that optimized thin-film materials, and a mobile robotic chemist in Liverpool that roamed a human-scale lab on wheels, hunting for better light-driven catalysts. And in 2023 the A-Lab at Lawrence Berkeley National Laboratory ran for seventeen days largely unattended, attempting to make dozens of predicted new compounds. Each step widened what “autonomy” could mean.
Each of these systems was built for a different science — genetics, organic chemistry, solar materials, inorganic powders. Yet they share one structural feature, and naming it precisely is the whole point of this article.
03 · Anatomy
Four steps in a circle, and a decision in the middle
Strip a self-driving laboratory down to its skeleton and you find a loop with four stages. Researchers call it the DMTA cycle, after its four verbs: Design, Make, Test, Analyze. It is simply the scientific method, drawn as a circle.
In the Design stage, the system chooses an experiment — what to mix, at what concentration, at what temperature. In the Make stage, robotic and fluid-handling hardware physically carries it out. In the Test stage, automated instruments measure the result. In the Analyze stage, software interprets that measurement and extracts what was learned. Then the loop closes: that learning flows back into the next Design, and the cycle begins again.
A human laboratory runs this same cycle, but with people standing at each station and, crucially, a person deciding what to try next. In a self-driving laboratory, every station is automated and the deciding is automated too. The closing of that loop — the moment where analysis becomes the next design without a person in between — is the entire idea.
Here is a structural analogy that holds up under inspection, not just as a poetic flourish. Consider your own body’s immune system. It does not check a fixed list against every invader. It generates candidate antibodies, tests them against what it encounters, keeps the ones that bind well, and refines them over successive generations — a closed loop of propose, test, select, refine. A self-driving laboratory runs the same kind of adaptive search, only its “antibodies” are experiments and its “binding” is a measured property. The parallel is genuine because both systems share the same logic: learn from each trial to make the next trial smarter.
The “deciding” software is the brain of the machine, and most often it runs on a method called Bayesian optimization. We will watch it work, step by step, a little later. For now, the one-sentence version is this: it builds a running best-guess of how the experiment behaves, keeps track of where it is most uncertain, and uses both to choose a next experiment that is likely to teach it the most. It is a way of searching a vast space of possibilities without testing every point — which matters, because the spaces are astronomically large.
04 · The Distinction
Why this is not just automation with a faster robot
Conventional automation follows a list a human wrote (an open loop). A self-driving lab writes the list itself, choosing each experiment from the results of the last (a closed loop) using a decision engine, usually Bayesian optimization. The leap is not speed — it is judgment.
It is easy to assume a self-driving laboratory is simply a conventional automated lab with better robots. The two even share the same DMTA cycle. But there is a clean line between them, and it falls on a single question: who chooses the next experiment?
A conventional automated laboratory is an open loop. A human scientist plans a batch of experiments — say, a grid of two hundred combinations — and the machine executes that plan exactly, quickly, and tirelessly. When the batch finishes, the data comes back to the human, who studies it and designs the next batch. The automation is real and valuable, but the intelligence sits outside the machine. The robot never decides anything; it follows a script.
A self-driving laboratory is a closed loop. After each experiment, its software analyzes the fresh result and chooses the next experiment in light of it — with no human in that inner cycle. The loop is closed because the output of analysis becomes the input to design, automatically. The system is, in the field’s language, decision-driven rather than merely script-driven.
This distinction has a consequence that is larger than convenience. Researchers draw a line between hardware autonomy — robots that physically do tasks — and software autonomy — programs that decide which tasks are worth doing. A robot can pour liquids faster than a human, but it almost never does something a skilled human physically could not. The transformative part is the software autonomy, because the progress of science is shaped less by how fast experiments run and more by which experiments get chosen. A self-driving laboratory is significant precisely because it automates the choosing.
That is why practitioners describe SDLs as a structural reorganization rather than an efficiency upgrade. When a machine selects its own experiments, the rate-limiting step of discovery is no longer human attention. A small team can pursue many more lines of inquiry at once, and the searchable space of materials or molecules opens up in a way that a hand-planned grid could never match. The change is not “the same science, faster.” It is a different division of labor between human judgment and machine judgment.
05 · Plain Language
A short glossary, before we go further
A handful of terms recur throughout this field. None is as forbidding as it sounds.
- Self-driving laboratory (SDL)
- A system that joins automated experimental hardware with decision-making software in a closed loop, so it can design, run, measure, and interpret experiments — and choose the next one — with little or no human input.
- Closed loop
- A cycle in which the output of one step automatically becomes the input to the next. Here: the result of an experiment directly shapes the choice of the following experiment, without a person in between.
- Open loop
- A cycle that does not feed back on its own. A machine runs a fixed plan; any adjustment requires a human to step in and revise the plan.
- DMTA cycle
- Design, Make, Test, Analyze — the four repeating stages of an experimental loop, essentially the scientific method drawn as a circle.
- Bayesian optimization
- A search method that keeps a running best-guess of how a system behaves, tracks where it is most uncertain, and uses both to pick the next experiment likely to teach it the most.
- Active learning
- The broad idea of a model choosing which data to gather next, rather than being handed a fixed dataset — letting the system request the experiments it most needs.
- Exploration vs. exploitation
- The core trade-off in any smart search: test where results already look promising (exploit), or test where you know least, in case something better is hiding (explore).
- High-throughput experimentation
- Running very many experiments in parallel or rapid succession. It supplies raw speed, but on its own it does not decide what to run — that is the difference from an SDL.
- Materials Acceleration Platform (MAP)
- A name often used for self-driving laboratories aimed specifically at discovering new materials, frequently the most autonomous class of such systems.
06 · Worked Example
Watching the loop think: finding the strongest recipe
Abstract definitions only go so far. Let us run a deliberately simple problem through the loop, by hand, so the “deciding” stops being mysterious.
Suppose we want the hardest possible version of a new material. We can change one thing: the temperature at which we bake it, anywhere from 20 °C to 100 °C. Each baked sample gets a hardness score from 0 to 100. There is some best temperature that gives the hardest result, but we do not know it, and every test costs time and material — so we want to find it in as few tries as possible.
A conventional automated lab might brute-force this: bake a sample at every 5 degrees and compare. That is 17 experiments, run on a fixed script. A self-driving lab does something cleverer. Watch.
Start with a few spread-out guesses. The system bakes three samples to get its bearings: 40 °C → 35, 70 °C → 62, 95 °C → 48. Already a shape is forming — the middle looks better than the edges.
Draw a best-guess curve, and a band of doubt. The software fits a smooth curve through those three points (its current belief about hardness) and, just as importantly, marks where it is least sure — the wide gaps it has not sampled, such as the stretch between 70 °C and 95 °C.
Balance exploring and exploiting. It now weighs two instincts: test near 70 °C, where things already look good (exploit), or test the uncertain gap, in case a hidden peak sits there (explore). It picks 80 °C as a smart compromise — close to the good region, but unmeasured.
Run it, then update. The result comes back: 80 °C → 71. A new best. The curve redraws itself, now peaking somewhere just below 80 °C, and the band of doubt around that region shrinks.
Close in. Guided by the updated curve, it tries 77 °C → 73, then 75 °C → 72. The gains are flattening — a sign it is near the top. It concludes the best recipe is about 77 °C, reached in five experiments instead of seventeen.
That is the whole secret, in miniature. Now imagine the recipe has not one knob but ten — temperature, three ingredient ratios, mixing speed, curing time, and more. A brute-force grid would explode into millions of experiments; no lab could run them. The closed loop, by choosing each next experiment intelligently, can find an excellent recipe after a few hundred. Multiply that efficiency across an entire field and you begin to see why the people who build these systems talk about compressing work that once took months into days. Reports in the research literature describe acceleration on the order of tens to a thousandfold in the best cases, alongside large reductions in cost and waste — though, as we will see, those headline numbers deserve a careful reading.
07 · The World as It Is
Out of the demo, into the messy real
Self-driving laboratories have moved from proof-of-concept to working infrastructure in several fields. In materials science, systems hunt for better battery components, solar absorbers, and catalysts. In drug discovery, autonomous platforms now optimize how a compound is formulated and screen candidates around the clock; companies have built modular, closed-loop labs that retrain their models on each round of results. Open-source decision engines, shared on public code repositories, let academic and industrial labs plug an experiment planner into their own robots — a sign the field is standardizing rather than reinventing itself each time.
The most cinematic public demonstration came from Berkeley’s A-Lab in 2023. Running largely unattended for seventeen days, it set out to make a list of inorganic compounds that computer models had predicted should be stable but that no one had yet synthesized. It planned recipes by mining the scientific literature with language models, ran them robotically, read the results with pattern-recognition software, and adjusted its approach as it went — performing on the order of twenty experiments a day. The team reported synthesizing dozens of genuinely new compounds in that window, a pace no human lab could match.
And then something instructive happened — instructive precisely because it complicates the triumphant version of the story. After publication, an outside materials chemist scrutinized the results and questioned whether the automated analysis had correctly confirmed that the target compounds were really made. A public debate followed; the lab’s leaders defended the system’s purpose while conceding that a careful human could refine the same measurements to a higher standard. In time, the journal issued a correction, and the headline tally of new compounds was revised downward. The point of the autonomous lab, its builders argued, was to show what such a system can do on its own — not to claim it already matches the best human expert.
It was our objective to show what an autonomous laboratory can achieve.
— Prof. Gerbrand Ceder, A-Lab, responding to critics (2023)That episode is not a scandal; it is the field growing up in public. It draws the real boundary of today’s self-driving laboratories: they are extraordinary at searching a space of possibilities quickly, and still imperfect at the final, judgment-laden act of verifying what they found. A human scientist reading a measurement brings context, suspicion, and the willingness to distrust a clean-looking result. Teaching a machine that kind of skepticism is much harder than teaching it to pour, heat, and measure.
08 · Stakes
When the instrument also reasons, who is accountable?
Every technology that automates a human act also relocates a human responsibility, and a laboratory that designs its own experiments relocates several at once. These questions are not an appendix to the science; they are braided through it.
Consider reproducibility — the bedrock of science. In one sense, self-driving labs are a gift to it: a machine records every parameter of every run in exhaustive, machine-readable detail, far more faithfully than a tired researcher scribbling in a notebook. Adam’s creators argued years ago that formalizing experiments this way makes science more reusable and trustworthy. But the A-Lab episode shows the other edge of the blade. When an automated pipeline interprets its own results, an error in that interpretation can be reproduced perfectly and confidently, at scale. Precision is not the same as correctness, and a closed loop can be precisely wrong.
Then there is judgment itself. A Bayesian optimizer chases whatever goal it is given — maximize hardness, minimize cost. It does not pause to ask whether the goal is the right one, or whether a promising result is too good to be true. Those are human acts. As experimental choice moves into software, the values embedded in that software — what it optimizes for, what it ignores, when it flags doubt — become quiet but consequential decisions made by whoever wrote it.
There are quieter costs, too. If junior scientists spend their early years supervising autonomous platforms rather than performing experiments by hand, a kind of tacit knowledge — the feel for when something is off — may thin out across a generation. And these systems are expensive, which risks concentrating the fastest discovery in the best-funded institutions, widening rather than closing the gap between them and everyone else. Some teams are working deliberately against this, building low-cost “frugal” platforms to keep autonomous experimentation within reach of smaller labs. Whether that openness wins out is a choice the field is making now, not a foregone conclusion.
None of this argues against self-driving laboratories. It argues for holding two facts at once: the loop can be a powerful engine of discovery, and the loop has no conscience. The accountability stays exactly where it always was — with the people who build the system, set its goals, and decide how much to trust what it returns.
09 · What Remains Open
The hard problems the demos politely skip
It is tempting, watching a robot discover a material overnight, to imagine human scientists are nearly obsolete. The honest view is the opposite: the systems work best precisely where the problem is already well-defined, and they struggle exactly where science is most itself.
The first open problem is generality. Most self-driving labs are exquisitely tuned to one narrow task — these films, those reactions, that class of compound. Move the same machine to a genuinely different chemistry and much of its hardware and software must be rebuilt. A robot that can flexibly improvise across the messy variety of real experimental work, the way a graduate student does, remains far off.
The second is the unforeseen. A closed loop optimizes within the space of possibilities it was given. It is superb at finding the best point on a map; it is not designed to notice that the map is wrong, or that the most interesting thing is just off its edge. Many of science’s greatest moments came from an anomaly someone refused to dismiss — a contaminated plate, an unexpected glow. Whether a machine can be built to be productively surprised, rather than merely efficient, is among the field’s deepest unsolved questions.
The third is trust, and the A-Lab debate is its clearest illustration. For autonomous discovery to be believed, the verification stage must become as rigorous as the search stage — the machine must learn not just to find an answer but to doubt it. The likeliest near future is therefore not the empty laboratory but the collaborative one: humans setting the questions and the standards of proof, machines running the relentless search between, each doing what it does best.
10 · Reflection
A new shape for an old method
Strip away the robotics and the language models, and a self-driving laboratory is something quietly radical: the scientific method itself — observe, hypothesize, test, revise — lifted out of the human mind and printed into a machine that can run it without rest. Conventional automation gave us hands that never tire. The self-driving laboratory gives us a loop that chooses, and that small addition changes the arithmetic of discovery.
It is not, as its careful builders insist, a replacement for the scientist. It is a new division of labor, one that hands the tireless search to the machine and keeps the harder things — what is worth asking, what is worth believing — for the human. The empty room at two in the morning is not a vision of scientists made obsolete. It is a vision of their attention freed to point at the questions that still need a mind.
Which leaves the question worth carrying out of this article. If we build laboratories that can answer our questions faster than we can pose them, the bottleneck of science shifts from finding answers to choosing which questions deserve one — and that was always the part no machine could do for us. So: when the search becomes effortless, will we get any wiser about what is worth searching for?
— · Sources
Sources & further reading
- National Institute of Standards and Technology (NIST). Autonomous Laboratories. nist.gov/autonomous-laboratories
- Tom, G., et al. (2024). Self-Driving Laboratories for Chemistry and Materials Science. Chemical Reviews. doi.org/10.1021/acs.chemrev.4c00055
- Vriza, A., et al. (2025). Autonomous “self-driving” laboratories: a review of technology and policy. Royal Society Open Science. royalsocietypublishing.org · rsos.250646
- Review (2025). Self-driving laboratories with artificial intelligence: a process systems engineering perspective (DMTA cycle). Computers & Chemical Engineering / ScienceDirect. sciencedirect.com · S0098135425002698
- King, R. D., et al. (2009). The Automation of Science (Robot Scientist “Adam”). Science / Aberystwyth University announcement. aber.ac.uk · Robot Scientist
- University of Cambridge (2009). Robot scientist becomes first machine to discover new scientific knowledge. cam.ac.uk · robot scientist
- MacLeod, B. P., et al. (2020). Self-driving laboratory for accelerated discovery of thin-film materials (“Ada”). Science Advances. doi.org/10.1126/sciadv.aaz8867
- Szymanski, N. J., et al. (2023). An autonomous laboratory for the accelerated synthesis of inorganic materials (A-Lab). Nature. nature.com · s41586-023-06734-w
- Chemistry World (2024). New analysis raises doubts over autonomous lab’s materials discoveries. chemistryworld.com · A-Lab analysis
- R&D World (2025). Self-driving cars are hitting the streets — is your lab up next for automation? (industry deployments). rdworldonline.com · autonomous labs
Acceleration and cost-reduction figures (tens- to thousandfold faster; large cuts to cost and waste) are reported potentials drawn from the SDL research literature and depend heavily on the specific task; they should be read as demonstrated best cases, not guarantees. The A-Lab’s originally reported tally of new compounds was later revised following a published correction.