Self-Driving Labs: AI Takes Over the Experiment

Atinary's Boston lab opened in February 2026 with autonomous platforms that design, run, and analyze their own experiments continuously. A concurrent Nature paper asked whether robot labs could replace biologists. The honest answer: not yet, and not in the ways most people assume.
Self-Driving Labs: AI Takes Over the Experiment
Image: Picsum Photos (CC0)

Introduction

On February 11, 2026, a startup called Atinary quietly opened what it calls a Scientific Discovery Factory in Boston — two fully autonomous platforms that design their own experiments, execute them with robotic arms, analyze the results, and then immediately decide what to test next. No biologist in the loop. No waiting for a researcher to read a printout. The system runs 24 hours a day in a closed cycle called Design-Make-Test-Analyze-Learn.

At almost exactly the same time, a Nature news article asked whether self-driving robot labs could replace biologists entirely. The question landed differently than it might have two years ago. Concrete deployments are now common enough that it no longer sounds theoretical.

What a Self-Driving Lab Actually Does

The core idea is a closed feedback loop. An AI model proposes an experiment — which compound to synthesize, which cell condition to test, which protein variant to screen. Robotic hardware executes it. Sensors and instruments capture the output. The AI reads that output and uses it to update its model of the problem, then proposes the next experiment. This cycle runs continuously, without a human initiating each step.

Atinary’s Boston lab focuses on small-molecule synthesis and catalysis, targeting pharmaceutical R&D from early discovery through process development. Its technology stack integrates instruments from ABB, Agilent, Bruker, Chemspeed, and Mettler-Toledo, all coordinated through Atinary’s no-code AI platform. MIT Professor Stephen Buchwald — described by the company as the most cited chemist in the world for ten consecutive years — sits on the scientific advisory board.

The architectural pattern is not unique to Atinary. A 2025 review published in Royal Society Open Science maps these systems onto an autonomy scale from Level 1 (machine assistance with human control) to Level 5 (a fully autonomous AI researcher). The most capable current deployments sit at Level 4: systems that can independently form hypotheses, design and run experiments, and iterate — but still require humans to define the problem and set boundaries.

Biology’s Early Benchmarks

The two most cited biological self-driving labs are Adam and Eve, developed at Aberystwyth and Cambridge universities. Adam, demonstrated in 2009, autonomously identified three genes encoding orphan enzymes in yeast lysine biosynthesis — a discovery it made by generating hypotheses, running experiments, and updating its model without human intervention between cycles. Eve, in 2015, screened compounds against malaria targets and identified TNP-470 as a promising treatment lead.

More recent systems have pushed further. The SAMPLE platform, reported in 2024, improved enzyme thermostability by 12°C while searching less than 2% of the full combinatorial landscape — an efficiency that would be essentially impossible through exhaustive manual testing. Novartis’s MicroCycle system, also 2024, autonomously synthesizes compounds, performs purification, and runs biochemical assays in a single integrated workflow.

In cell biology, one published SDL ran 143 cell culture conditions across 111 days and achieved an 88% improvement in the production of retinal pigment epithelial cells. That kind of systematic exploration across a large experimental space is precisely where autonomous systems outperform human-led research: they do not get bored, they do not make transcription errors, and they do not skip conditions that seem unlikely.

The Speed and Scale Argument

Argonne National Laboratory’s Polybot system screened 90,000 material combinations in weeks — work that would have required months of manual effort. LabGenius’s EVA™ platform designed, produced, and characterized panels of up to 2,300 multispecific antibodies in six weeks. These are not theoretical projections. They are documented production runs.

A July 2025 paper in Nature Chemical Engineering from North Carolina State University showed another efficiency gain: switching from steady-state flow experiments to dynamic flow collection increased data output by 10x. The system captured one measurement every 0.5 seconds, versus a single data point every 10 seconds with prior methods. As Milad Abolhasani, the ALCOA Professor who led the work, described it: “It’s like switching from a single snapshot to a full movie of the reaction as it happens.” The system identified optimal material candidates on the first attempt after training.

McKinsey has estimated that comprehensive AI and automation in pharmaceutical R&D could reduce cycle times by more than 500 days and cut overall R&D costs by approximately 25%. Those numbers are contested — they depend heavily on how broadly you define “AI and automation” — but even conservative interpretations suggest that the speed advantage of SDLs is real and significant.

Where Humans Still Matter

The Nature article that framed the debate was careful about its conclusions. Philip Romero, a protein engineer, called autonomous lab systems “the future of biology,” but researchers quoted alongside him were quick to note the limitations. Current SDL hardware struggles with tasks requiring fine motor dexterity — moving fragile samples, handling irregular geometries, working with live organisms that behave unpredictably. And SDLs perform poorly on experiments without a clear optimization target: they need a measurable signal to optimize toward.

A telling example from the same article: a PhD student at Northwestern University spent four months and 40 experimental sessions manually testing 1,231 combinations to find a cell-free protein synthesis recipe at least six times cheaper than existing methods. A well-designed SDL could have covered that search space faster. But identifying that this was the right problem to work on, designing the initial protocol, and interpreting what the 6x cost reduction actually means for downstream applications — those steps still required a scientist.

The vortx.ch article on AI tools for academic research workflows in 2026 made a similar point about AI in research more broadly: the technology is most effective when it handles high-volume, well-defined subtasks while humans focus on problem selection and interpretation. Self-driving labs are a physical instantiation of exactly that division of labor.

IP, Safety, and What Gets Harder

Two underappreciated problems are starting to surface. The first is intellectual property. Under current patent law in every major jurisdiction, AI systems cannot be named as inventors. If an SDL discovers a novel compound or process autonomously, the resulting IP landscape is murky. Companies are navigating this by keeping humans nominally involved in the inventive step — a structural workaround that may not survive legal challenge as autonomy increases.

The second is safety. The Royal Society Open Science review recommends that every SDL deployment include mandatory human review of experimental plans before execution and a physical “kill switch” that supervisors can use to halt autonomous experiments immediately. For chemistry labs handling reactive or hazardous compounds, this is not optional. Defining the appropriate scope of autonomy — what the system can do without asking — turns out to be as much a governance problem as a technical one.

Conclusion

Self-driving laboratories have moved from proof-of-concept to production deployments faster than most researchers expected. The throughput numbers are real, the pharmaceutical partnerships are funded, and the debate about replacing biologists — while premature — reflects how seriously the field is taking the trajectory. The honest summary is that these systems are excellent at exploration within a well-defined space and poor at everything else that makes scientific research valuable. The labs that figure out how to pair autonomous experimentation with human problem-framing will run circles around those that don’t.

Further Reading

Don’t miss on GenAI tips!

We don’t spam! We are not selling your data. Read our privacy policy for more info.

Don’t miss on GenAI tips!

We don’t spam! We are not selling your data. Read our privacy policy for more info.

Share the Post:

Related Posts

Claude Code vs Copilot vs Devin: Which Agent Wins?

The AI coding assistant market has split into three tiers — inline completion, terminal agents, and fully autonomous cloud agents. Claude Code, GitHub Copilot, and Devin represent each tier clearly. Here is how their benchmark scores, pricing, and real-world performance stack up in March 2026.

Read More

Why 95% of Enterprise GenAI Pilots Never Reach Production

MIT’s 2025 GenAI Divide report found that 95% of enterprise AI pilots fail to deliver measurable P&L impact. The culprits aren’t the models — they’re organizational: poor data quality, misallocated budgets, and AI tools that never learn. Here’s what separates the 5% that make it to production.

Read More