---
title: "Self-driving labs: what they are, what's proven | Quantlix"
description: "A self-driving lab lets AI choose experiments and robots run them. Here is what has been shown since 2023, what was disputed, and what a factory can borrow."
canonical: "https://quantlix.com/en/blog/self-driving-labs-ai-experiments/"
last-updated: 2026-11-21
---

Source: https://quantlix.com/en/blog/self-driving-labs-ai-experiments/
Language: en
Description: A self-driving lab lets AI choose experiments and robots run them. Here is what has been shown since 2023, what was disputed, and what a factory can borrow.

Perspective Physical AI

# What is a self-driving lab, and does it work yet?

A self-driving lab lets AI choose experiments and robots run them. Here is what has been shown since 2023, what was disputed, and what a factory can borrow.

[Ali Muhammad](https://quantlix.com/en/about/leadership/ali-muhammad/)

Founder and Chief Executive Officer

Nov 21, 2026 (2026-11-21) 14 min read

![A laboratory workbench with a microscope and instruments, with the headline “When the lab runs itself”.](https://quantlix.com/_astro/insight-self-driving-labs-ai-experiments.xqxXQnT2_vIx2K.webp)

01 / Engineering notes

Field notes: Blog

**On this page**

The short answer

A self-driving lab is a laboratory where AI chooses each experiment, robots or automated instruments run it, and the measured result feeds back so the AI can choose again. Since 2023 these labs have run real experiments, but some claims were disputed, and people still set the questions and check the results.

Picture a laboratory at three in the morning, lights low and nobody in. A robot arm draws a few drops from one vial, adds them to another and slides the sample into an instrument. Minutes later a program reads the measurement, compares it with everything tried so far and picks the next experiment. The arm reaches for a fresh vial.

That scene is an illustration, but it is close to what a **self-driving lab** does: AI chooses the next experiment, robots or automated instruments carry it out, and the result goes straight back so the AI can choose again. Below we set out what has been shown since 2023 (with dates and sources), where a claim was disputed, and which part of the idea a factory can borrow without a single robot arm.

## How does a self-driving lab work?

A self-driving lab repeats one four-step loop: choose an experiment, run it, measure the result, update what the AI knows. Then it starts again.

1. **Choose.** The AI picks the experiment it expects to teach it the most, given everything tried so far.
2. **Run.** A robot or an automated instrument carries it out.
3. **Measure.** The result is recorded automatically, in the same way every time.
4. **Update.** The AI folds the result into its understanding and picks the next experiment.

The clever part is the choosing. A person working by hand usually changes one thing, looks at what happens, then changes the next. A machine does not have to go in order. Scientists call this approach **active learning**, which simply means letting the results so far decide what to try next.

The robots are the less glamorous half, but they matter. A robot can repeat a step the same way and log it each time, which makes results easier to check and repeat.

## What did the first self-driving labs show in 2023?

Two results in late 2023, both described in the journal *Nature*, showed that AI-guided systems could run real laboratory work: Berkeley Lab’s A-Lab and Carnegie Mellon’s Coscientist.

On 29 November 2023, [Lawrence Berkeley National Laboratory](https://www.eurekalert.org/news-releases/1009276) announced the A-Lab, a facility where AI guides robots in making new solid materials. The lab reported that over 17 days of independent operation it produced 41 new compounds out of an attempted 58. Its AI wrote the synthesis recipes by combing through scientific papers, then adjusted them as results came in. Those figures are the lab’s own.

On 20 December 2023, [Carnegie Mellon University](https://www.cmu.edu/news/stories/archives/2023/december/cmu-designed-artificially-intelligent-coscientist-automates-scientific-discovery) described Coscientist, a system built on large language models including OpenAI’s GPT-4 and Anthropic’s Claude. Given a plain-language request, it searched the web and equipment manuals, chose a plan and sent instructions to automated instruments, including a remotely operated lab run by Emerald Cloud Lab. The university’s release called it the first time a non-organic intelligent system had designed, planned and executed a chemistry experiment. The chemistry it ran was the synthesis of compounds already known, so it demonstrated the method rather than a new discovery.

The Carnegie Mellon team also tested whether the system could be pushed into making hazardous chemicals. Its lead, Gabe Gomes, said: “we have a responsibility to acknowledge what could go wrong and provide solutions and fail-safes.”

## Why was the A-Lab’s result disputed?

In March 2024, a group of chemists including Leslie Schoop and Robert Palgrave published a peer-reviewed critique in [*PRX Energy*](https://journals.aps.org/prxenergy/abstract/10.1103/PRXEnergy.3.011002) concluding that the A-Lab work did not establish any new materials. Their words: “no new materials have been discovered in that work.”

They reviewed the reported products and named four recurring shortfalls. Two stand out. First, they wrote that automated analysis of powder X-ray diffraction data (the pattern an X-ray makes when it passes through a powder, used to identify what the material is) “is not yet reliable.” Second, the predictions neglected that different elements can share the same position in a crystal. The authors estimate that “two thirds of the claimed successful materials” are probably known materials in a disordered form.

A self-driving lab is only as trustworthy as the way it measures and judges its own results.

That is the lesson of the episode, and it is one every factory engineer knows. Speed without careful checking produces wrong answers faster. A claim made in the open and challenged in the open is also science working as intended. We read the critique’s abstract, not the full paper, so treat this as a summary of the dispute and not a ruling.

## What has 2026 added?

In 2026, AI-generated ideas reached lab tests in several fields, one AI-linked drug entered a late-stage trial, and projects that close the whole loop were announced. Four examples, each with who is saying it.

**Google’s Co-Scientist.** On 19 May 2026, [Google DeepMind](https://deepmind.google/blog/co-scientist-a-multi-agent-ai-partner-to-accelerate-research/) said its latest Co-Scientist research had been published in *Nature*. Co-Scientist is a multi-agent system built with Gemini: several AI agents generate, critique, rank and refine research hypotheses. Google’s page describes collaborators testing its suggestions. For liver fibrosis (scarring), Google says the system pointed to overlooked drug candidates and that one “successfully blocked 91% of a scarring-linked response in lab tests”, in work Google says was published in *Advanced Science*. That is a lab result, not a patient result. We could not open the *Nature* paper, so we do not say which findings it contains.

**An AI-linked medicine in a phase 3 trial.** On 10 September 2026, [Insilico Medicine](https://insilico.com/news/isn1009261-insilico-medicine-doses-first-patient-genesis-ipf-3) announced the first patient dosed in GENESIS-IPF-3, a phase 3 trial (the large final testing stage) of rentosertib for idiopathic pulmonary fibrosis, a lung disease. The trial plans to enroll 320 participants at 47 centers in China. The company says AI identified the biological target and that the molecule was developed using generative AI. The drug is investigational and not approved by any regulator, and a phase 3 trial can still fail. The company says an earlier phase 2a study was published in *Nature Medicine* in 2025.

**A closed-loop lab for yeast.** In September 2026, Chalmers University of Technology in Sweden reported a “closed-loop AI laboratory” working on brewer’s yeast, described in the *Journal of the Royal Society Interface* ([DOI 10.1098/rsif.2026.0043](https://dx.doi.org/10.1098/rsif.2026.0043)). [Phys.org’s report](https://phys.org/news/2026-09-ai-scientist-autonomously-generates-validates.html) of 30 September says the system can pick biological questions, propose experiments, assess the results and update its understanding with “minimal human intervention”. It also says human scientists remain essential for defining research priorities. The report names no specific discovery, and we could not open the paper, so we cannot say what the system found.

**SPIRE at SLAC.** On 8 October 2026, [SLAC National Accelerator Laboratory](https://www6.slac.stanford.edu/news/2026-10-08-slac-leads-new-project-build-self-driving-laboratory-future) described SPIRE, a platform for self-driving experiments at U.S. Department of Energy photon science facilities, covering sample handling, beam control and real-time data analysis. Parts already run. Robotic arms mount and center protein crystals, machine learning reconstructs the X-ray beam from detector images, and in a Department of Energy demonstration AI agents ran a battery-imaging workflow with no human involved. A second beamline now has agents monitoring the instrument for weeks while a person stays at the controls. SLAC contrasts the old way of tuning a beam, which “consumed hours of highly competitive beamtime”, with a beam picture that now reaches scientists “in about five minutes”. That is not a like-for-like comparison, and SLAC gives no timeline for the full system.

## How strong is the evidence for each claim?

Most of these claims come from the people who made them, and only some have been checked by outsiders. The table separates what each source shows from what it does not.

| Project | Date | Who reports it | Kind of evidence | What it does not show |
| --- | --- | --- | --- | --- |
| Coscientist | Dec 2023 | Carnegie Mellon | *Nature* paper; details here come from the university’s release | A new discovery (it made known compounds) |
| A-Lab | Nov 2023 | Berkeley Lab | *Nature* paper, then disputed in a peer-reviewed critique (Mar 2024) | That the compounds were new |
| Co-Scientist | May 2026 | Google DeepMind | *Nature* paper (we did not open it); lab tests described by Google | Patient benefit; any overall success rate (Google’s page gives none) |
| Rentosertib | Sep 2026 | Insilico Medicine | Company announcement of a trial start | Approval, or the phase 3 result |
| Yeast lab | Sep 2026 | Chalmers, via Phys.org | Journal paper (we did not open it) and a news report | Any specific finding (none named) |
| SPIRE | Oct 2026 | SLAC | Project announcement | A finished system, or a measured speed-up from SPIRE itself |

Fairly summed up: machines have proposed ideas that were then tested at the bench, run experiments with little steering and sped up routine steps. In the sources we read, we found no case where one found something important that its makers did not expect and outsiders then confirmed. That is the result to watch for.

## What can a self-driving lab not do yet?

A self-driving lab cannot choose the question, speed up slow experiments or reliably judge its own work. Four limits, plus safety.

- **It cannot choose the question.** SLAC says researchers dictate the scientific goals, and the Chalmers report says human scientists remain essential for defining priorities. In our reading, a machine will run a thousand experiments on the wrong question as readily as on the right one.
- **It slows down when results are slow.** The loop shines when a test finishes in minutes or hours. When a result takes months, as with a new crop variety or a medicine in human trials, the loop runs at human speed again.
- **It can fool itself.** The A-Lab dispute turned on automated analysis. If the way of judging results is flawed, the machine reports the wrong thing with confidence.
- **It is only as good as its records.** A system learns from everything it has tried, so patchy records mean patchy learning.

Safety stays with people and hard-wired systems. SLAC says its hard-wired safety systems will remain in place, with AI operating inside those boundaries, and lists guardrails as something the team is still examining.

## What can a factory borrow from a self-driving lab?

A factory can borrow the loop (try, measure, adjust, try again) without any robots. What it needs is consistent measurement and complete records.

Plants already do a version of this by hand. Engineers use a structured method called **design of experiments** to change several settings at once and work out which ones really matter. What AI adds is the chance to keep that loop running on live data and to suggest the next adjustment from everything learned so far. A lab hunts for new knowledge and a plant hunts for steadier quality or lower waste, but the machinery underneath is close kin. Our guides to [testing a change in a digital twin first](https://quantlix.com/en/blog/digital-twins-food-manufacturing-supply-chain/) and to [what physical AI is](https://quantlix.com/en/blog/what-is-physical-ai-robots-in-factories/) cover the neighboring pieces.

Here is an illustration, not a real plant or a result. A textile dye house often misses a customer’s shade on the first try. A batch comes out too dark or too dull and is re-dyed, which costs water, energy, chemicals and a day. The dyeing masters are very good, and much of what they know lives in their heads: how a little more salt or a few degrees will shift the color on this fabric.

Now imagine the dye house writes down every batch the same way: recipe, fabric, water temperature, time in the machine, and the final color measured by a spectrophotometer (an instrument that reads color as numbers). After enough batches there is a real record. A simple model trained on it could suggest a starting recipe for each new order, with a note of how confident it is. The dyeing master stays in charge, accepting the suggestion, adjusting it or ignoring it. Each batch adds to the record.

That is a self-driving lab with the robot arm left out and a skilled person kept firmly in the loop. The same pattern could fit a fryer on a snack line or the hold time in a dairy.

## What should you get right before borrowing the idea?

Fix the measurement first, then automate. Four mistakes are common, and each has a simple guard.

- **Automating before you measure.** If results are judged by eye, or recorded differently by each shift, a model will learn the noise.
- **Chasing one number and breaking another.** A loop told only to cut energy may let quality slip. Name everything that matters and set limits it must never cross.
- **Letting it touch the machine without guard rails.** Any system that suggests settings should work inside limits a qualified person has set, leave the machine’s own safety functions untouched, and have a person approve changes until the team trusts it.
- **Throwing away the failures.** The batches that went wrong are often the most valuable records you have.

Where a process is validated or regulated, as in food or medicines, a changed setting still goes through your own change control and quality procedures. This article is not regulatory advice. If your loop lives in a laboratory, the same record-keeping applies, and our [life sciences page](https://quantlix.com/en/industries/life-sciences/) describes how we approach research and lab data.

## Where to start

You can begin this month with a notebook or a spreadsheet and no new equipment.

1. **Pick one test or adjustment** your team repeats often.
2. **Decide how its result is measured,** then measure it the same way every time: same instrument, same method, same unit.
3. **Record every attempt:** what was changed, what happened, who did it and when.
4. **Keep the failures,** and note why you think they failed.
5. **Review the record together** after a while. You will probably spot patterns nobody had noticed, and you will have the foundation any future AI system would need.

If you would like a second pair of eyes, tell us about one machine or one line and we will reply with a short note on what we would measure first. You can [reach us here](https://quantlix.com/en/contact/). Our guide to [starting with AI in a factory](https://quantlix.com/en/blog/how-to-start-with-ai-in-a-factory/) covers the wider first steps.

## Key takeaways

1. A self-driving lab repeats one loop: the AI chooses an experiment, a robot runs it, the result is measured, and the AI updates before choosing again.
2. In March 2024 a peer-reviewed critique disputed the A-Lab's claim of new materials, showing that how results are judged matters as much as speed.
3. In 2026 Google, Insilico Medicine, Chalmers and SLAC each reported progress. Most figures are the organizations' own, and none of the sources we read shows a surprise discovery.
4. A plant can borrow the loop without robots: measure the same way every time, record every attempt including failures, and keep a person approving changes.

## Questions people ask

### What is a self-driving lab?

A self-driving lab is a laboratory where an AI system picks which experiment to run next, robots or automated instruments carry it out, and the measured result goes straight back to the AI. The loop repeats without a person directing every step. Scientists still set the goals and judge what the results mean.

### Do self-driving labs work yet?

For narrow, repeatable experiments, yes. Carnegie Mellon's Coscientist ran chemistry through automated instruments in 2023, and SLAC reports AI agents running a battery-imaging workflow with no human involved. What is not yet shown is a surprising discovery confirmed by outsiders, and some early claims were disputed.

### What was the controversy about the A-Lab?

In November 2023 Berkeley Lab reported that its A-Lab made 41 new compounds out of 58 attempted in 17 days. In March 2024 other chemists argued in PRX Energy that the automated analysis was unreliable and that no new materials were established. The case shows that checking results matters as much as speed.

### Has AI helped design a medicine that reached human trials?

Insilico Medicine says AI identified the target and helped develop rentosertib, a drug for idiopathic pulmonary fibrosis. In September 2026 the company announced the first patient dosed in a phase 3 trial planned for 320 participants in China. The drug is investigational and not approved by any regulator, and the trial could still fail.

### What does a self-driving lab have to do with manufacturing?

The method is the same. A factory that measures its process, adjusts a setting, and measures again is running the same loop. The difference is the goal: a lab hunts for new knowledge, while a plant hunts for steadier quality or lower waste. Both depend on consistent measurement and complete records.

### Can a small plant use this idea without robots?

Yes. The two essentials are measuring a result the same way every time and recording every attempt, including the failures. With those in place, a simple model can suggest the next setting to try while a skilled person decides whether to use it. Robots can come later, if they ever make sense.

## Sources

1. [Google DeepMind, Co-Scientist: a multi-agent AI partner to accelerate research (19 May 2026)](https://deepmind.google/blog/co-scientist-a-multi-agent-ai-partner-to-accelerate-research/)
2. [Insilico Medicine, news release on the first patient dosed in the GENESIS-IPF-3 trial (10 September 2026)](https://insilico.com/news/isn1009261-insilico-medicine-doses-first-patient-genesis-ipf-3)
3. [SLAC National Accelerator Laboratory, news release on the SPIRE self-driving laboratory project (8 October 2026)](https://www6.slac.stanford.edu/news/2026-10-08-slac-leads-new-project-build-self-driving-laboratory-future)
4. [Phys.org, report on the Chalmers University of Technology AI scientist (30 September 2026)](https://phys.org/news/2026-09-ai-scientist-autonomously-generates-validates.html)
5. [Carnegie Mellon University, news story on Coscientist (20 December 2023)](https://www.cmu.edu/news/stories/archives/2023/december/cmu-designed-artificially-intelligent-coscientist-automates-scientific-discovery)
6. [Lawrence Berkeley National Laboratory, A-Lab news release, republished by EurekAlert (29 November 2023)](https://www.eurekalert.org/news-releases/1009276)
7. [Leeman et al., Challenges in High-Throughput Inorganic Materials Prediction and Autonomous Synthesis, PRX Energy 3, 011002 (7 March 2024), abstract page](https://journals.aps.org/prxenergy/abstract/10.1103/PRXEnergy.3.011002)

Sources checked on Nov 21, 2026.

- Perspective
- Physical AI

Written by

## Ali Muhammad

Founder and Chief Executive Officer · Quantlix

[See full profile](https://quantlix.com/en/about/leadership/ali-muhammad/)

[Talk to Ali](https://quantlix.com/en/contact/)

[More blog](https://quantlix.com/en/blog/)

## Keep reading

[More blog](https://quantlix.com/en/blog/)

- ![Orange industrial robots along an assembly line, with the headline “What a robot needs before it can work a shift”.](https://quantlix.com/_astro/insight-what-is-physical-ai-robots-in-factories.TGUip7Qv_1CPJjd.webp)

  Perspective

  Physical AI Nov 21, 2026 (2026-11-21)

  ### [What is Physical AI, and are robots ready for factory work?](https://quantlix.com/en/blog/what-is-physical-ai-robots-in-factories/)

  Physical AI is AI that sees, decides and moves things. See what factory robots do today, what the leading labs report, and how to judge a robot for your line.

  15 min read Read
- ![A close-up of a circuit board, with the headline “Test the change before you make it”.](https://quantlix.com/_astro/insight-digital-twins-food-manufacturing-supply-chain.BB-QJoEq_Dta0c.webp)

  Playbook

  Food manufacturing Nov 21, 2026 (2026-11-21)

  ### [What is a digital twin, and how does a food company use one?](https://quantlix.com/en/blog/digital-twins-food-manufacturing-supply-chain/)

  A digital twin is a working copy of a line, plant or supply chain kept in step with real data. See what food makers ask it, who uses one and how to start small.

  15 min read Read
- ![A circular conveyor in a large, bright factory hall, with the headline “The model is the easy part”.](https://quantlix.com/_astro/insight-start-ai-in-a-factory.BsmMIis5_Z1XFbFd.webp)

  Playbook

  Industrial AI Nov 21, 2026 (2026-11-21)

  ### [How should a factory start with AI?](https://quantlix.com/en/blog/how-to-start-with-ai-in-a-factory/)

  Start AI in a factory with one line and one costly problem. Before you buy, check five things: machine data, wrong answers, an owner, internet loss and access.

  15 min read Read

## Need the version tailored to your platform, risk model and roadmap?

Bring us the context. We will turn the patterns in this library into a pragmatic delivery path your team can use.

[Plan a workshop](https://quantlix.com/en/contact/)
