What makes a chemistry lab self-driving?

A useful way to read an autonomous-chemistry paper is to follow one decision from beginning to end. What information reached the system? What did it choose? Which measurement decided whether the choice worked? Our reading of three studies suggests that these questions deserve separate answers.

Choosing the next experiment

In an August 2026 paper, Ranković and colleagues present GOLLuM, which trains a language model together with a Gaussian process, a statistical predictor that also represents uncertainty. The authors evaluate 23 optimization tasks beginning with ten low-performing examples. Their method ranks first on average across the tested alternatives. These are benchmark results, not a demonstration that the method already controls a general-purpose laboratory. The authors also flag scaling costs for large campaigns and difficulties representing some complex structures as text. Read the GOLLuM study.

For a reader assessing a proposed deployment, we would ask for an additional record: every suggested experiment, every rejected suggestion, and the rule that determined the next measurement. That is an editorial test of inspectability, not a performance result from this paper.

Connecting choices to equipment

RoboChem-Flex, published in April, approaches the physical workflow through modular hardware and software. The authors test six case studies, including photocatalysis and biocatalysis, and support both autonomous and human-assisted configurations. Their approximately US$5,000 entry configuration uses existing shared analytical equipment and human involvement. It should therefore not be read as the price of an entirely independent analytical laboratory. Read the RoboChem-Flex study.

The procurement question we would put beside that figure is simple: which instruments, maintenance work and operator time are already available? A component price becomes more useful when the reader can see what sits outside it. We have not priced or assembled this platform.

Deciding what counts as a hit

An earlier study, published in November 2024, links mobile robots with synthesis equipment, liquid chromatography–mass spectrometry and benchtop nuclear magnetic resonance. Its heuristic decision-maker combines analytical evidence and checks reproducibility before advancing hits. Domain experts choose the chemistry in advance. This is an experimental demonstration of bounded autonomous decisions, not evidence that the robots independently chose the research programme. Read the mobile-robot study.

Our interpretation

We would evaluate autonomy as a set of demonstrated responsibilities: proposing work, executing it, interpreting a measurement and responding to an uncertain outcome. These papers examine different responsibilities; none of the comparisons above establishes that their components work together.

For the next generation of reports, our preferred evidence would include a run that stalls, an analysis that disagrees with another instrument, and a successful recovery. The instructive moment may be the one when the system decides that it does not yet know enough to continue.