How Do Self-Driving Laboratories Choose Their Next Experiment?

A self-driving laboratory uses the results of one experiment to help select the next. Automation moves equipment; feedback changes the experimental plan. These are different capabilities, and neither requires a robot shaped like a person.

What closes the loop?

The researcher defines an objective, allowable materials, operating limits and a budget. A planning algorithm then selects a candidate experiment. Instruments execute the procedure, measurements return data, and an analysis step updates what the planner knows. This design-run-measure-decide cycle repeats until a stopping condition is reached.

In a hypothetical mixture experiment, the objective might be a specified optical response. The system could adjust ingredient ratios after each measurement. A higher reading is useful only if it measures the intended property.

How does software reach the instruments?

Instrument drivers translate software requests into device-specific commands. Task queues coordinate execution; permissions restrict access. NIST's instrument-control documentation illustrates this separation between client, server, driver and queue.

Results also need sample identities and metadata. A measurement detached from the preparation that produced it cannot reliably guide another experiment. NIST's data-system documentation describes shared sample identifiers across instrument servers.

Why can a better score be misleading?

A sensor can drift, a sample can be contaminated, or a candidate can interfere with the assay. Repeating a misleading measurement only produces more misleading data. Counter-screens test alternative explanations, while an orthogonal assay uses a different measurement approach. The Assay Guidance Manual explains why confirmation belongs in a screening workflow.

Where do people remain responsible?

Researchers define acceptable risks, validate methods and decide whether a result is scientifically meaningful. A digital twin can support planning, but only within its modeled assumptions. Human participation need not disappear when routine decisions are delegated.

As the review of self-driving laboratories emphasizes, autonomy spans different levels. A successful bounded workflow does not demonstrate a laboratory that can investigate any question safely on its own.