Wiki · Frontier · concept
Self-driving lab
A lab where AI proposes experiments, robots run them, and results feed back — closing the loop with minimal human intervention.
A self-driving lab is a laboratory where an AI proposes experiments, robots run them, and the results feed straight back into the AI. Instead of a human designing each run, reading the result, and deciding what to try next, the whole design→build→test→learn cycle turns on its own — the machine picks the next experiment, executes it, measures the outcome, and updates its beliefs, with people stepping in only at the edges.
How it works
At the core is an active-learning loop. A model holds a hypothesis about the system — which material, molecule, or condition might work — and chooses the experiment expected to be most informative, not just the one most likely to succeed. A robotic platform then carries it out: dispensing reagents, heating a sample, running a synthesis. Instruments measure what happened, the result flows back, and the model revises its next pick. Each turn sharpens the model, so later experiments get smarter. Real systems include Berkeley’s A-Lab, which runs robotic inorganic-materials synthesis; Coscientist from Carnegie Mellon, which uses a language model to plan and execute chemistry; and mobile chemistry robots that roam a lab and operate standard equipment.
Why it matters (for bio × AI)
A self-driving lab compresses the experiment cycle from weeks to hours, and it closes the gap that in-silico predictions keep hitting: a model can propose thousands of candidates, but something still has to test them in the real world. When the wet lab is automated and folded into the loop, it stops being the bottleneck — prediction and validation run at the same tempo. This is the frontier of agentic science, where an AI doesn’t just analyze data but drives the apparatus that produces it.
The honest caveat
“Autonomous” is bounded. Humans still frame the problem, choose the chemistry, define what counts as success, and set the safety limits. These systems are impressive within a narrow, well-instrumented domain; they don’t yet invent their own questions. The loop runs itself — the scientist decides what the loop is for.
Related
See single-cell foundation model for the kind of predictive model whose hypotheses a self-driving lab could put to the test.