The field is quietly abandoning “lights-out” for shared control — AI does the heavy planning and execution, a scientist keeps the judgment.
The most credible autonomous-lab systems in 2026 aren't fully hands-off — they're human-in-the-loop. AI handles the data-heavy planning, execution, and verification; an experienced scientist sets the goal, makes real-time judgment calls, and reviews the output. This isn't a limitation to apologise for — it's the design that actually works.
Early autonomous-lab hype pitched fully self-running science. In practice, the strongest results kept a person in the loop: the Virtual Lab's AI agents worked under a human researcher's high-level steering [1], and 2025 work on shared human–AI control explicitly keeps decisions with experienced researchers who are used to acting on limited data [2].
AI is strong at bounded execution and weak at open-ended judgment — choosing which question matters, spotting a result that's “off” in a way no rule anticipated, and deciding when to stop. Reviews of self-driving labs make the same point: autonomy accelerates the loop, but the scientific direction remains a human responsibility [3].
Shared control only works if you can trust each step. That's why vision-verified execution matters: when every action is confirmed and logged, a clean result means the steps actually happened — and the scientist can review with evidence, not faith.
The goal isn't to remove the scientist. It's to remove the busywork between the scientist and the answer.
On a self-driving bench, you describe the protocol in plain language, the robot executes it vision-verified, and you stay in charge of the science — reviewing, adjusting, deciding what's next. See the approach →
