Autonomous experimentation is real in narrow domains and growing fast — but “lights-out science” is still mostly marketing. Here's the honest map.
Self-driving labs are real where the problem is tightly scoped — automated synthesis, reaction optimisation, and now agent-designed biology. What's not real yet is a general, hands-off lab that sets its own research agenda. The market is scaling quickly, but the credible model in 2026 is a closed loop with a human in charge of the science.
A self-driving lab (SDL) closes the loop: an AI proposes an experiment, a robot runs it, instruments capture the data, and the AI updates its plan — repeat, with minimal human intervention [5]. The key word is loop: it's not just automation that repeats a script, it's automation that decides what to try next.
The evidence is concrete. Berkeley's A-Lab ran autonomous inorganic synthesis for 17 days [2]; an LLM agent (Coscientist) optimised real chemical reactions [1]; and in 2025 the Virtual Lab's AI agents designed nanobodies that were experimentally validated [4]. These work because the goal and success signal are crisp.
What marketing often implies — an autonomous lab that picks its own questions and needs no scientists — doesn't exist. Even the headline results were steered by humans and, in A-Lab's case, needed correction after outside scrutiny [3]. Autonomy still fails on open-ended judgment and ambiguous results.
The commercial signal is unambiguous: the laboratory-automation market was worth roughly $8–9 billion in 2025 and is projected to reach about $20–24 billion by the mid-2030s [6].
For most labs, the win isn't a megalab or a cloud facility — it's a self-driving bench you own: describe the protocol, the robot runs it vision-verified, and you keep the science. See the approach →
