The same foundation-model wave reshaping language and robotics is arriving in the lab — as agents that reason about protocols and perceive a real bench.
Foundation models — large models trained broadly then adapted to many tasks — are reaching the bench in two forms: language-and-reasoning agents that turn a described protocol into an executable plan, and perception models that let a robot see a real deck. Together they're what makes “speak science, not robot” possible on the messy, variable bench that resisted automation for decades.
Reasoning/agent models (LLMs with tools) interpret a plain-language protocol into steps, choose methods, and coordinate a run — the capability behind agent systems that have planned and executed real experiments [1] [2]. Perception models give a robot vision: detecting plates, tips, liquids, and misalignments so it can act on a real, unstructured deck instead of assuming a perfect one.
Factories are structured; a lab bench is not — every lab is laid out differently, labware shifts, reagents look alike, and protocols change weekly. That variability is why bench work resisted automation for forty years. Broadly-trained perception plus reasoning is what makes an unstructured bench tractable: the system adapts to the world as it is, rather than demanding the world be perfect [3].
Fixed automation gets faster at one task; a foundation-model approach gets broader — because perception and language are general, a new assay is a new description, not a new integration. That's the shift from a machine that does one thing to an operator that handles many.
Automation repeats. Foundation-model physical AI perceives, interprets, and adapts.
The practical form is a rail-mounted robot that reads your protocol, sees your deck, and runs the workflow vision-verified — on the instruments you already own. See how that works →
