Google Research has described its latest push to make AI agents at home in the physical wet lab. In a blog post, the team outlines three projects centered on VERID, a framework built on open-source PyLabRobot that acts as a semantic digital twin of a laboratory workcell. VERID combines synchronized multi-camera views with live 3D state tracking, allowing agents to check deck layouts, well volumes, and pipette-tip inventory before acting, and to simulate protocols in fast virtual dry runs before touching reagents.
VERID's safety checks also block physically impossible or logistically invalid commands. The team reports cases where the digital twin caught dye stratification caused by dispensing too close to the well floor, let an agent recover from a plate-reader failure without re-pipetting, and rejected an 8-channel aspiration that would not fit in a narrow trough, prompting the agent to replan. In a closed-loop self-play campaign, an agent iteratively designed and ran experiments to reconstruct blinded dye mixtures, improving accuracy more than fivefold by identifying its own optical and fluidic artifacts.
To move from liquid-handling demonstrations toward molecular design, Google Research used DNA aptamers as a sandbox for lab-in-the-loop discovery. Each design-test-analyze cycle takes 7-10 days, with candidate sequences screened against target proteins on magnetic beads and read out directly by sequencing. The long-term vision is that AI agents will not only ideate and analyze, but also run lab experiments, learn from the results, and self-improve, while remaining driven by human scientific agency.