Google Research unveils VERID, a digital twin to ground AI agents in wet-lab reality

Google Research published a framework on October 16 for connecting AI agents to wet-lab robots without letting them spill the reagents. It is called VERID, for VERified Infrastructure for Discovery Laboratories, and it is a live 3D digital twin of a lab workcell. It tracks robotic arm geometry, well volumes, reagent identities and pipette tip inventory. It is built on the open-source PyLabRobot framework.

The pitch is four capabilities: live state tracking, virtual dry runs that simulate protocols in seconds, hardware safety checks that block invalid commands, and synchronized video and action logs.

The numbers come from Google's own post. In a closed-loop campaign, an agent reconstructed blinded dye mixtures from camera images and improved accuracy more than fivefold. On DNA aptamers for α-thrombin, hit rates went from 0% in Round 1 to 26.0% and then 42.1% at 98.5% specificity. On myoglobin, agents chased raw binding, specificity fell from 75.0% to 19.1%, then recovered to 72.4% after they revised their own rules.

Context: Anthropic recently released a Model Hardware Standard to link agents to lab devices, so the plumbing race is on.

My take: the myoglobin result is the honest part. The agent fell for a sticky-binder artifact, and the wet-lab readout caught it. That is the point of grounding. What the material does not say is when anyone outside Google can use VERID.