Google Research introduces a digital replica of a laboratory to check an AI agent’s commands

Google Research introduces a digital replica of a laboratory to check an AI agent’s commands

On 16 October, Google Research introduced VERID, a digital model of a laboratory setup. Software uses the model to check commands for a liquid-handling robot before it begins an experiment.

A laboratory robot can receive the correct sequence of steps and still collide with labware or try to draw liquid from an empty well. In the VERID technical report, eight pipette tips spanned at least 72,0 millimetres, while the tray was 71,2 millimetres long. A check in the digital replica blocked the operation, and the agent revised its plan.

VERID maintains a three-dimensional model that tracks the robot’s position, labware geometry, liquid volumes in wells, reagents, and the supply of replacement pipette tips. The agent queries the setup’s current state and tests a sequence of actions in the virtual replica. Before a command reaches the robot, the system checks it again against the physical setup. Video and logs allow each action to be matched to what happened in the laboratory workcell.

Model Hardware Standard gives an agent a common interface for reading instruments and adjusting permitted settings. VERID checks whether a specific command can be carried out with the robot and labware in their current positions.

The model also helps with reviewing completed experiments. In one case, synchronized video and logs showed that dye was forming layers because the pipette tip was positioned near the bottom of the well. After a communication failure with the plate reader, the saved state allowed the robot to resume where it had stopped.

Alongside VERID, Google describes two independent studies on molecular discovery. In an experiment with DNA aptamers, short DNA strands capable of binding to a target protein, agents proposed sequences over three rounds, with sequencing results guiding selection in the next round. By the third round, the proportion of new sequences with statistically supported target binding reached 33,5% for myoglobin and 42,1% for thrombin after correction for binding to the beads.

In a separate study, the computational search focused on a local region of the target protein. Four selected variants were tested using yeast display, in which a protein is presented on the surface of a yeast cell and its binding is measured. The estimated binding affinity for the target was 3,0–9,1 nM.

In PRAXIS, agents used laboratory measurements to select the next batch of protein variants. VERID adds an earlier step: checking whether the robot can execute the plan before measurements begin. This takes an agent’s proposal from a digital plan through robot movement to a measured result.