VERID checks laboratory agent commands against a live physical workcell

Laboratory agents can generate code that is syntactically valid but physically impossible. Google's VERID work places a stateful workcell representation between an agent's planned procedure and the actual equipment. The twin tracks robot geometry, deck layout, well contents, reagents and tip inventory, simulates proposed steps, and uses camera observations and execution traces to reconcile its state after permitted actions.

In Google's reported demonstrations, simulated checks rejected an eight-channel pipette operation against an incompatible trough. Another sequence used video to recognize poor dye mixing associated with dispensing depth and supported a revised experiment; a failed plate-reader run could be handled without blindly repeating liquid transfers. These are controlled examples of grounding model output in externally checked physical constraints, not proof that the model is always right.

The safety property is only as strong as the simulation, its sensor inputs and the hardware interlock. A collision-free simulated move can still be scientifically useless, and an inaccurate state estimate can misclassify a hazard. Human approvals and biosecurity screening remain part of the reported workflow. No arbitrary, unattended wet-lab deployment is demonstrated.