Skip to content

SYSTEM Cited by 1 source

VERID

What it is

VERID (VERified Infrastructure for Discovery Laboratories) is Google Research's semantic digital twin of a wet-lab workcell: a "live, stateful 3D volumetric entity tree that tracks everything from robotic arm geometry to individual well volumes, reagent identities, and pipette tip inventory." It is the systems scaffolding that lets AI agents drive physical lab experiments safely — closing the gap between an agent's high-level reasoning and the error-prone physical reality of liquid-handling robots. It is built on top of the open-source PyLabRobot framework.

VERID exists because, for physical actuation, "the safety of a command depends on the live physical context as well as the code." An agent can emit code that is:

  • physically impossible — e.g. lowering an 8-channel pipette into a narrow trough that is too short;
  • logistically invalid — e.g. aspirating from a depleted well;
  • scientifically compromised — e.g. by unmodeled fluidic effects (dispensing too close to the well floor causing uneven stratification).

A static validator over the code text catches none of these. Only a stateful model of the deck — the digital twin — can, because the answer to "is this command safe?" is a function of live state, not syntax. (Source: sources/2026-10-16-google-grounding-ai-agents-in-biological-reality)

Four core capabilities

Capability What it does
Live 3D state tracking Agents check deck layout, well volumes, and available tips before acting. The twin is the authoritative live state the agent queries.
Fast virtual dry runs Simulate a protocol in seconds to catch collisions or empty wells before touching reagents — a physical-world simulate-then-execute.
Hardware safety checks Block invalid commands before they reach the robot — a fail-closed actuation boundary.
Synchronized video + action logs Capture a complete record of every run for research rigor, reproducibility, and debugging — an execution trace as a first-class artifact.

Together these form a verify-before-actuate gate: the agent never sends unverified code directly to the liquid-handling robots; every protocol is dry-run against the twin and validated before any reagent is touched.

Why the twin is more than a safety gate

The persistent twin state makes failures recoverable and replannable, not merely rejectable:

  • Mid-run recovery. In one campaign the persistent digital-twin state "allowed the agent to recover mid-run from a plate-reader failure without re-pipetting" — the twin held enough live state to resume rather than restart.
  • On-the-fly replanning. When the simulator rejected an 8-channel aspiration that physically wouldn't fit a narrow trough, the agent replanned on the fly, designing a targeted factorial diagnostic that isolated a subtle bias (residual tip-wall carryover) replication alone would have missed.
  • Self-correction from traces. In a closed-loop self-play campaign an agent iteratively ran robot experiments to reconstruct blinded dye mixtures from overhead camera images, identifying its own optical and fluidic artifacts across runs to improve accuracy >5×.

Reproducible execution plus full trace logs — capturing procedural variables like dwell times, evaporation, and plate layout — are framed as the groundwork for agents to "eventually be able to separate biological signals from experimental noise as they explore and refine many candidates in parallel across rounds." (Source: sources/2026-10-16-google-grounding-ai-agents-in-biological-reality)

Safety posture

VERID is the technical mechanism behind Google's Frontier-Safety-Framework / bioresilience stance for autonomous labs. Three layers:

  1. Simulation before execution. "Agents never send unverified code directly to the liquid handling robots. Every protocol is dry-run in simulation and validated against the digital twin before touching reagents." — the fail-closed actuation rule.
  2. Guardrails on what gets designed/run. Every designed sequence goes through biosecurity screening; testbeds are deliberately well-controlled, non-infectious, no-gain-of-function systems chosen to catch drift early. See concepts/ai-agent-guardrails.
  3. Human oversight. Humans "directly authorize and review designed sequences, verify wet-lab protocols, and supervise robot runs" — a concepts/human-in-the-loop stance applied to physical actuation.

Relationship to other Google "digital twin" work

VERID is Google's second wiki-tracked "digital twin": the RLM / Borg work uses a digital-twin backtesting framework that replicates cluster state as a training- data source and authoritative fallback. Both are stateful replicas used to make decisions safely without hitting the real system first — one for cloud scheduling (software), one for lab robotics (physical). The shared shape: simulate against a live replica, fall back / block when the real action would be unsafe or expensive.

Caveats

VERID's wiki coverage is a blog-level sketch. The twin's data model, the simulator's physics fidelity, the state-update / consistency protocol (how the twin stays in sync with real sensor readings mid-run), dry-run latency beyond "seconds," and the workcell's throughput/scale (robots, concurrent protocols) are not in the source post. Architectural depth lives in the linked report (verid_semantic_lab.pdf), which is not ingested.

Seen in

  • sources/2026-10-16-google-grounding-ai-agents-in-biological-reality — Grounding AI agents in biological reality (Google Research, 2026-10-16). The system's introduction: semantic digital twin, four core capabilities, simulate-before-execute safety gate, mid-run recovery/replanning, closed-loop aptamer + de-novo protein-binder campaigns.
Last updated · 854 distilled / 2,448 read