Google Research has described a set of projects that connect AI agents to physical laboratory robots, so that an agent can propose an experiment, run it, read the results and design the next round. In a post dated October 16, researchers Pooja Rao, Petar Sirkovic and Annalisa Pawlosky present three pieces of work: a safety framework called VERID, a closed-loop study with DNA aptamers, and an approach to designing protein binders against a cancer-related target. These are company research reports, and we have not seen independent replication; the post does not say whether the three reports have been peer reviewed.

The problem: agents are good at ideas, weak at physics

The authors name two obstacles. First, frontier models “still do not reliably reason about 3-dimensional geometry and physical state,” which is a problem when commanding the liquid-handling robots that move reagents, track volumes and run molecular biology protocols. Second, while agents are getting good at generating hypotheses and designing molecules on a computer, real experimental feedback is slow, which makes lab readouts the bottleneck. The search space is vast: a single protein with 100 amino acids has at least 20 to the power of 100 possible sequences, so designers need many cycles of testing to find good candidates.

VERID: a digital twin that keeps the robot honest

VERID, short for VERified Infrastructure for Discovery Laboratories, is described as a semantic digital twin of the lab workcell. It is a live, stateful 3D entity tree that tracks robot arm geometry, individual well volumes, reagent identities and pipette tip inventory. It is built on the open-source PyLabRobot framework and provides four capabilities, according to the post: live 3D state tracking so an agent can check the deck layout, volumes and tips before acting; fast virtual dry runs that simulate a protocol in seconds to catch collisions or empty wells; hardware safety checks that block invalid commands before they reach the robot; and synchronized video and action logs that record each run for reproducibility and debugging.

The reason is that code can be syntactically valid yet physically impossible, for instance lowering an eight-channel pipette into a trough that is too short, or aspirating from a depleted well. In the reported case studies, video and execution traces caught uneven dye layering caused by dispensing too close to a well floor, and the persistent twin let an agent recover from a plate-reader failure without repeating the pipetting. In a closed-loop self-play campaign, an agent reconstructed blinded dye mixtures from overhead camera images and improved accuracy more than fivefold by identifying its own optical and fluidic artifacts.

Aptamers as a sandbox

DNA aptamers, short sequences that bind proteins, are a convenient test bed because a full design-test-analyze cycle takes only 7 to 10 days: DNA can be synthesized in pooled libraries and read out directly by sequencing. Over three rounds on two targets, myoglobin and alpha-thrombin, agents raised confirmed on-target hit rates to 33.5% and 42.1% respectively, according to Google’s report, though each target hid a different failure mode.

On alpha-thrombin, round one produced no binders because fixed primer overhangs disrupted folding in 48% of sequences; agents introduced clamped flanking stems in round two and diversified in round three, reaching 26.0% and then 42.1% hit rate at 98.5% specificity. On myoglobin, agents chasing raw binding strength heavily enriched G-quadruplex motifs, 63% of designs, which turned out to be sticky, non-specific binders. Specificity fell from 75.0% to 19.1%, and the true hit rate from 10.2% to 3.7%. After reading that round’s data, the agents revised their own rules, capped promiscuous motifs, subtracted bead binders and moved to other scaffolds, recovering specificity to 72.4% and lifting the on-target hit rate ninefold to 33.5%. This is the clearest example of the loop working: the wet-lab data corrected a mistake the in-silico design would not have revealed.

Protein binders

Proteins are harder, because they must be expressed biologically, which adds selection, expression and folding steps to each loop. For the third project, the team applied its workflow to fibroblast activation protein alpha (FAP), a large receptor overexpressed in solid tumors. Per the post, in just a few experimental runs the top designs produced binders with single-digit nanomolar affinity, a level of tight binding typical of drug-like molecules. Binding is only one step toward a medicine; it says nothing about safety, delivery or effect in patients.

Why it matters

Autonomous labs are one of the most plausible ways AI could speed up science, because the real constraint is often experimental throughput rather than ideas. The work also shows a design principle that applies beyond biology: agents need a trustworthy model of the physical state and a way to test commands safely before executing them. Readers new to the field can start with our overviews of AI in drug discovery and how AI is accelerating science, and with what AI agents are.

What to watch next

  • Peer-reviewed versions and open data releases for the three reports.
  • Independent labs reproducing the aptamer and FAP results.
  • Whether VERID or similar twins are adopted for other lab robots.
  • Safety frameworks for autonomous labs working with biological materials.
  • Progress from binders to functional tests in cells or animals.

Key takeaways

  • Google Research connects AI agents to lab robots to run closed design-test loops.
  • VERID is a digital twin that simulates and checks agent commands before they reach hardware.
  • Agents improved aptamer hit rates over three rounds and corrected their own design mistakes using wet-lab data.
  • Protein designs against FAP reached single-digit nanomolar affinity, per Google.
  • The results are company-reported and early.

Frequently asked questions

What is a digital twin of a lab?

A live software model of the physical workcell that tracks positions, volumes and supplies, so commands can be simulated and checked before running.

What is an aptamer?

A short strand of DNA or RNA that folds into a shape that can bind a specific target such as a protein.

Does this mean AI has designed a drug?

No. Strong binding to a target is an early step; it does not show a molecule is safe or effective.