Source-linked AI summary
Semantic Representation for Agentic Self-Play in Automated Laboratories
Yves Delley, John Coller, Zachary Cutts, Alexander Daryin, Mathias Bellaiche, Jun W. Kim, Oscar Donosa, Srinidhi B. Seshadri, Tanner Clary, Sandy Klemm, Avinatan Hassidim, Yossi Matias, Pooja Rao, Petar Sirkovic, Annalisa Pawlosky
TL;DR
Biological search requires both navigation of vast combinatorial spaces and reliable physical experimentation, but computational ideation and laboratory execution remain difficult to connect. VERID places a persistent semantic digital twin between agents and the workcell, and dye-based case studies demonstrate protocol error prevention, recovery, and diagnosis of physical artifacts.
Problem
Biological discovery must navigate vast search spaces while translating experimental designs into reliable physical execution, a gap existing agentic platforms still address with fixed workflows.
Source: §1 · p. 1Method
VERID uses a persistent, queryable semantic digital twin to connect reasoning agents with the workcell through simulation, trace persistence, and safe execution.
Source: §1 · p. 3, §2.1 · p. 3Results
Dye-and-water case studies demonstrated pre-execution error prevention, stateful recovery, and post-hoc diagnosis of subtle physical artifacts.
Source: §1 · p. 3, §3.1 · p. 6, §3.2 · p. 7Takeaways & Limitations
The case studies support VERID as infrastructure for separating systematic experimental bias from small biological signals and closing computational-experimental feedback loops.
Source: §4 · pp. 9–10, §5 · p. 12Whether these gains lead to significant biological breakthroughs remains to be demonstrated.
Source: §5 · p. 12
Abstract
from arXiv · showhide
Navigating complex, high-dimensional biological search spaces necessitates scalable physical and computational intelligence to explore beyond local optima through iterative self-play across ideation, experimentation, and automation. Translating experimental designs into reliable physical execution remains bottlenecked by static laboratory automation and unmodeled physical variability. To address this, we introduce VERified Infrastructure for Discovery Laboratories, or VERID, a foundational infrastructure designed to bridge the gap between the high-level reasoning of artificial intelligence (AI) agents and complex, error-prone physical realities of wet labs. VERID is a persistent, high-fidelity semantic digital twin of the laboratory workcell that provides four core services: queryable introspection for agentic reasoning, time-compressed simulation for protocol validation, comprehensive trace persistence for a complete scientific record, and a framework for safe execution. We present three case studies using aqueous dye and deionized water mixtures to validate the optical and fluidic foundations of VERID. These studies demonstrate how VERID enables pre-emptive detection of critical errors in experimental protocols, stateful mid-run error recovery and post-hoc identification of subtle physical failures. The simulation functions as a discrete accounting engine, allowing agents to safely and inexpensively explore experimental designs in silico and refine experimental strategies through iterative self-play. VERID also captures structured multimodal experimental data for downstream analysis, knowledge mining and agent training. We argue that this semantic layer is an essential prerequisite for deploying autonomous agents in the biochemical and biomedical sciences, ensuring data integrity, operational safety, and paving the way for tackling previously intractable grand challenges in biology.
1. Introduction
Autonomous discovery must navigate vast biological search spaces while reliably executing experiments in the physical lab. VERID is introduced as a persistent semantic digital twin to connect computational reasoning with physical execution and address the limits of existing approaches.
1. Introduction
- Biological discovery requires navigating immense, rugged search spaces and executing laboratory experiments with high fidelity, adaptability, and reproducibility.The introduction describes these as intertwined challenges that have largely been addressed independently.Source: §1 · p. 1
- Existing autonomous platforms that couple computational and experimental work still rely on fixed, pre-programmed workflows rather than adaptable agentic protocols.Biological search spaces are combinatorially large, and random-library screening can limit exploration to local hill climbing.Source: §1 · pp. 1–2
- In silico predictions can fail experimentally, while manual protocols and static automation constrain throughput, adaptability, or reproducibility.The introduction reports that 60% to >80% of computationally top-ranked de novo protein binder designs typically drop off in experimental testing.Source: §1 · p. 2
- VERID introduces a persistent semantic digital twin with queryable introspection, time-compressed simulation, multimodal trace persistence, and safe execution.Dye-and-water case studies benchmark foundational fluidic and optical operations and demonstrate pre-execution error prevention, stateful recovery, and causal debugging of physical artifacts.Source: §1 · p. 3
2. Results: The Semantic Agentic Lab Twin (VERID)
VERID places a persistent semantic digital twin between reasoning agents and the robotic workcell. Its services let agents inspect live state, test protocols virtually, preserve multimodal traces, and gate hardware actions against physical context.
2. Results: The Semantic Agentic Lab Twin (VERID)
- VERID abstracts hardware communication while persisting general workcell state, separating equipment from protocols to support use across labs, hardware, and consumables.Its server-managed digital twin tracks deck items, their geometry, status, and semantic contents across command blocks and instruments.Source: §2 · p. 3
2.1. Adaptation for Agentic Reasoning
- A structured, queryable API exposes live workcell state so agents can ground instructions in physical constraints rather than infer state from logs or camera pixels.Queries can return well volume and liquid height or identify tips and troughs with sufficient capacity.Source: §2.1 · p. 3
2.2. Efficient Simulation for Exploration
- Before dispatch, VERID clones the current twin state and runs a proposed command block as a time-compressed, deterministic discrete-state simulation.The simulator tracks mass balance, tip state, geometry, and elapsed time, allowing protocol refinement without consuming reagents.Source: §2.2 · pp. 3–4, Figure 1 · p. 4
2.3. Trace Persistence for a Complete Scien tific Record
- VERID synchronizes twin snapshots, time-tagged imagery and telemetry, and execution logs against a common clock to create an auditable research dossier.The dossier links protocol code to hardware commands, state transitions, and video frames for debugging, reproducibility, and agent learning.Source: §2.3 · p. 4
2.4. Safe Execution
- VERID checks mechanical trajectories, volumetric bounds, and chemical compatibility against current twin state, filtering unsafe actions before hardware dispatch.The simulation is re-synced with physical state and commands are re-gated in a verify-then-execute framework.Source: §2.4 · p. 5
3. Results: Case Studies
Dye-and-water case studies test VERID through protocol refinement, simulation-gated planning and diagnosis, and constrained self-play. They report near-target dilution after refinement, diagnosis of a wall-film carryover artifact, and improved mixture identification across multi-cycle challenges.
3. Results: Case Studies
- The three case studies use dye and deionized-water mixtures to demonstrate protocol debugging, multi-step fault recovery, and closed-loop self-learning while avoiding biological complexity.The self-learning challenge varied access to prior learnings and whether intermediate results could be inspected before further tasks were dispatched.Source: §3 · p. 5
3.1. Case
- 1.92×–1.99× average step dilution factors restored near-ideal 2-fold dilution, and the refined protocol reduced tube-to-trough differences to ≤0.02× across four dyes.The protocol used 1.5 mm bottom clearance and vertical convective mixing; persistent twin state also let the agent resume readout after an instrument-state error without re-pipetting.Source: §3.1 · p. 6, Figure 2 · p. 6
3.2. Case Study 2: Simulation-Gated Protocol Planning and Multi-Step Diagnosis
- The simulation rejected an 8-channel trough entry because the 72.0 mm tip span exceeded the 71.2 mm trough length, prompting a 7-channel mix and two 4-channel sampling passes.The semantic model also identified that 15.7 mL trough capacity made in-place dilution impossible for 12 mL target volumes under the stated stock-concentration uncertainty.Source: §3.2 · p. 6, Figure 3 · p. 7
- After a communication failure interrupted the fourth dye, persistent twin state allowed the run to continue from the fault point instead of repeating completed work.The agent then designed a factorial diagnostic across two runs, with the second run resuming on persisted Columns 5–8 without setup overhead.Source: §3.2 · p. 7
- Reversing the 7-channel trough mix shifted the absorbance deficit from Channel 8 to Channel 1, identifying wall-film carryover rather than Channel 8 under-aspiration.The carryover added +4.9 µL (+40.6%) during well mixing and explained the 14.5% offset between 24 µL and 12 µL dilution levels.Source: §3.2 · p. 7
3.3. Case Study 3: End-to-End Physical Self-Play in Automated Laboratories
- On Challenge 2B, MC-2B reached 58 mnats mean loss, more than a five-fold improvement over MC-2A, by comparing full dilution trajectories rather than single spiked wells and calibrating uncertainty.MC-2A's loss rose from 247 to 307 mnats across four cycles. MC-2B used a sparse initial dilution survey with pre- and post-dispense blank imaging, followed by four-point serial dilutions.Source: §3.3 · p. 8
- In Challenge 1, MC-1 achieved 85 mnats versus SC-1’s 205 mnats through within-run feedback alone, while the single-cycle SC-2A→SC-2B lineage stalled at 142→162 mnats.The case study compared solvers under fixed physical budgets, with no access to ground-truth identities or Oracle scores.Source: §3.3 · p. 8
- Across Challenges 2A and 2B, agents used residuals to identify camera, pipetting, and optical artifacts, then adapted measurement and transfer bounds and their camera model.In Challenge 2B they also detected vertical concentration gradients in unmixed deep-well source plates and Yellow–Blue–Green twin degeneracy, motivating dilution trajectories and uncertainty calibration.Source: §3.3 · p. 9
4. Discussion
The case studies empirically validate VERID’s core services and show how its infrastructure connects high-level scientific reasoning with automated wet-lab execution. The architecture targets systematic experimental bias and closes the computational-experimental loop through structured trajectories and autonomous learning.
4. Discussion
- The use cases empirically validate VERID’s core services and connect high-level scientific reasoning with automated wet-lab execution.Source: §4 · p. 9
- MC-2A ended at 307 mnats final loss, while MC-2B recorded 58 mnats mean loss after inheriting sample-blind learnings and conducting a two-cycle dilution survey.The figure also reports autonomous physical and optical discoveries from internal consistency residuals.Source: Figure 4 · p. 9
- VERID addresses systematic experimental bias in physical assay execution and closes the computational-experimental loop through structured trajectories and autonomous learning.Source: §4 · p. 10
4.1. Isolating Systematic Experimental Bias in High-Throughput Assays
The discussion distinguishes stochastic noise from systematic bias, which can be confounded by instrument, labware, fluid, and protocol variation. VERID’s execution traces and targeted controls help agents identify physical causes that replication alone cannot correct.
4.1. Isolating Systematic Experimental Bias in High-Throughput Assays
- High-dimensional assay artifacts can masquerade as biological signal, while replication mitigates unbiased random error but not systematic bias.The discussion identifies instrument tolerances, labware geometry, fluid properties, and untracked protocol variation as sources of confounding.Source: §4.1 · p. 10
- VERID records execution covariates, including channel assignments, tip reuse, aspiration heights, meniscus clearances, dwell times, and synchronized video, for agent inspection.Agents can correlate anomalies with execution parameters and pause a campaign for targeted factorial controls on the same deck state.Source: §4.1 · p. 10
- In Case Studies 1 and 2, trace inspection and controlled channel-assignment inversion isolated under-dilution and inner-wall film carryover that replication could not fix.The carryover reproducibly biased Channels 1–7 by +40.6% and made Channel 8 appear defective.Source: §4.1 · p. 10
4.2. Data Synchronisation and Persistence for Knowledge Extraction
VERID synchronizes multimodal experiment data into a unified trace dossier. The dossier can support agentic analysis and downstream learning, while large-scale indexing and benchmarking remain future work.
4.2. Data Synchronisation and Persistence for Knowledge Extraction
- VERID synchronizes data streams, digital-twin snapshots, command traces, video, telemetry, and endpoint readouts onto a unified timeline in a single trace dossier.Source: §4.2 · p. 10
- The experiment-specific data layer can support improved experimental design, hypothesis refinement, or reinforcement learning of the underlying AI model.Source: §4.2 · p. 10
- Automated indexing and knowledge extraction across archived trace dossiers, and systematic benchmarking of downstream impact, remain future work.Source: §4.2 · p. 10
4.3. Closing the Loop for Self-Play Agentic Learning
VERID supports agents in accumulating transferable expertise about wet-lab robotics by learning from unexpected physical behavior. The learning signal is self-supervised prediction error, and the agent learns unmodeled physical residuals within a semantic, constraint-aware world model.
4.3. Closing the Loop for Self-Play Agentic Learning
- Agents can identify gaps in their assumptions and unexpected physical behavior, then distill operational rules and observation models to initialize later iterations.Source: §4.3 · p. 10
- Learning uses self-supervised prediction error—the mismatch between expected and observed physical outcomes—rather than a task-specific extrinsic reward function.Source: §4.3 · p. 11
- Because the digital twin enforces volumetric conservation and geometric constraints, agents learn unmodeled residuals such as tip-wall carryover, density stratification, and optical cross-coupling.Discoveries are stored as natural-language operating bounds and analytical models, which agents can inspect when selecting verified regimes or planning targeted diagnostics.Source: §4.3 · p. 11
4.4. Closing the Loop for Biological Hill-Climbing
VERID’s trace dossiers give optimization procedures access to experimental execution details as well as measured biological objectives. These details include intermediate readouts and procedural covariates relevant to selecting subsequent experiments.
4.4. Closing the Loop for Biological Hill-Climbing
- Measured biological objectives can also feed back into domain-adapted or general black-box optimization procedures.The passage gives drug-repurposing combinations, small-molecule dosing of IDPs, and cellular perturbations as possible applications.Source: §4.4 · p. 11
- A suitably sophisticated optimizer can use intermediate readouts and procedural covariates from trace dossiers when selecting the next experiment.Readouts may flag sample-preparation issues, while covariates such as dwell times and plate positions may correlate with bias or measurement uncertainty.Source: §4.4 · p. 11
4.5. Safety and Responsible Innovation
VERID combines machine-checkable operational boundaries with human review and proposed biosecurity screening. Its trace dossiers support scientists’ inspection of anomalous runs, while the biosecurity extension is framed as a further safeguard for synthesis workflows.
4.5.1. Operational Safety
- VERID encodes operational boundaries as machine-checkable invariants and enforces them at the pre-execution simulation gate.These include capacity, clearance, agitation-speed, and chemical-compatibility limits; mechanical and chemical constraints screen for collision, contamination, and hazardous-reaction risks.Source: §4.5.1 · p. 11
4.5.2. Human-in-the-Loop (HITL) Review
- Synchronized trace dossiers let scientists inspect anomalous steps and record structured annotations that feed tacit laboratory know-how into the training archive.The dossiers align code execution with gantry video and well-level state histories.Source: §4.5.2 · pp. 11–12
4.5.3. Automated Biosecurity
- For de novo nucleic-acid or protein synthesis, the authors propose extending verification with lineage tracking, physical-chemistry rules, and automated biosecurity review.They describe function-aware screening as a safeguard because generated functional sequences may have low identity to known pathogens and evade homology-based screens.Source: §4.5.3 · p. 12
5. Conclusion
VERID places a persistent, queryable semantic digital twin between reasoning agents and the physical workcell. The case studies demonstrate protocol-error detection, stateful recovery, and diagnosis of subtle physical biases; significant biological breakthroughs remain a next step.
5. Conclusion
- VERID’s case studies show simulation-based detection of geometric and volumetric errors, stateful recovery from interruptions, and factorial diagnosis of fluid stratification and tip-wall carryover.The authors describe trace dossiers as auditable execution trajectories and structured records.Source: §5 · p. 12
- Demonstrating that these gains yield significant biological breakthroughs on grand-challenge questions remains future work.The authors characterize separating systematic noise from small biological signals as a physical prerequisite for agents to experiment and learn at scale.Source: §5 · p. 12
6. Methods
The methods describe a robotic workcell, an agent-facing client–server software architecture, and a scoring rule for mixture-composition predictions. The score combines relative-composition and total-concentration errors across four target wells.
6.1. Robotic Workcell and Reagents
- Case studies used a Hamilton liquid handler, a four-wavelength plate reader, a robotic arm, and cameras, with deionized water and retail food-colorant solutions as model reagents.The workcell included 96-channel and independent 8-channel pipetting heads.Source: §6.1 · p. 12
6.2. Software Architecture and Agent Inter face
- Agents inspect the semantic workcell state, test commands on a forked virtual replica, and dispatch commands that are re-verified against the live state before execution.Completion artifacts include state snapshots, logs, video, and assay readouts; supervised runs used an operator approval gate locked to the verified script.Source: §6.2 · pp. 12–13
6.3. Scoring Rules for Case Study 3
- Each target-well loss sums composition and concentration errors, and the run score is their mean across four target wells, reported in millinats.Composition error is KL divergence between true and scored tempered dye distributions; concentration error is a weighted squared logit error on clipped water fraction.Source: §6.3 · p. 13