Self-Driving Laboratories
Self-driving labs, also called autonomous laboratories, close the experimental loop. A model proposes an experiment, automated hardware runs it, the result returns to the model, and the model proposes the next experiment. The autonomy is not in the robotics alone; it is in the loop. Evidence spans closed-loop chemistry and materials systems, while current biology examples remain lab-in-the-loop rather than autonomous robotic loops. Coscientist demonstrated the pattern for Pd-catalysed cross-coupling chemistry with GPT-4 as planner (Boiko et al., 2023). Virtual Lab demonstrated multi-agent biology research that produced experimentally validated SARS-CoV-2 nanobodies (Swanson et al., 2025). Robin extended the biology branch into lab-in-the-loop therapeutic-candidate discovery for dry age-related macular degeneration (Ghareeb et al., 2026). A-Lab reported autonomous discovery of dozens of inorganic materials and then drew published commentary and a later PRX Energy critique that illustrate how easily novelty claims can run ahead of validation (Szymanski et al., 2023; Neilson, 2023; Leeman et al., 2024). The capability is real for bounded optimisation; general autonomous discovery is not.
Use this chapter to:
- Close the loop between hypothesis generation, experiment planning, robotic execution, measurement, and model update.
- Autonomy is bounded by protocol validity, instrument calibration, search space, safety rules, and whether the measured endpoint is meaningful.
Prerequisites: Robotic Lab Automation and Cloud Labs for the hardware layer; Evaluation Principles for Life Sciences AI for the prospective-validation discipline.
Summary: Close the loop between hypothesis generation, experiment planning, robotic execution, measurement, and model update. Self-driving systems are real in selected chemistry and materials tasks; broad autonomous biology remains much more constrained.
Key point: Autonomy is bounded by protocol validity, instrument calibration, search space, safety rules, and whether the measured endpoint is meaningful. Open question: whether closed loops reproduce across laboratories, endpoints, organisms, and failure conditions.
Bottom line: Self-driving laboratories connect experimental design, automation, chemistry, synthetic biology, biomanufacturing, agents, and research governance.
What is this field trying to solve? Close the loop between hypothesis generation, experiment planning, robotic execution, measurement, and model update.
What is the core idea? Autonomy is bounded by protocol validity, instrument calibration, search space, safety rules, and whether the measured endpoint is meaningful.
What is the current state of the field? Self-driving systems are real in selected chemistry and materials tasks; broad autonomous biology remains much more constrained.
What do we know, and what remains open? Known reference points include Adam, Eve, Coscientist, Virtual Lab, Robin, A-Lab, Ada, ChemOS, mobile robotic chemists, cloud labs, optimization benchmarks, and automated assay platforms. What remains open is whether closed loops reproduce across laboratories, endpoints, organisms, and failure conditions.
Why does this matter? Self-driving laboratories connect experimental design, automation, chemistry, synthetic biology, biomanufacturing, agents, and research governance.
Introduction
A self-driving laboratory has three parts: an inference loop that proposes the next experiment, hardware that executes the experiment, and an evaluation step that returns the result to the inference loop. The composition is older than current AI: Adam, built by Ross King and colleagues at Aberystwyth, demonstrated autonomous functional-genomics hypothesis testing in 2009 with logic-programming-based reasoning (King et al., 2009). Adam’s successor Eve focused on drug screening. The historical lineage matters because the conceptual contribution of self-driving labs predates the current LLM wave by more than a decade.
The current wave has three branches:
- Chemistry and materials automation has peer-reviewed demonstrations for bounded tasks: thin-film optimization (Ada, MacLeod et al., 2020), mobile robotic chemistry (Burger et al., 2020), Pd-catalysed coupling (Coscientist, Boiko et al., 2023), and workflow orchestration (ChemOS, Roch et al., 2020). These studies do not establish production readiness across laboratories or reaction classes.
- Materials autonomous discovery is a prominent example, with caveats: A-Lab’s reported 41 new inorganic materials in 17 days (Szymanski et al., 2023) drew published commentary and a later PRX Energy critique arguing that many “new” materials required stronger phase-identification and novelty evidence (Neilson, 2023; Leeman et al., 2024). The episode is a clear case study of how autonomous-discovery novelty claims can outrun validation.
- Agentic biology research has peer-reviewed, lab-in-the-loop examples: Virtual Lab’s SARS-CoV-2 nanobody design with human-executed validation (Swanson et al., 2025) and Robin’s dry age-related macular degeneration candidate work (Ghareeb et al., 2026). These are adjacent to, but not evidence of, autonomous robotic biology loops.
This chapter applies the evidence framework to each branch and explains the historical context that current claims are read against.
Demonstrated capability
Chemistry automation: the mature branch
Coscientist is a peer-reviewed example of LLM-planned chemistry (Boiko et al., 2023). The system used language-model planning, web and documentation search, code execution, and automated experimental tools. The experimental case included Pd-catalysed cross-coupling reactions. The evidence supports orchestration of bounded, known chemical procedures, not general autonomous chemical discovery.
A practical self-driving lab is a design-make-test-analyse system, not only a model attached to instruments. Reviews of autonomous chemical experimentation emphasise that chemical production, characterization, calibration, and exception handling remain bottlenecks even when optimisation logic is automated (Seifrid et al., 2022). This is why the strongest papers report the physical loop, the control policy, and the validation result together.
A 2026 Nature Reviews Chemistry assessment identifies scalability, generalizability, and provenance-complete experimentation as the interdependent requirements for moving self-driving laboratories from focused demonstrations toward shared infrastructure (Canty and Abolhasani, 2026).
AutoLabs provides a current test of the agent-to-hardware boundary. The 2026 Scientific Reports study evaluated 20 agent configurations across five benchmark experiments and reported an F1 score above 0.89 on its challenging multi-plate chemistry benchmark when multi-agent orchestration and iterative self-correction were combined (Panapitiya et al., 2026). The system generated hardware-ready XML for one high-throughput liquid-handler platform. This is evidence for protocol translation and platform-specific execution, not autonomous discovery or transfer to other laboratory systems.
Ada (MacLeod et al., 2020) is an earlier and more focused example: a self-driving laboratory for thin-film materials discovery using Bayesian optimisation as the loop driver. The Science Advances paper reported optimisation of organic photovoltaic films with hundreds of automated cycles. The conceptual contribution was Bayesian-optimisation-driven autonomy at production scale before the LLM era.
ChemOS (Roch et al., 2020) is the orchestration layer beneath several of these systems. The PLOS ONE paper described an open-source workflow framework that handles experiment specification, scheduling, hardware integration, and result handling. ChemOS is a piece of infrastructure rather than a research result, and it appears in the methods sections of many subsequent self-driving lab papers.
The mobile robotic chemist (Burger et al., 2020) addressed a different constraint: untethered robotics that can move between instruments and conduct multi-instrument workflows. The Nature paper demonstrated an autonomous chemistry workflow that ran continuously for over a week and explored a photocatalyst search space.
Adjacent lab-in-the-loop agentic biology
Virtual Lab (Swanson et al., 2025) is a peer-reviewed example of multi-agent biology research. The Nature paper described a PI agent coordinating specialist agents to design 92 nanobodies against SARS-CoV-2 variants; selected designs were tested experimentally. Virtual Lab did not autonomously execute a robotic wet-lab feedback loop, so it should not be counted as self-driving-laboratory evidence.
Robin (Ghareeb et al., 2026) is the next major evidence point. It combined literature-search agents and data-analysis agents to generate hypotheses, propose experiments, interpret results, and update hypotheses. In the reported dry age-related macular degeneration workflow, Robin identified ripasudil and KL001 as candidates, proposed follow-up RNA-seq analysis, and generated the report’s main-text hypotheses, analyses, and figures. The important qualifier is the phrase lab-in-the-loop: humans still executed the wet-lab work, and replication remains the threshold for stronger claims.
Outside these systems, biology automation has been demonstrated for narrower tasks such as directed-evolution loops and assay optimization. Published autonomous-laboratory evidence is more extensive in chemistry and materials than in open-ended biology. The assay, hardware, metadata, and exception-handling layers all constrain transfer.
A-Lab and the novelty-claim discipline
A-Lab (Szymanski et al., 2023) reported 41 new inorganic materials autonomously produced and characterised in 17 days. The Nature 2023 paper was the highest-profile autonomous-discovery result of the year and drew immediate attention. Subsequent commentary and a later PRX Energy critique from solid-state chemists argued that novelty claims required stronger treatment of known phases, disorder, polymorphism, and automated X-ray diffraction interpretation (Neilson, 2023; Leeman et al., 2024).
The episode is the clearest current case study of the novelty-claim discipline that self-driving labs require:
- Autonomous platform throughput is not autonomous discovery. Running 1,000 experiments per day produces 1,000 experiments per day. Whether any of them is a discovery depends on what was already known.
- Automated characterisation can be wrong in characteristic ways. Powder X-ray diffraction pipelines, machine-learning structure-solving, and database matching each have failure modes that look like new materials.
- Independent expert review is part of validation. A novelty claim that does not pass scrutiny from the relevant subdiscipline is not a discovery.
A-Lab does not invalidate autonomous-laboratory work. It shows what the validation bar looks like.
Historical lineage: Adam, Eve, and the pre-LLM era
Adam (King et al., 2009), built at Aberystwyth, was the first robot scientist to autonomously generate functional-genomics hypotheses, design experiments to test them, and update its beliefs from results. The reasoning engine used logic programming and abductive inference. Adam ran for months with minimal human intervention and produced hypotheses about gene function in yeast that were independently verified. Eve, Adam’s successor, applied the same closed-loop architecture to drug screening.
The Adam and Eve work is conceptually upstream of the current LLM-driven systems. The closed loop, the integration with hardware, the question of what counts as discovery, and the validation discipline were all explored at Aberystwyth more than a decade before Coscientist. Current systems should be read in continuity with this lineage, not as a discontinuous emergence from LLM scaling.
Agentic chemistry orchestration: ChemCrow
ChemCrow (M. Bran et al., 2024) wraps an LLM around chemistry-specific tools (RDKit, retro-route planners, property predictors, web search, hardware interfaces). The Nature Machine Intelligence paper demonstrated tool-using agents on chemical synthesis planning tasks, drug-discovery workflows, and materials-design queries. ChemCrow is conceptually adjacent to Coscientist: where Coscientist is task-specific and hardware-coupled, ChemCrow is tool-coupled and more general. Together they describe two complementary approaches to LLM-driven chemistry workflows. The agentic-orchestration layer is covered in detail in Agentic Science Workflows.
Evidence anchor summary
| Evidence Anchor | What It Supports | Practical Constraint |
|---|---|---|
| Coscientist | LLM-planned autonomous Pd cross-coupling chemistry | Reproduces known chemistry; novel-reaction discovery is separate |
| AutoLabs | Agent-generated, hardware-ready protocols on a bounded chemistry benchmark | One liquid-handler platform; protocol execution is not autonomous discovery |
| Virtual Lab | Adjacent multi-agent biology research with human-executed experimental validation | Not a robotic closed loop; one nanobody-design context |
| Robin | Adjacent lab-in-the-loop agentic research | Humans executed wet-lab work; replication and transfer remain open |
| A-Lab | Autonomous inorganic materials platform at scale | Novelty claims drew published commentary and PRX Energy critique; validation discipline required |
| Ada | Bayesian-optimisation-driven thin-film discovery | Pre-LLM workflow at production scale |
| ChemOS | Open-source orchestration layer for self-driving chemistry | Infrastructure, not a research result by itself |
| Burger mobile chemist | Untethered robotic chemistry across instruments | Single laboratory, single chemistry domain |
| ChemCrow | LLM-orchestrated chemistry tools | Tool-coupled rather than hardware-coupled |
| Adam (King 2009) | First closed-loop robot scientist (functional genomics) | Logic-programming reasoning, not deep-learning |
| Self-driving lab review (Canty 2026) | Current field assessment | Scalability, generalizability, and provenance remain infrastructure constraints (Canty and Abolhasani, 2026) |
Theoretical capability
Several capabilities are plausible but not yet routine.
Closed-loop biology across targets and assays. Biology assays can be slow, variable, and difficult to automate. Virtual Lab and Robin are agentic, lab-in-the-loop examples rather than autonomous robotic loops. Prospective, repeated closed-loop biology across multiple targets and assay classes remains an open evidence need.
Autonomous discovery of new reaction mechanisms. Reproducing known chemistry at high throughput is one thing; discovering new chemistry is another. Current systems can explore parameter spaces inside a defined reaction class; jumping to a genuinely new reaction class typically requires human chemists to define the search space.
Cross-laboratory reproducibility. Independent reproduction on a materially different autonomous platform would provide stronger evidence of transfer. Current work remains dominated by single-platform and single-laboratory demonstrations.
Autonomous in-vivo experimentation. Cell-free and in-vitro experiments dominate current self-driving lab work. In-vivo experimentation adds welfare, protocol, review, monitoring, assay, and robotics constraints. These are governance and technical limits, not one interchangeable bottleneck.
Truly autonomous goal-setting. Current systems take a human-defined goal (optimise this property, find a binder to this target) and execute toward it. Setting the goal itself, deciding which problem to work on, remains human. Whether this is a desirable target for autonomy is an open question.
Beyond current capability
A few framing claims are not supported by current evidence.
Self-driving labs make scientists obsolete. They do not. Goal definition, experimental judgement, novelty assessment, and integration across subdisciplines remain human. The autonomous platforms accelerate the parts of science that are well-characterised; they do not replace the parts that are not.
Autonomous labs reliably discover new biology without expert oversight. Current evidence does not support that claim. The A-Lab episode illustrates that high-throughput platforms can produce candidate discoveries that require expert review and independent validation. Validation is part of the scientific claim, not a downstream formality.
Closed-loop optimisation generalises across science. It does for narrow, well-characterised optimisation problems (thin films, photocatalysts, reaction yields). It does not yet for open-ended discovery, where the right question is part of the problem.
Robot scientists are a recent invention. They are not. Adam and Eve at Aberystwyth demonstrated the closed-loop concept in 2009. Current LLM-driven systems are an architectural advance over Adam’s logic programming, not a discontinuous new capability.
Evidence that would change the assessment
Autonomous-lab claims become more promising when they reproduce across independent facilities, with pre-specified objectives, complete loop logs, and external confirmation of novelty or function. For biology, the threshold is prospective closed-loop optimisation that improves a validated assay across multiple targets or cell systems without hiding exception handling, failed loops, or human intervention.
Implications for research and program decisions
For researchers and program leaders evaluating self-driving lab claims:
- Count the loop iterations. A platform that ran once is automation; a platform that ran a closed inference-experiment-update loop many times is autonomy. The number of loop iterations should be in the paper.
- Read the validation layer. What experiments were run to verify novelty, correctness, and reproducibility? A-Lab’s experience shows what happens when this layer is thin.
- Match the chemistry-to-biology gap. Bounded chemistry and materials loops have more mature demonstrations than autonomous biology loops. A platform that works for thin films does not directly transfer to nanobody discovery, and vice versa.
- Read the hardware layer alongside the algorithm layer. Cloud-lab integration (Emerald, Strateos), robotic-arm reliability, and assay-format constraints often dominate what is actually possible. The published algorithm sits on top of a hardware stack that bounds its real performance.
- Treat single-laboratory results as proof of concept, not as a community capability. A self-driving result that works in one lab and does not reproduce in another has demonstrated something narrower than the abstract suggests.
- Cite the historical lineage. Current self-driving lab papers that ignore Adam, Eve, Ada, ChemOS, and the Häse 2019 review tend also to overclaim novelty. The honest framing locates current work in the trajectory.
- Pair this chapter with Agentic Science Workflows. The orchestration-agent layer (ChemCrow, Coscientist’s planner, Virtual Lab’s PI-agent) is conceptually distinct from the autonomous-loop layer, even though both appear in the same papers.