Lab automation & self-driving laboratories

Experimental R&D is limited by how many well-chosen experiments a team can run and analyse. A self-driving laboratory (SDL) closes the loop: an algorithm (typically Bayesian optimisation, increasingly an LLM agent on top) proposes the next experiments, liquid handlers and robots run them, instruments measure, and the result updates the model. The benefit is fewer experiments to reach a target, 24/7 operation and complete, reproducible data. Most real deployments are partial: automated workcells with AI-guided design, not fully autonomous labs.

Problem & value

• Today: scientists design experiments by intuition or classic design of experiments (DoE), pipette or program robots by hand, transfer data between instruments, ELN and spreadsheets, and decide the next round days later. Protocols live in heterogeneous vendor software; metadata is often lost.
• AI value: sample-efficient search of large parameter spaces, automatic protocol generation, continuous operation, and structured provenance that makes data reusable for model training (see SCI-003, SCI-014).
• Evidence: a 2026 review in Nature Reviews Chemistry describes the history and state of SDLs and stresses that autonomy levels differ widely [1]. Trade sources note that "self-driving" is used loosely by vendors, and published speed-ups are mostly from narrow, well-defined optimisation tasks, so treat vendor claims as unverified until reproduced on your own chemistry.

Variants

• Closed-loop optimisation workcell: one process (reaction conditions, formulation, media, protein expression) optimised by Bayesian optimisation; the most common and most mature variant.
• Modular multi-instrument SDL: mobile robots or rail systems move samples between synthesis, analysis and characterisation stations (materials, organic synthesis).
• Cloud lab / remote lab: experiments are specified in code or a web UI and executed in a provider's facility (e.g. Emerald Cloud Lab) [4].
• Biology and pharma: high-throughput screening, assay development, cell-culture and strain engineering with liquid-handling platforms.
• Private use: not applicable.

How it works

1. goal + constraints + search space
2. experiment planner (Bayesian optimisation / active learning / LLM agent)
3. protocol compiler (JSON/XML worklist
4. robot and instrument drivers)
5. scheduler & orchestrator (resources, timing, error handling)
6. execution (liquid handler, synthesis robot, mobile robot, instruments)
7. analysis (peak integration, spectra interpretation, image/vision QC)
8. data layer (ELN/LIMS, provenance, FAIR metadata)
9. model update
10. next batch of experiments

• Planner: Gaussian-process or neural surrogate models with acquisition functions, multi-objective and batch variants, optionally warm-started from literature or foundation-model priors. LLM agents translate natural-language goals into protocols and code and call planner and instrument tools.
• Orchestration: an instrument-abstraction layer (SiLA 2, OPC UA LADS, vendor APIs) and a scheduler handle device heterogeneity, parallel runs and recovery from failed steps.
• Analysis in the loop: automated readouts must be trusted without a person looking; use QC checks, controls and replicate wells.
• Human-in-the-loop: scientists define objective, safe parameter bounds and hazard rules; approve new protocol classes and reagents; review anomalies and results before decisions with consequences. Hard safety interlocks are implemented in hardware and PLC logic, never only in the AI layer.

Data preparation

Data needed: historical experiments with parameters and outcomes, instrument and device descriptions, protocols, and live measurements. Typically these sit in ELN/LIMS systems, instrument PCs with vendor formats, Excel sheets, file shares and paper lab notebooks.

Typical gaps: free-text protocols, parameters not recorded, failed runs never logged, inconsistent units and sample IDs, vendor-locked formats, and no machine-readable instrument interface.

Preparation steps: 1. Assess readiness and map which data feeds the optimiser (PREP-001, GUIDE-003). 2. Digitise paper notebooks and logbooks at the source (PREP-008, PREP-006). 3. Define a schema for experiments, units and sample IDs; tidy past results (PREP-011, PREP-019). 4. Expose instruments via APIs/SiLA 2 and export from ELN/LIMS (PREP-005). 5. Record failures and metadata for FAIR provenance.

Tools: ELN/LIMS exports, OME/standard formats, Python data pipelines.

Share of effort: about 40-60 percent, plus hardware integration.

Implementation paths

Which path for whom

• Enterprise: Path B, because the planning and data layer must integrate with existing ELN/LIMS, instruments and workcells and no single product covers that.
• Developer: Path C, because open tools (BoTorch/Ax, SiLA 2, workflow engines) give full control over the closed loop.
• Public: Path C, because research institutes need open, reproducible, FAIR-compliant stacks and sovereign hosting.

Path A — Buy (off-the-shelf; good for hardware and a first workstation)

• Integrated platforms: Chemspeed (Swiss; with SciY/Bruker announced an open SDL platform combining vendor-agnostic automation, lab management software and AI orchestration at SLAS 2026) [2]; Automata (UK, LINQ orchestration platform for multi-vendor labs); Hamilton (CH), Tecan (CH), Beckman Coulter (Danaher), Opentrons, Formulatrix (liquid handling and workstations); HighRes Biosolutions and Biosero (BioNova) for workcell integration.
• Lab informatics: LIMS/ELN suppliers such as Benchling, Genedata (CH; screening and bioprocess), Scispot, Labguru, LabVantage, LabWare, and Sapio. Check EU hosting and open APIs.
• Cloud labs: Emerald Cloud Lab for remote execution [4].
• Choose A when: you have no automation team; start with one automated workstation and software that exposes APIs, and add the planning layer on top.

Path B — Build on cloud APIs

• Planner: use open-source Bayesian optimisation libraries (BoTorch/Ax) [3] as the core; add a reasoning LLM (Claude, Gemini, GPT, or Mistral as EU provider) through structured tool calls for protocol drafting, anomaly explanation and literature lookup. Run it with EU-region endpoints.
• Orchestration: Prefect, Dagster, Temporal or n8n for workflows; SiLA 2 or vendor SDKs for devices.
• Choose B when: you already own automation hardware and need a custom planning and data layer.

Path C — Self-host (open source)

• Optimisation and active learning: BoTorch/Ax, Atlas, Summit, scikit-optimize, Optuna; chemistry-specific tools such as ChemOS and ChemSpyd. Lab control: SiLA 2 reference implementations, PyLabRobot (liquid handlers), Opentrons API, OpenTrons/Jubilee open hardware.
• Language layer: open-weight models (Qwen, Llama, Mistral, gpt-oss families) served with vLLM for protocol generation and log analysis; scientific foundation models as priors (see SCI-003, SCI-018).
• Hardware: HW1 suffices for classical Bayesian optimisation and instrument control; HW2/HW3 for local LLM agents and vision models. The laboratory hardware dominates cost, not the compute.

Path D — Custom training

Worth it when you operate a long-running, repeated experimental campaign in one domain: train domain surrogates on your own experiment history, add vision models for liquid and crystal state recognition, or fine-tune an LLM on your protocol corpus. Needs hundreds to thousands of well-annotated runs with negative results included.

State of the art (2026)

• Algorithms: batch, multi-objective, constrained Bayesian optimisation with mixed (categorical and continuous) spaces is standard; transfer learning across campaigns is active research.
• Agents: LLM agents that read literature, write protocols and operate instrument APIs are now demonstrated in academic and startup systems; funding for "AI science factory" ventures such as Lila Sciences and Periodic Labs is large, but independent, peer-reviewed benchmarks of end-to-end autonomy remain scarce.
• Interoperability: SiLA 2 (SiLA consortium) and OPC UA LADS are converging standards for lab devices; vendor orchestration layers are moving toward open backbones [2].
• Limits and failure modes: robot reliability and calibration drift; sparse, noisy, biased data (publication bias, missing failed runs); surrogate models misled by measurement artefacts; reward hacking when the objective omits cost, safety or reproducibility; LLM hallucinated protocol steps or unit errors; narrow generalisation, since each new chemistry needs new hardware and analytics. Capital cost and integration effort are the real bottlenecks.

Shared resources & combinations

• Upstream design: SCI-002 (drug candidates), SCI-003 (materials), SCI-004 (retrosynthesis), SCI-014 (formulation) propose candidates that an SDL validates experimentally.
• Analysis: SCI-018 (spectral interpretation), SCI-008 (microscopy and image analysis).
• Agentic layer: SCI-011 (AI scientist agents) for hypothesis generation; ROBO-003 and ROBO-010 for robot programming and simulation.
• Shared stack: optimisation-solver, tabular-ml, agent-framework and workflow-automation are reusable; a single HW1/HW2 server can run planner, orchestration and LLM for several workcells.

Effort, cost & KPIs

• Path A: 3-12 months to a productive workcell; hardware from tens of thousands of euros for a liquid handler to seven-figure sums for multi-instrument platforms (vendor quotes required).
• Path B/C: 6-18 months for a custom closed loop, with a team of an automation engineer, a data scientist and domain scientists. Cost drivers: hardware, integration, consumables, maintenance, validation.
• KPIs: experiments to reach target vs. DoE/manual baseline, uptime and unattended hours, failed-run rate, cost per data point, time per design-make-test-analyse cycle, share of runs with complete FAIR metadata, reproducibility across replicate runs.

People & change

• Involve early: management (investment is large), lab and R&D leads, lab technicians and scientists as daily users, IT and IT security for instrument networks, data owners of ELN/LIMS, legal/compliance for GxP, safety and IP questions.
• Works council: depends. Lab robots and software that log which technician ran what can be suited to monitor behaviour and performance (§87(1) no. 6 BetrVG, BAG: objective suitability); changed work processes trigger information under §90(1) no. 3. See GUIDE-004. Public research: Personalrat.
• Data protection: mostly no personal data; staff logins and audit trails are personal. Involve the DPO and sign an AVV with cloud vendors (GUIDE-005).
• AI literacy: under AI Act Art. 4 (since 27 Jul 2026) employer and provider must take measures to support staff AI literacy; plan a short training (see GUIDE-006).
• Pilot and communication: start with one workstation and one optimisation task, run it parallel to manual work, and involve technicians in protocol design (GUIDE-009, GUIDE-008).
• Typical concerns: job loss for technicians (address: shift to supervision and method development), safety of unattended runs.
• Human control: scientists set objectives and safety limits, approve new protocol classes and can stop the loop at any time.

Risks, regulation & compliance

• EU AI Act: typically minimal risk; AI literacy (Art. 4) still applies. Not high-risk unless output feeds a regulated product decision (e.g. medical device or safety component), in which case the product rules apply. Scientific research exemption (Art. 2(6)) covers systems developed solely for research.
• GDPR: normally no personal data; operator logs and cloud-lab accounts contain staff data; human samples or patient-derived material trigger GDPR and ethics-committee rules.
• Machinery and safety: Machinery Regulation (EU) 2023/1230 for robots and workcells, CE marking, risk assessment, interlocks, chemical handling rules (CLP, REACH, workplace hazardous-substance rules).
• Dual-use and biosecurity: autonomous design and synthesis of hazardous chemicals or pathogens is a misuse risk; apply allow-lists of reagents, screened sequences and approval gates, and follow dual-use export-control rules (Regulation (EU) 2021/821).
• GxP: in pharma, data integrity (ALCOA+), computerised-system validation (EU GMP Annex 11, 21 CFR Part 11) and audit trails apply to automated records.
• Mitigations: hard-wired safety limits, simulation/dry-run before execution, versioned protocols, full provenance, independent verification of key findings.

Getting started

1. Pick one well-defined optimisation problem (3-10 parameters, fast and quantitative readout).
2. Audit ELN/LIMS and instrument interfaces; make data machine-readable and FAIR first.
3. Automate one workstation end to end and demonstrate repeatability without AI.
4. Add a Bayesian optimisation loop with safe bounds and compare against DoE on the same budget.
5. Introduce an LLM assistant for protocol drafting only after the loop is stable, with approval gates.

Sources

1. Nature Reviews Chemistry (2026) — The past, present and future of self-driving laboratories — https://www.nature.com/articles/s41570-026-00847-2
2. Bruker — Chemspeed and SciY announce self-driving laboratory platform (2026) — https://ir.bruker.com/press-releases/press-release-details/2026/Chemspeed-and-SciY-Announce-SelfDriving-Laboratory-Platform-Integrating-Automation-Analytics-and-AI-Orchestration/default.aspx
3. BoTorch — Bayesian optimization in PyTorch — https://botorch.org/
4. Emerald Cloud Lab — https://www.emeraldcloudlab.com/
5. SiLA consortium — SiLA 2 standard — https://sila-standard.com/
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Regulation (EU) 2023/1230 (Machinery Regulation) — https://eur-lex.europa.eu/eli/reg/2023/1230/oj