Self-Driving Laboratories

• Self-Driving Laboratories are autonomous platforms that combine modular hardware, advanced computing, and AI to conduct iterative, closed-loop scientific experiments.
• They integrate standardized automation hardware with orchestration software and decision-making algorithms to optimize experimental workflows efficiently.
• SDLs accelerate discovery in materials science, chemistry, biology, and engineering by enabling rapid, reproducible experimentation with minimal human intervention.

A self-driving laboratory (SDL) is an autonomous platform that integrates automation hardware, advanced computing, and artificial intelligence to execute closed-loop scientific workflows. SDLs iteratively propose, carry out, and analyze experiments with the goal of accelerating discovery in domains including materials science, chemistry, biology, and engineering. Distinct from conventional high-throughput or materials acceleration platforms, SDLs explicitly couple physical automation (such as robotic sample handling and instrumentation) with autonomous decision-making algorithms, resulting in minimal human intervention throughout experiment planning, execution, data analysis, and iteration (Maffettone et al., 2023, Adesiji et al., 8 Aug 2025, Ginsburg et al., 2023).

1. Architecture and Core Components

SDLs employ a modular architecture, abstracting both hardware and software layers for flexibility, scalability, and reusability. Core components include:

• Automation hardware: Modular instruments (pipetting robots, synthesizers, plate handlers, mobile manipulators), robotic arms, sensors, and dispensers. Hardware is frequently encapsulated in standardized “modules” that expose uniform interface methods (e.g., device actions such as pipette, transfer, measure) (Ginsburg et al., 2023, Vescovi et al., 2023).
• Orchestration and workflow software: YAML or JSON configuration files specify workcells (comprising sets of modules with defined spatial relationships) and workflows as sequences of module actions. Workflow executors execute these action sequences, logging metadata, statuses, and outputs (Vescovi et al., 2023, Ginsburg et al., 2023).
• Decision-making and AI: Algorithms propose new experiments based on prior results. Bayesian optimization, active learning, evolutionary/genetic algorithms, or reinforcement learning agents iteratively suggest experiment conditions to maximize objectives under resource constraints (Adesiji et al., 8 Aug 2025, Martin et al., 2022, Ginsburg et al., 2023).
• Data management and provenance: Automated data acquisition, real-time FAIR-compliant storage, and distributed compute integration. Provenance-aware systems routinely log all steps, outcomes, and metadata for reproducibility, auditability, and off-line analysis (Deucher et al., 24 Jun 2025, Maffettone et al., 2023).
• Digital twins and simulation: Digital replicas of both hardware configurations and processes provide environments for workflow debugging, vision model training, and simulation-augmented experimentation (Vescovi et al., 2023, Torresi et al., 17 Dec 2025).
• Cloud/HPC coupling: Remote compute and storage resources handle AI inference, model training, simulation, and workflow execution at scale, interconnected via APIs and workflow managers (e.g., Globus, funcX, nanoHUB workflows) (Vescovi et al., 2023, Deucher et al., 24 Jun 2025).

This modularity and abstraction allow hardware exchange (e.g., swapping pipettors or imaging devices) and relocation of applications between workcells without changing overall logic—configurations, not code, are typically updated (Ginsburg et al., 2023, Vescovi et al., 2023).

2. Autonomy, Optimization, and Experiment Planning

SDLs realize closed-loop autonomy by combining experiment proposal, execution, and analysis within unified feedback cycles.

• Experiment proposal: Algorithms generate new candidate conditions based on observed objective response data. Bayesian optimization with Gaussian process surrogates and Expected Improvement (EI) or Upper Confidence Bound (UCB) acquisition functions are widely used (Adesiji et al., 8 Aug 2025, Martin et al., 2022). Evolutionary algorithms may also be used, as in color-matching benchmarks, relying on selection, crossover, and mutation (Ginsburg et al., 2023).
• Formulation of the optimization problem: Experiment variables (e.g., reagent concentrations, temperature, genome edits) are optimized under box and sum constraints. Objectives may be scalar (e.g., yield, mobility, color closeness) or vector-valued for multi-objective settings (Ginsburg et al., 2023, Martin et al., 2022).
• Batching and parallelism: SDLs may execute experiments in serial or in asynchronous parallel pipelines, with multiple samples flowing through multi-stage processes (e.g., mixing, reaction, characterization). Parallelism increases throughput but introduces delayed feedback, requiring optimizers to account for “pending” experiments (Wen et al., 2023).
• Multi-stage and proxy-aware decision-making: Extensions to standard Bayesian optimization handle multi-stage workflows, where intermediate (proxy) measurements are available. Multi-stage BO with nested acquisition functions uses proxies to prune unpromising candidates, improving time- and cost-efficiency (Torresi et al., 17 Dec 2025).
• Handling constraints and failures: Hard constraints, penalty methods, or learned feasibility models filter infeasible or unsafe actions. Execution failures are logged as informative signals, updating feasibility surrogates and triggering corrective cycles (Chen et al., 25 Jan 2026).
• Metrics for benchmarking: Key metrics include acceleration factor (AF: reduction in experiment count to a target outcome vs. random or baseline), enhancement factor (EF: relative objective improvement at a fixed experiment count), and additional system-level measures such as time-without-humans (TWH) and commands-without-humans (CCWH) (Adesiji et al., 8 Aug 2025, Ginsburg et al., 2023).

SDLs therefore deliver both sample efficiency and system-level autonomy, with architectures able to support continuous, multi-day autonomous operation across a range of physical workflows.

3. Representative Applications and Domains

SDLs have been deployed across a spectrum of sciences and engineering, including:

• Materials discovery: E.g., optimization of thin-film hole mobility via modular robot platforms (SCARA arms, pipetting, deposition, inline optical/electrical measurement), achieving 3x mobility improvement in 30 hours and uncovering structure–performance trade-offs difficult to identify manually (MacLeod et al., 2019).
• Color-matching benchmark tasks: Fully automated pigment mixing to minimize color distance, illustrating end-to-end automation, device-independence (via modular YAML configurations), and optimization within hardware constraints (Ginsburg et al., 2023).
• Combinatorial synthesis and in-situ characterization: ML-guided composition mapping (e.g., co-sputtered thin films) with in-situ sensors, Bayesian active learning, real-time model updating, and calibration-free mapping in minutes (Jarl et al., 6 Jun 2025).
• Synthetic biology: Genome engineering SDLs (liquid handling, microfluidics, colony picking, multi-omics screening), autonomous Design–Build–Test–Learn (DBTL) cycles, and Bayesian optimization of pathway or microbiome design (Martin et al., 2022).
• Collaborative, distributed SDLs: Cross-lab digital twins, FAIR-compliant data repositories (e.g., ResultsDB), and on-demand, web-driven experiment suggestion and data ingestion for low-cost, replicable experiment pipelines (Deucher et al., 24 Jun 2025).
• Safety and error handling: Visual LLM–driven safety monitoring of robots and personnel (PPE, fire, accident detection) with real-time actionable feedback for hazard avoidance (Munguia-Galeano et al., 7 Aug 2025), and deep-learning–driven substrate manipulation error correction (Fontenot et al., 4 Dec 2025).
• Tool-using agentic systems: Cognitive multi-agent architectures (AutoLabs, AILA) that decompose complex experimental requests into controlled protocols, integrate tool-assisted calculations, and carry out iterative self-correction to yield hardware-ready instructions or conduct scientific instrument operation via language agents (Panapitiya et al., 30 Sep 2025, Mandal et al., 2024).

SDLs thus span domains from small-molecule chemistry to complex biological engineering and multimodal data-driven materials design.

4. Data Management, Provenance, and FAIR Principles

SDLs depend on rigorous, automated data-handling infrastructure:

• Automated logging: All experiment step metadata—parameters, timings, outputs, images—are logged automatically for full provenance (Ginsburg et al., 2023, Deucher et al., 24 Jun 2025).
• FAIR data infrastructure: SDLs leverage continuous, not just post-hoc, FAIR data practices, including schema-driven metadata capture, open APIs, and integration with persistent repositories (e.g., ResultsDB, ALCF Community Data Co-Op) (Deucher et al., 24 Jun 2025, Maffettone et al., 2023).
• Interoperability: Standard ontologies, OGC GeoTIFF conventions, and mappings to ORCID/ROR are used to integrate across labs and hardware (Deucher et al., 24 Jun 2025).
• Data-centric vision and error handling: Pipetting checkpoints and rare error detection (e.g., bubble detection) are achieved through real-plus-virtual data engines with human-in-the-loop curation and automated confidence-based routing, yielding 99.6% classifier accuracy (Liu et al., 1 Dec 2025).
• Digital twins: Synchronization of module/workcell specifications between real and virtual hardware allows seamless migration and testing, with virtual environments supporting workflow debugging and vision training (Vescovi et al., 2023).
• Quality assurance and continuous integration: Ongoing QC, version-tracked data models, and automated CI/CD pipelines for data schemas underpin reproducibility and reliability (Maffettone et al., 2023).

Such foundations promote auditability, enable reproducibility, and facilitate multi-site scientific collaboration.

5. Benchmarking, Evaluation, and Comparative Metrics

Proper evaluation of SDLs requires domain- and system-level benchmarks:

Metric	Definition	Example/Note
Acceleration Factor (AF)	Number of baseline experiments to solution divided by SDL experiment count	Median AF ≈ 6 in MS/chemistry (Adesiji et al., 8 Aug 2025)
Enhancement Factor (EF)	Best achieved outcome (SDL) divided by baseline after n experiments	Peaks at ~10–20 experiments/dimension
Time Without Humans (TWH)	Autonomous wall-clock time between required human interventions	Used in cross-facility benchmarks (Ginsburg et al., 2023)
Commands Without Humans (CCWH)	Number of device commands executed autonomously	387 commands in color-matching task (Ginsburg et al., 2023)
Throughput	Number of completed experiments per hour	15.6 wells/hour (color-matching) (Ginsburg et al., 2023)
Task-level Accuracy	Percent correct on benchmark suite operation domains (doc, analysis, calculation, hybrid)	Examples: 80–92% for GPT-4o (AFM) (Mandal et al., 2024)
Regret	Difference between achieved and optimal objective at iteration n	Lower regret indicates faster convergence
Error detection and correction	Ratio of auto-detected/corrected operation failures	e.g., 98.5% substrate placement (Fontenot et al., 4 Dec 2025)

SDLs are systematically benchmarked using standardized test cases (e.g., color-matching, composition mapping), public competition suites (e.g., AFMBench), and framework-level metrics that emphasize both sample efficiency (objective per trial) and operational reliability (autonomy, error recovery) (Adesiji et al., 8 Aug 2025, Ginsburg et al., 2023, Mandal et al., 2024).

6. Limitations, Challenges, and Opportunities

Despite substantial progress, SDLs face persistent technical, operational, and societal challenges:

• Data standardization and management: Proprietary formats, lack of robust ontologies, and insufficient application of FAIR principles remain barriers (Maffettone et al., 2023).
• Hardware–software integration: Non-standard APIs, sample-exchange complexities, and underdeveloped error-handling approaches constrain plug-and-play automation (Maffettone et al., 2023).
• Algorithmic challenges: Handling high-dimensional parameter spaces, noise, delayed feedback, and complex protocol branching are active research areas (Wen et al., 2023, Adesiji et al., 8 Aug 2025, Torresi et al., 17 Dec 2025).
• Scalability: Transitioning from single-system prototypes to distributed “science factories” supporting massive parallelism, robust supply/waste logistics, and multiple concurrent workflows remains a target (Vescovi et al., 2023, Maffettone et al., 2023).
• Human–AI collaboration: Enabling interactive, explainable, and safe agentic autonomy, with human-in-the-loop for constraint editing and reward shaping, is an unresolved area (Chen et al., 25 Jan 2026, Panapitiya et al., 30 Sep 2025, Mandal et al., 2024).
• Safety and compliance: Novel hazards (robot–fire interactions, PPE monitoring, laboratory accidents) demand specialized, vision-language–driven approaches (Munguia-Galeano et al., 7 Aug 2025).
• Education and training: Workforce development in modular SDL programming, digital twin curation, and data management is a community priority (Maffettone et al., 2023).

Research directions prioritize multi-modal/uncertainty-aware representation learning, dynamic protocol optimization with constraints, failure-aware adaptation, and shared community benchmarks/DDFs with precise reproducibility specifications (Chen et al., 25 Jan 2026).

7. Generalization and Future Outlook

SDL architectures are widely extensible. Core software can drive diverse hardware platforms by swapping compliant modules; digital twins enable migration between physical and virtual configurations for rapid prototyping and training; and collaborative frameworks (e.g., distributed FAIR databases, nanoHUB workflows) facilitate community-sourced learning and optimization (Vescovi et al., 2023, Deucher et al., 24 Jun 2025). As standardization, modularity, and interoperability improve, SDLs are poised to become foundational infrastructures for accelerated scientific discovery and machine-driven hypothesis generation across scales and disciplines.

Concrete best practices include: embracing fully declarative module/workcell interfaces (Vescovi et al., 2023, Ginsburg et al., 2023), enforcing rigorous data provenance (Deucher et al., 24 Jun 2025, Liu et al., 1 Dec 2025), embedding advanced AI/ML in experiment planning (Adesiji et al., 8 Aug 2025, Martin et al., 2022), and developing robust benchmarking and error-handling protocols (Wen et al., 2023, Munguia-Galeano et al., 7 Aug 2025, Fontenot et al., 4 Dec 2025).

SDLs represent a transformative convergence of automation, data-centric modeling, and agentic AI, with ongoing progress fundamentally linked to modular design, scalable orchestration, and open community infrastructure (Maffettone et al., 2023, Ginsburg et al., 2023, Adesiji et al., 8 Aug 2025, Vescovi et al., 2023, Chen et al., 25 Jan 2026).