A pulse on self-driving labs

90-second orientation

First, separate automation from autonomy.

Automation performs a procedure. Autonomy uses evidence to decide what happens next. A self-driving laboratory needs both, with boundaries that scientists can see and change.

Automation

“Run this protocol.”

The steps are defined in advance. The machine may be very capable without choosing the next experiment.

Autonomy

“Use the result to choose.”

The system senses state, evaluates evidence and selects or adapts an action inside a defined authority boundary.

Check yourself. A robot runs a fixed protocol and emails the result. Is it autonomous? No. The loop closes only when the result can change the next experimental action.

Plan

Question → target → experiment

Execute

Robots, instruments and sample logistics

Measure

Sensors, assays and quality control

Learn

Analysis → model update → next experiment

What the room saw. measurement is often the most mature link; physical execution and exception recovery remain the weakest. The field is advancing, but not evenly.

Paper result

Paper 1 • Cooper et al., 2025

A systems map for AI and robotics in natural-science laboratories

This ICRA workshop paper gives the discussion a map of the whole system. It draws attention to dependencies that a single-device demonstration can hide.

Integration & scale

Heterogeneous devices, legacy software and data systems must act like modules rather than one-off engineering projects.

Digital twins

Models of devices and workflows can support simulation, commissioning, diagnosis and explanation. Their fidelity has to be tested against the lab.

Levels of autonomy

Ask what decisions the system owns, under which conditions, and what role people retain. “Human in the loop” is a deliberate design choice.

Foundation models

Language and multimodal models can read, translate and interpret. Classical ML, optimization and control remain essential.

Standardization

Shared interfaces, data structures and laboratory objects reduce integration cost. They work only when vendors and institutions adopt them.

Responsible operation

Reproducibility, safety, sustainability, dual use and workforce skills belong inside the architecture.

Biggest strengths

Defines autonomy as a socio-technical system.
Puts humans, safety and standards beside throughput.
Explains why reliability compounds across long workflows.
Avoids treating foundation models as the only intelligence that matters.

Biggest weaknesses

Expert synthesis rather than systematic evidence.
Limited quantification by domain or lab size.
Economics, maintenance and change management need more empirical treatment.
“Democratization” remains easier to state than to deliver.

The scientific opportunity is a better division of work, with human judgement used where it earns its keep.

Paper result

Paper 2 • Szymanski et al., Nature 2023

A-Lab closes the loop for inorganic materials

A-Lab linked computed targets, literature-derived recipes, robotic solid handling, furnaces, X-ray diffraction, machine interpretation and active learning for a 17-day autonomous campaign.

Worked example. A first recipe produces no useful target phase. The diffractometer records that failure, the analysis updates the evidence and ARROWS³ chooses a different route. That change of action closes the loop.

Choose

Computed candidate targets

Synthesize

Powder dosing, mixing and heating

Measure

Automated X-ray diffraction

Adapt

Phase analysis and ARROWS³

The denominator changes the story

Confirmed targets

63%

Inconclusive

Not obtained

30%

Recipes producing target

105

30%

Target-level result: 36/57 confirmed. Recipe-level result: 105/353 produced the target. Both are true; they answer different questions.

Why it matters

A-Lab integrated the full evidence loop instead of automating one task. It also turned failures into scientific evidence: slow kinetics, model error, volatility, amorphous products and metastability became diagnoses rather than invisible waste.

What it does not prove

There was no matched human or random-search baseline. Majority target phase is not necessarily a pure, useful material. Seventeen days is not long-term operations, and one inorganic platform is not a template for every wet lab.

A common trap. Whenever a system is called autonomous, ask: autonomous over which decisions, for how long, with what preparation, under what conditions, and with what recovery path?

Meeting pulse

BiB pulse on lab automation

The room supplied the friction coefficients.

These are anonymized, paraphrased themes from the 23 September 2026 discussion. They record what this group saw; they are not a representative survey.

Layer	What the papers anticipate	What practitioners encounter	Useful response
Planning	AI proposes experiments and mines literature.	A plan fails if it does not know local inventory, SOPs, device limits or authorization.	Ground models in machine-readable capabilities and explicit uncertainty.
Execution	Flexible robots run complex protocols.	Samples arrive in odd containers; consumables vary; devices jam; assays change.	Normalize inputs, expose device state, design retries and exception queues.
Measurement	Automated analytics generate reproducible evidence.	Queues, calibration drift, ambiguous signals and hand-offs still dominate.	Inline QC, orthogonal evidence, calibration checks and provenance.
Learning	Active learning chooses informative next experiments.	Objectives are unstable, multi-objective or tacit; data are sparse and biased.	Bound the search and preserve scientist-defined stop rules.
Digital twins	Simulation helps plan and explain the lab.	A twin of discovery biology can be confidently wrong.	Start with scheduling, device state, error injection and process scale-up.
Standards	Common interfaces enable modularity.	Legacy devices, partial implementations and vendor lock-in persist.	Standardize capability and data layers; maintain tested adapters.
People	Humans retain creativity and oversight.	Trust turns on agency, space, roles, maintenance and mid-run intervention.	Co-design authority, escalation and evidence review.
Safety	Automation removes people from hazards.	AI intent, translation and biased learning create additional failure modes.	Use interlocks, authorization, monitoring, safe states and auditable evidence.

What the community added

Six observations from people working with the gap

These blocks collect recurring ideas from the discussion. They are anonymised and paraphrased, with no claim that every participant held the same view.