- Most self-driving labs operate at Level 3, requiring human oversight for anomalies.
- No laboratory has yet achieved Level 5 full autonomy.
- Berkeley's A-Lab controversy shows speed gains can come at the cost of analytical accuracy.
Ross King saw something that bothered him. The laboratories at his university sat empty at night, expensive equipment idle while scientists slept.
Forty years ago, that observation launched his quest to build machines that could run experiments around the clock - which eventually grew into what we now call "self-driving laboratories".
His robot scientist Adam became the first machine to autonomously discover scientific knowledge in 2009. Now King is building Genesis, designed to run a thousand parallel experiments daily.
But a comprehensive review published in 2025 by Royal Society Open Science reveals an uncomfortable truth about the field King helped create: most self-driving laboratories operate at Level 3 autonomy - conditional, not full.
The machines can run experiments, but they still need humans when anything unexpected happens.
What is a self-driving laboratory?
A self-driving laboratory combines robotics and artificial intelligence to run scientific experiments with minimal human intervention. Robots handle physical tasks like mixing chemicals or moving samples, while AI decides what experiments to try next based on results. Think of it as a feedback loop: the machine runs an experiment, analyzes outcomes, and plans the next step - potentially around the clock.
Five Levels of Laboratory Autonomy
The review, authored by Alexander Tobias and Adam Wahab at the MITRE Corporation, adapts the automotive industry's autonomy scale to scientific laboratories. The framework runs from Level 1 (assisted operation) through Level 5 (full autonomy requiring no human intervention).
| Level | Name | Description |
|---|---|---|
| 1 | Assisted operation | Machine assistance with laboratory tasks |
| 2 | Partial autonomy | Proactive scientific assistance, e.g. protocol generation |
| 3 | Conditional autonomy | Minimum to qualify as an SDL. Autonomous performance of at least one cycle of the scientific method. Requires human intervention for anomalies |
| 4 | High autonomy | Hypothesis tester capable of automating protocol generation, experiment execution, data analysis and results-driven hypothesis adjustment |
| 5 | Full autonomy | Full automation of the scientific method (AI researcher) |
The review found no laboratory has achieved Level 5.
The classification cuts through marketing language that conflates automation with autonomy. A system that pipettes liquids automatically is not the same as one that formulates hypotheses independently. Most current self-driving laboratories excel at optimization - finding the best parameters within a defined space - rather than genuine discovery.
This distinction matters because the term "self-driving" implies capabilities these systems do not possess. Like autonomous vehicles that still require human oversight, these laboratories handle routine operations well but falter when conditions deviate from expectations.
Key figure
Level 3
Where most "self-driving" labs actually operate - conditional autonomy requiring human oversight for unexpected situations
Berkeley's A-Lab and the Quality Question
The tension between speed and quality surfaced dramatically in the debate over Berkeley's A-Lab, which claimed to synthesize 41 novel materials in 17 days with a 71 percent success rate.
Subsequent analysis challenged that claim. Robert Palgrave, examining the lab's crystallographic analysis, concluded the work was inadequate. "The Rietveld refinement was very bad, very beginner, completely novice human level," he told Chemistry World in a 2024 interview. "That led to them misidentifying things in some cases."
Gerbrand Ceder, the principal scientist behind A-Lab, acknowledged the limitations while defending the project's goals. "We have no doubt that a human can perform a higher-quality refinement on these samples," he wrote. "However, it was our objective to show what an autonomous laboratory can achieve."
The exchange touches a fundamental question: faster processing means little if the analysis cannot reliably identify what was actually made.
Collaborators, Not Replacements
Milad Abolhasani, the ALCOA Professor of Chemical and Biomolecular Engineering at North Carolina State University, frames the technology's purpose differently than the hype suggests.
Imagine if scientists could discover breakthrough materials for clean energy, new electronics, or sustainable chemicals in days instead of years.
Milad Abolhasani, Alcoa Professor & University Faculty Scholar, NC State University
"Self-driving labs will serve as collaborators for human researchers, significantly reducing the time and cost required to reach scientific solutions," he told Phys.org. "They will not replace the unique expertise or creativity of human researchers."
His team recently demonstrated a dynamic flow technique that collected ten times more data than conventional approaches in tests on quantum dot synthesis - like switching from a single snapshot to a full movie, as he describes it.
The method addresses a practical bottleneck: gathering enough quality data to train the machine learning systems that guide automated experiments.
Will self-driving laboratories ever reach level 5 and become fully autonomous? As of today, no lab has managed to reach higher than level 3, which means human oversight is still needed. (Science Reader).
Patent Law Has Not Caught Up
Beyond technical limitations, self-driving laboratories face a legal obstacle. Patent offices worldwide require human inventors. Courts in the United States, United Kingdom, European Union, and Australia have rejected applications naming the AI system DABUS as inventor.
This creates a peculiar situation. A self-driving laboratory might discover a valuable new material, but if no human made the key inventive contribution, the discovery may be legally unpatentable. South Africa granted one DABUS patent, suggesting the legal landscape could shift.
For now, the gap between what machines can discover and what can be commercially protected remains unresolved.
The Field Moves Forward Cautiously
The U.S. Department of Energy launched AMP2 in December 2025, an 1,800-square-foot autonomous-capable platform for microbial research at Pacific Northwest National Laboratory. A larger facility, spanning 32,000 square feet with over 100 instruments, is planned for 2030.
Energy Secretary Chris Wright pressed laptop keys rather than cutting a ribbon to inaugurate the system - a small ceremony that acknowledged the technology's nature. "Humans drive science and innovation," Wright said at the launch. "They always have and they always will."
The Royal Society classification provides a framework for evaluating claims as the field develops. When a laboratory announces "autonomous" capabilities, the relevant question becomes: which level?
The answer, for now, is usually three.
Sources
- Primary Research: Tobias, A.V. and Wahab, A. (2025). Autonomous self-driving laboratories: a review of technology and policy implications. Royal Society Open Science, 12(7):250646. https://pmc.ncbi.nlm.nih.gov/articles/PMC12368842/
- Additional Context:
- New analysis raises doubts over autonomous lab's materials discoveries (Chemistry World)
- Q and A: Self-driving labs and the era of scientific acceleration (Phys.org)
- Fast forward for self-driving labs (NC State News)
- Energy Department launches AI-driven biotechnology platform (U.S. Department of Energy)
Fact Check: Claim-by-Claim Verification Verified
The article accurately represents the key findings of the Royal Society review and correctly characterizes the state of self-driving laboratory technology, despite using common science communication conventions of simplification and dramatic framing.
Commentary
- The article characterizes current self-driving labs as operating at "Level 3" while noting they excel at "optimization rather than genuine discovery." The Royal Society review distinguishes between Level 3 (conditional autonomy with human intervention for anomalies) and Level 4 (high autonomy including hypothesis testing). The article's characterization is simplified but not inaccurate for general audiences—most deployed systems do cluster around Level 3, though some notable examples (Adam, Eve, Genesis) operate at Level 4 according to the source
- The article's discussion of Milad Abolhasani's quote about "collaborators" is accurate to the Phys.org article cited, though the framing emphasizes the collaborative role over autonomous discovery capabilities
- The 71% success rate for Berkeley's A-Lab is correctly reported as what the original paper claimed, and the article appropriately contextualizes this within the subsequent critique—accurately representing both the claim and the controversy without endorsing either
Sources used for verification
Academic/Peer-reviewed:
- Autonomous 'self-driving' laboratories: a review of technology and policy implications - PMC/NIH
- Adam discovers yeast functional genomics autonomously - Nature (2009)
Other reliable sources:
- New analysis raises doubts over autonomous lab's materials discoveries - Chemistry World
- Q and A: Self-driving labs and the era of scientific acceleration - Phys.org
- Fast forward for self-driving labs - NC State University News
- Energy Department launches AI-driven biotechnology platform - U.S. Department of Energy
Fact-checked by Perplexity Sonar Pro on 2026-01-23
