self-driving labs pharma · self-driving laboratories
Self-Driving Labs in Pharma: Closed-Loop R&D That Works
August 5, 2026
Updated August 5, 2026
35 min read
A 2026 analyst guide to self-driving labs in pharma R&D, covering closed-loop experimentation, the AI methods behind it, 12+ named vendors, AstraZeneca's iLab, and Insilico Medicine's autonomous discovery pipeline.
- 01A 2025 preprint found a median 'acceleration factor' of 6 across self-driving lab studies, but as an unreviewed preprint this multiplier should be treated with caution.
- 02UC Berkeley's A-Lab, the most scrutinized SDL case study, received a formal Nature correction in January 2026 after outside chemists challenged its novelty claims.
- 03GMP's expectation of fixed, validated procedures conflicts with continuously learning AI models, a tension buyers should resolve with regulatory affairs before scaling.
- 04Global lab automation market-sizing estimates diverge by billions of dollars across research firms, so vendor market-size claims should be checked against methodology.
- 05Recursion's REC-1245 moved from target identification to IND-enabling studies in under 18 months while synthesizing roughly 200 compounds, an early disclosed pharma SDL result.
- 06A peer-reviewed UCL study found a self-driving formulation robot tested 7 times as many formulations as a human formulator in just 25% of the human time, supporting the sample-efficiency case for Bayesian optimization.
Executive Summary
A self-driving laboratory (SDL) is a closed-loop research system in which robotics execute experiments, sensors capture the results, and a machine-learning model decides what to try next, with little or no human intervention in the decision loop. Nature Synthesis defines an SDL as "a machine-learning-assisted modular experimental platform that iteratively operates a series of experiments selected by the machine learning algorithm to achieve a user-defined objective" ([1]), and a 2026 Nature Reviews Chemistry survey describes the field as having "progressed from narrowly focused automation tools to multipurpose discovery platforms in which algorithms propose, execute and interpret experiments with limited human intervention" ([2]). That is the essential distinction from traditional pharma lab automation: liquid handlers and high-throughput screeners execute a fixed protocol designed by a human, while an SDL closes the design-make-test-analyze (DMTA) loop itself, using Bayesian optimization, active learning, reinforcement learning, or large language model (LLM) agents to choose the next experiment ([3]) ([4]).
Commercially, the market is still small relative to pharma R&D spending overall but growing quickly. Grand View Research sizes the global lab automation market at $8.27 billion in 2024, projected to reach $18.39 billion by 2033 ([5]), while MarketsandMarkets projects the narrower AI-in-drug-discovery segment, valued at $3.92 billion in 2025, to grow from $5.09 billion in 2026 to $17.56 billion by 2031 ([6]). McKinsey estimates investors have put more than $50 billion into over 500 AI-driven R&D companies since 2015 ([7]), a pool that includes cloud-lab operators like Emerald Cloud Lab and Strateos, robotics vendors like Automata, Opentrons, and Chemspeed, and newer autonomy-focused entrants such as Flagship Pioneering's Lila Sciences ($200 million in committed seed capital) ([8]) and Medra ($52 million Series A) ([9]). This spending is set against a backdrop in which the Tufts Center for the Study of Drug Development's benchmark estimate puts full development cost at $2.6 billion per approved drug ([10]) and clinical success rates from Phase I have averaged 7.9% across the industry ([11]).
Pharma-specific deployments are already producing disclosed, if early, results. Recursion Pharmaceuticals says its REC-1245 program moved "from target identification to IND enabling studies in under 18 months" while synthesizing roughly 200 compounds ([12]). AstraZeneca's iLab program, running since 2017, aims to "identify potential drug candidates in half the time it takes today" ([13]). Insilico Medicine's Life Star robotics facility and its newer LabClaw autonomy system have supported a wholly-owned pipeline of 30 assets, 10 of which have received Investigational New Drug (IND) clearance ([14]). Eli Lilly and NVIDIA announced a joint investment of up to $1 billion over five years in January 2026 to connect Lilly's "agentic wet labs" with computational dry labs for round-the-clock experimentation ([15]), and Merck (MSD) struck a multi-year, up-to-$1-billion agentic-AI partnership with Google Cloud spanning R&D, manufacturing, and commercial functions ([16]).
The academic literature is more measured than vendor press releases. A 2025 preprint benchmarking review of self-driving labs reported a median "acceleration factor" of 6 across the studies it reviewed, with wide variation; this finding has not been established through peer review ([17]). The most publicized case, UC Berkeley's A-Lab, received a formal Nature correction in January 2026 after outside chemists challenged its novelty claims, with the authors acknowledging their "original claims of material novelty were subject to misinterpretation" ([18]). Barriers to wider pharma adoption include data-standardization gaps, a lack of interoperability protocols such as SiLA ([19]), and a tension between continuously learning AI models and the fixed, validated procedures that Good Manufacturing Practice (GMP) regulation expects, which one peer-reviewed regulatory analysis calls a direct conflict between "continuous learning and opacity of decision-making processes" and "the principles of GMP" ([20]). For pharma and life-sciences organizations evaluating this space, the practical task is not choosing a single winning platform but sequencing a defensible pilot, and independent advisors such as intuitionlabs.ai, a life-sciences and AI consultancy, frame their role as helping clients conduct exactly this kind of "technology assessment" before committing capital to closed-loop infrastructure ([21]).
Introduction and Background
Pharmaceutical research and development has long followed an iterative cycle: design a candidate molecule, make it, test it, analyze the result, and use that analysis to design the next candidate. Industry shorthand calls this the design-make-test-analyze (DMTA) cycle, and a 2025 preprint on lab-automation agents describes it as "an iterative framework that drives the optimization of drug candidates from initial concept to clinical development" ([3]). For decades, each turn of that cycle has depended on a bench scientist to interpret data and choose what to try next, a bottleneck that shows up directly in industry economics: the Tufts Center for the Study of Drug Development's widely cited 2014 study put the fully capitalized cost of bringing a new drug to market at $2,558 million in 2013 dollars ([10]), a figure Nature Reviews Drug Discovery confirmed was up 145% from the prior 2003 estimate ([22]). Deloitte's most recent "Measuring the Return from Pharmaceutical Innovation" study found average research and development (R&D) cost per approved asset reached $2.23 billion in 2024 among the top 20 biopharma companies ([23]), even as the industry's internal rate of return on R&D improved to 7.0% in 2025, its third consecutive year of gains ([24]).
Against that backdrop, a distinct category of laboratory technology, the self-driving laboratory (SDL), has moved from academic materials-science demonstrations toward pharma-relevant deployment. Nature Reviews Chemistry frames the shift plainly: SDLs "merge autonomous experimentation, advanced reactor engineering, robotics and artificial intelligence to accelerate scientific knowledge creation" ([25]), and a companion 2025 review in the same journal calls the shift "a paradigm shift in scientific research" analogous to autonomous vehicles ([26]). The concept is not brand new: a 2010 peer-reviewed paper describing the "Robot Scientist" concept, and the specific system named Adam, noted that such a system "generates hypotheses from a computer model of the domain, designs experiments to test these hypotheses, runs the physical experiments using robotic systems, analyses and interprets the resulting data, and repeats the cycle" ([27]), and Adam itself had already identified twelve genes in yeast metabolic pathways ([28]). What changed over the following decade was the maturity of machine learning for experiment planning, cloud-accessible robotics, and, most recently, large language model (LLM) agents capable of writing and executing their own experimental protocols.
This report is written for pharmaceutical, biotech, and life-sciences leaders evaluating whether, and how, to invest in self-driving lab technology as of August 2026. It defines the term precisely and distinguishes it from conventional lab automation, walks through the robotics, software, and AI methods that make up an SDL, surveys the commercial vendor landscape, documents how major pharmaceutical companies and AI-native biotechs are deploying closed-loop experimentation today, and reviews the data on cost, market size, and measured outcomes, including the more skeptical academic findings that vendor marketing tends to omit.
Global lab automation market size in 2024, per Grand View Research
Time Recursion's REC-1245 took from target ID to IND-enabling studies
Target compounds A-Lab synthesized in 17 days of unattended operation
Formulations tested by a self-driving robot vs. a human formulator
What Is a Self-Driving Laboratory? Definitions, Taxonomy, and the Comparison to Traditional Automation
Core Definition and the Autonomy Spectrum
The most cited technical definition of a self-driving laboratory (SDL) comes from a 2023 Nature Synthesis review, which states that "an SDL is a machine-learning-assisted modular experimental platform that iteratively operates a series of experiments selected by the machine learning algorithm to achieve a user-defined objective" ([1]), a definition the same paper visualizes with a dedicated figure titled "Conventional versus self-driving labs" ([29]). A 2025 peer-reviewed review in Royal Society Open Science goes further, arguing that "today's most capable SDLs automate nearly the entire scientific method, from hypothesis generation, experimental design, experiment execution and data analysis, to drawing conclusions and updating hypotheses" ([30]). That review, and the broader literature it surveys, treats "closed-loop experimentation" as a near-synonym for the SDL concept, listing it alongside "cloud laboratories" and "autonomous science" as core descriptive terms for the field ([31]).
For pharma specifically, a 2026 Nature Synthesis perspective is direct about the relevance: "Self-driving laboratories, which incorporate robotics, advanced sensing, automated feedback control and artificial intelligence, are revolutionizing the conduct of biopharmaceutical research" ([32]). The same paper describes the mechanism precisely: "By coupling high-throughput experimentation with real-time data analytics, these autonomous systems enable rapid exploration of chemical space and adaptive optimization that far outpaces conventional workflows" ([33]), and illustrates a pharma-relevant application through an "autonomous crystallization SDL platform for closed-loop experimentation" ([34]).
Not every autonomous system is created equal, and the literature has begun formalizing a spectrum. The Royal Society Open Science review positions "robotic liquid handlers" and "data analysis software" as tools of the lowest, "assisted operation" rung of laboratory autonomy ([35]), while a 2026 review in Applied Physics Express defines true SDLs by their closed loop: they "integrate (i) machine-learning-based inference and selection of the next experimental conditions with (ii) experiments using robots" ([36]). A 2025 peer-reviewed UK study of a self-driving pharmaceutical formulation robot puts the distinction even more concretely, noting that "automation has widely been used to increase the throughput of established assays or manufacturing processes" while a self-driving system instead "interprets the results of those experiments, predicts which experiment to perform next, then executes that experiment" ([37]) ([38]).
Self-Driving Labs Versus Traditional Lab Automation
This is the direct answer to how self-driving labs differ from traditional lab automation: a traditional automated lab, whether a liquid-handling robot, a high-throughput screener, or a barcode-tracked sample store, executes a protocol that a human designed in advance and does not change its own plan based on intermediate results. A self-driving lab closes that loop: software interprets each result and autonomously selects the next experiment, using an algorithm rather than a person to make the decision. The Chemical Reviews survey of the field frames the payoff of this distinction plainly, noting that SDLs "hold the potential to greatly accelerate research in chemistry and materials discovery" precisely because they combine automated execution with autonomous planning rather than automated execution alone ([39]). A trade-press account of pharma SDL adoption describes the legacy state of affairs bluntly: in conventional process development, "experimentation is often labor-intensive and sequential, data are captured inconsistently, and decisions rely heavily on precedent and expert intuition" ([40]).
The AI and Machine Learning Methods That Close the Loop
Several distinct machine-learning approaches are used to make the "next experiment" decision, and understanding which one a given platform uses matters because it determines cost, speed, and applicability:
- Bayesian optimization is the most common method precisely because it is sample-efficient: the Applied Physics Express review explains that it "can identify near-optimal solutions with fewer experiments than random search" ([4]), which matters directly in pharma where each wet-lab run consumes scarce compound and instrument time.
- Active learning selects which data point (or experiment) would most reduce model uncertainty next, and is grouped with Bayesian optimization in a 2026 survey as a method "for sample efficient experiment selection" ([41]).
- Reinforcement learning (RL) and planning extend the loop across multi-step protocols; the same survey describes RL and planning as tools "for long horizon protocol optimization" ([42]), extending Bayesian methods to sequences of linked experimental steps rather than single-shot optimization.
- Large language model (LLM) agents are the newest addition, described in the Applied Physics Express review as a source of "expanded autonomy via large language model-based artificial intelligence (AI) agents" ([43]), grounded in a landmark demonstration discussed in the Case Studies section below: Carnegie Mellon's Coscientist system, which used GPT-4 to design and execute its own chemistry experiments ([44]).
None of these methods is a free lunch. A 2026 Digital Discovery paper identifies a specific limitation of the workhorse method: "standard Bayesian optimisation relies on fixed experimental workflows with predefined parameters and objective functions" ([45]), meaning it cannot easily adapt mid-run if intermediate results suggest a different protocol entirely, a real constraint for the branching decision trees typical of drug synthesis.
The Architecture of a Self-Driving Lab: Robotics, Software, and Data Pipelines
Liquid handlers, robotic arms, automated synthesis reactors, and analytical instruments linked by conveyors or mobile robots form the physical base.
Schedules instrument time, routes samples, and exposes a programmable interface so an AI planner can submit an experiment as code.
The Bayesian optimization, active learning, reinforcement learning, or LLM-agent layer that most differentiates a self-driving lab from a merely automated one.
Structured, machine-readable data must flow back from every instrument; this is widely regarded as the weakest link in the stack.
When robotics, orchestration, AI decision-making, and standardized data all interoperate, the system closes the design-make-test-analyze loop autonomously.
Widespread adoption remains limited by high costs, complex infrastructure, and limited accessibility.
An operating SDL is not a single machine; it is a stack of at least four layers that must work together, and each layer has its own commercial ecosystem, discussed by category below.
Robotic execution hardware. At the base sits the physical apparatus: liquid handlers, robotic arms, automated synthesis reactors, and analytical instruments (liquid chromatography-mass spectrometry, nuclear magnetic resonance, X-ray diffraction) linked by conveyors or mobile robots. UC Berkeley's A-Lab, one of the most detailed public descriptions of this layer, runs three robotic arms and eight furnaces working from roughly 200 powder precursors inside a 600-square-foot lab that operates around the clock ([46]), testing between 100 and 200 samples per day ([47]).
Orchestration and lab-management software. This layer schedules instrument time, routes samples, and exposes a programmable interface so an AI planner (or a human scientist) can submit an experiment as code rather than a pipetting instruction sheet. Emerald Cloud Lab's own description of the category states that "cloud labs are fully software controlled, highly automated life science laboratories" accessible over the internet ([48]). Independent software vendors compete in this layer as well: Artificial, Inc.'s aLab Suite counted Thermo Fisher and Beam Therapeutics as customers "using its software directly and in partnership" as of its 2021 Series A ([49]), while Synthace's Antha platform reported that "seven of the top 10 global pharmaceutical companies have also adopted Synthace's next-gen R&D cloud platform" by late 2021 ([50]).
The AI decision layer. As described above, this is the Bayesian optimization, active learning, reinforcement learning, or LLM-agent model that selects the next experiment. It is the layer that most differentiates a self-driving lab from a merely automated one, and it is where most of the recent academic and vendor activity is concentrated.
Data capture and standardization. Every layer above depends on structured, machine-readable data flowing back from instruments, and this is widely regarded as the weakest link. The Nature Reviews Chemistry survey names it explicitly as an open requirement: SDLs still need "interoperable data and metadata standards, modular and integrable experimental hardware, and trustworthy artificial intelligence agents" before they can mature further ([51]). The International Society for Pharmaceutical Engineering (ISPE) points to specific protocol work underway to close this gap, including "interoperability protocols, such as the Standardization in Lab Automation (SiLA)" ([19]).
Cost and complexity remain real barriers to assembling this stack. A 2026 Nature Synthesis paper introducing a lower-cost, modular self-driving platform states directly that "widespread adoption remains limited by high costs, complex infrastructure and limited accessibility" ([52]), a point echoed by the roughly $5,000 bill of materials that same platform, called RoboChem-Flex, was explicitly engineered to hit in order to democratize access ([53]).
The Vendor and Platform Landscape
The commercial market spans cloud-accessible remote labs, modular robotics vendors, orchestration software companies, and a newer wave of autonomy-native startups. Table 1 below summarizes the major named vendors, their business model, and disclosed financial scale as of August 2026, drawn from official vendor pages, funding announcements, and business press.
| Company | Founded / HQ | Model | Disclosed Funding or Scale | Pharma-Relevant Detail |
|---|---|---|---|---|
| Emerald Cloud Lab | 2010s, US | Fully remote, software-controlled "cloud lab" scientists operate over the internet ([48]) | Raised total financing of $13.5 million as of its 2014 launch, per Fierce Biotech ([54]); subscription plans start around $30,020/month on its live configurator ([55]) | Ran Carnegie Mellon's Coscientist experiment remotely; hosts a published biotech method-optimization case (Pragma Bio) |
| Strateos | Founded from Transcriptic/3Scan lineage, US | On-site and remote "Studio Lab" robotic cloud labs for synthesis, purification, and analysis | $56.1 million Series B (2021) led by DCVC and Lux Capital ([56]); two sites totaling over 14,000 sq ft and 200+ instruments ([57]) | Built and operated Eli Lilly's $90 million Studio Lab ([58]); ran weekly medicinal-chemistry assay cycles for Amgen ([59]) |
| Automata | UK | Modular lab robotics and "Automata Labs" operating-system software for AI-ready labs | $50 million Series B (2022, Octopus Ventures) ([60]); $45 million Series C (Jan. 2026, led by Dimension, with Danaher partnership) ([61]) | Deployed in NHS pathology labs running "40 fully automated stations 24/7" ([62]) |
| Opentrons Labworks | 2013, Brooklyn, NY | Benchtop liquid-handling robots (OT-2, Flex) with open-source protocol software | $200 million Series C (2021, SoftBank Vision Fund 2), valuing the company at $1.8 billion, up from $90 million a year earlier ([63]); OT-2 lists starting at $15,950 ([64]) | OT-2 used by "thousands of research organizations in more than 40 countries" ([65]) |
| Chemspeed Technologies | 1997, Fullinsdorf, Switzerland | Modular, "vendor agnostic self driving labs" for chemistry R&D and QC ([66]) | Acquired by Bruker Corporation (Nasdaq: BRKR) under a definitive agreement announced January 2024, closed March 2024 ([67]) | AstraZeneca reported its Chemspeed CATSCREEN 96 catalyst-screening system "did pay off itself within 6 months" ([68]); AbbVie published a high-throughput experimentation platform built on Chemspeed hardware ([69]) |
| Lila Sciences (Flagship Pioneering) | 2025, US | Autonomous labs spanning life, chemical, and materials sciences | $200 million in committed seed capital ([8]) | Positioned explicitly around building AI "superintelligence" for scientific discovery |
| Medra | 2025, US | "Physical AI Scientist" robots for biopharma R&D | $52 million Series A (Dec. 2025, led by Human Capital) ([9]) | Explicitly framed against the industry baseline that a new medicine still "takes 10-15 years and over $2B" |
Table 1 shows a market with two distinct commercial patterns: long-established robotics and cloud-lab vendors (Emerald Cloud Lab, Strateos, Opentrons, Chemspeed) that have accumulated a decade or more of operating history and disclosed pharma customers, and a newer wave (Lila Sciences, Medra, and smaller entrants such as ReactWise, which claims its models can achieve "cutting up to 95% of experimental work and accelerating these workflows by 30x" ([70])) explicitly marketing "self-driving lab" capability rather than lab automation as a category. Several additional names are worth noting even without a place in the table: Kebotix, a Cambridge, Massachusetts self-driving-lab pioneer for chemicals and materials discovery, raised $11.4 million in Series A funding in 2020 ([71]); Arctoris, founded at the University of Oxford in 2016, describes itself as the first fully automated, robotic laboratory dedicated to drug discovery and closed a £3.2 million seed round in 2019 ([72]); and Culture Biosciences, a cloud-bioprocessing company, closed a Series C round in December 2025 to expand its Stratyx 250 bioreactor and Console AI software, building on prior collaborations with Google Cloud ([73]).
“Self-driving laboratories, which incorporate robotics, advanced sensing, automated feedback control and artificial intelligence, are revolutionizing the conduct of biopharmaceutical research
Adoption in Pharma R&D: How Major Companies Are Deploying Closed-Loop Systems
Several large pharmaceutical companies and AI-native biotechs have disclosed named automation, AI-discovery, and laboratory-autonomy programs. The disclosures vary widely in measured outcomes and, in many cases, do not document the defining SDL feature: algorithmic selection of the next physical experiment.
AstraZeneca's iLab is the company's own in-house automation program, running since 2017 and now on a third-generation platform built with vendors BioSero and Zinsser Analytic, able to "synthesize and automatically purify several small-molecule compounds in parallel" with a stated goal to "identify potential drug candidates in half the time it takes today" ([13]). Separately, AstraZeneca signed an AI-led research collaboration with China's CSPC Pharmaceutical Group worth up to $5.3 billion, including a $110 million upfront payment, aimed at chronic-disease therapies ([74]).
Eli Lilly has pursued both physical lab automation and computational scale. Its 2020 Studio Lab, built with Strateos on a $90 million investment, packed synthesis, purification, analysis, and sample management into an 11,500 square foot San Diego facility ([75]). More recently, Lilly and NVIDIA announced a co-innovation AI lab in January 2026 with up to $1 billion in joint investment over five years, explicitly designed to connect "agentic wet labs" to computational "dry labs" for continuous, 24/7 experimentation ([15]), building on an October 2025 announcement that Lilly was building "the most powerful supercomputer owned and operated by a pharmaceutical company" using more than 1,000 NVIDIA B300 GPUs ([76]).
Merck (MSD) announced a multi-year agentic AI partnership with Google Cloud in April 2026 valued at up to $1 billion, deploying an agentic platform "across Merck's research & development, manufacturing, commercial and corporate functions" ([16]). Sanofi is investing CAD $294 million to expand its Toronto AI Centre of Excellence, adding 50 jobs on top of more than 150 roles created since the center was founded in 2022 ([77]). GSK is set to pay Relation Therapeutics up to $110 million to expand their existing AI-modeling drug discovery partnership ([78]), part of a broader industry push toward automated, AI-trained wet-lab discovery platforms. Genentech (part of the Roche group) frames its own approach as a "lab in a loop," explaining that "the 'lab in a loop' is a mechanism by which you bring generative AI to drug discovery and development," feeding wet-lab test results back to retrain the models ([79]), collaborating with AWS and NVIDIA on compute. Amgen says combining AI, automation, and biology in its generative-biology platform has "tripled protein engineering speed and cut discovery timelines in half" ([80]).
Among AI-native biotechs, Recursion Pharmaceuticals operates at what it describes as massive experimental scale, running "up to millions of wet lab experiments weekly, and massive computational scale" ([81]), and is publicly collaborating with HighRes Biosolutions and NVIDIA on what it calls "self-driving, high-throughput labs using advanced automation and orchestration" ([82]). Insilico Medicine launched LabClaw, which it calls the "pharmaceutical industry's first laboratory autonomy system built on a lightweight Agent-Guard architecture," with 5 AI agents and 28 skill modules spanning target discovery through wet-lab execution ([83]), and previously deployed a bipedal humanoid robot named "Supervisor" inside its fully robotic drug-discovery laboratory ([84]). And Isomorphic Labs, the Alphabet and DeepMind spinout, signed pharmaceutical partnerships with Eli Lilly (a $45 million upfront payment plus up to $1.7 billion in milestones) ([85]) and Novartis on the same day in January 2024, with combined deal value "worth nearly $3 billion to Isomorphic Labs, excluding any royalties" ([86]).
Not all disclosed results are recent. Sumitomo Dainippon Pharma (now Sumitomo Pharma) and Exscientia's AI-designed compound DSP-1181, targeting obsessive-compulsive disorder, entered Japanese Phase 1 trials after "requiring less than 12 months to complete the exploratory research phase, just a fraction of the typical average of 4.5 years" ([87]), one of the earliest publicly quantified pharma timeline reductions from AI-guided design, predating much of today's self-driving-lab terminology. On the regulatory side, the U.S. Food and Drug Administration's Center for Drug Evaluation and Research, together with the Center for Biologics Evaluation and Research and the Center for Devices and Radiological Health, opened formal discussion in May 2023 with a paper titled "Using Artificial Intelligence and Machine Learning in the Development of Drug and Biological Products" ([88]), signaling that regulators are now actively tracking this shift.
- 2017AstraZeneca iLab
In-house automation program aiming to identify drug candidates in half the time it currently takes.
- 2020Eli Lilly Studio Lab$90 million
Studio Lab built with Strateos packed synthesis, purification, analysis and sample management into an 11,500 sq ft San Diego facility.
- 2025Lila Sciences (Flagship Pioneering)$200 million
Launched with committed seed capital explicitly to build scientific 'superintelligence.'
- Dec 2025Medra$52 million
Series A raised to build 'Physical AI Scientist' robots, against the baseline that a new medicine takes 10-15 years and over $2B.
- Jan 2026Automata$45 million
Series C led by Dimension with a Danaher partnership, building the operating system for AI-ready labs.
- Jan 2026Eli Lilly & NVIDIA$1 billion
Co-innovation AI lab investment over five years connecting agentic wet labs with computational dry labs for round-the-clock experimentation.
- Apr 2026Merck (MSD) & Google Cloud$1 billion
Multi-year agentic AI partnership deploying an agentic platform across R&D, manufacturing, commercial and corporate functions.
Implementation Guidance: Evaluating and Adopting Self-Driving Lab Technology
Organizations considering an investment in closed-loop experimentation face a decision that is less about picking a single "winning" platform and more about sequencing a defensible, auditable pilot. Based on the vendor landscape and deployment patterns above, several practical evaluation criteria stand out:
- Match the autonomy level to the decision at stake. A liquid-handling robot executing a validated assay is not the same investment class as a closed-loop reaction-optimization system that changes its own protocol; conflating the two in a single "automation" budget line leads to mismatched expectations.
- Audit the AI method behind any "self-driving" claim. Ask whether the vendor's system uses Bayesian optimization, active learning, reinforcement learning, or an LLM agent, since each has different sample-efficiency, transparency, and validation characteristics, per the taxonomy above.
- Prioritize data standardization before scaling instrument count. As the Nature Reviews Chemistry survey notes, "interoperable data and metadata standards" remain an open requirement industry-wide ([51]); adopting protocols such as SiLA before or alongside hardware procurement avoids costly retrofits ([19]).
- Plan for GMP and validation tension early. A peer-reviewed regulatory analysis warns that AI's "continuous learning and opacity of decision-making processes" pose direct challenges to GMP principles ([20]); organizations should engage quality and regulatory affairs functions before, not after, a closed-loop system reaches a regulated process.
- Use FDA's Emerging Technology channels. The FDA established its Emerging Technology Program specifically "to assess innovative technologies, including AI/ML in pharmaceutical manufacturing" ([89]), and its Framework for Regulatory Advanced Manufacturing Evaluation (FRAME) initiative lists "ensure regulations and policy are compatible with future advanced manufacturing technologies" as a core priority ([90]), including open work on "standards for AI models used for process control and release testing" ([91]).
- Budget for pilot cost realistically. Public price points range widely across the stack, from an Opentrons OT-2 unit starting around $15,950 ([64]) to a full Emerald Cloud Lab subscription in the tens of thousands of dollars per month ([55]) to fully custom in-house builds costing tens or hundreds of millions of dollars, as Lilly's $90 million Studio Lab illustrates ([58]).
Independent technology assessment is a common step at this stage precisely because vendor claims and peer-reviewed measurements diverge (see the Data Analysis section below), and consultancies focused on regulated life-sciences technology, such as intuitionlabs.ai, describe their advisory work as "evaluation of current technology stack and recommendations for optimization" combined with "advisory services for maintaining compliance with industry regulations" ([21]) ([92]), a framing consistent with an adjacent-advisor role rather than a laboratory-automation vendor's.
Data Analysis and Evidence
Market-sizing estimates for laboratory automation vary meaningfully by scope and methodology, which is itself informative for buyers trying to benchmark vendor claims. Table 2 below compares three independent research firms' published figures for the broader lab automation market alongside the narrower AI-in-drug-discovery segment.
| Source | Market Segment | Base Year Value | Forecast Year Value | CAGR |
|---|---|---|---|---|
| MarketsandMarkets ([93]) | Global lab automation | $6.60B (2026) | $8.62B (2031) | 6.6% |
| Grand View Research ([5]) | Global lab automation | $8.27B (2024) | $18.39B (2033) | 9.3% |
| Fortune Business Insights ([94]) | Global lab automation | $10.07B (2026) | $20.71B (2034) | 9.43% |
| MarketsandMarkets ([6]) | AI in drug discovery | $5.09B (2026) | $17.56B (2031) | 28.1% |
| Grand View Research ([95]) | AI in drug discovery | $2.35B (2025) | $13.77B (2033) | 24.8% |
| Astute Analytica ([96]) | Lab automation and self-driving lab (combined) | $6.0B (2025) | $20B (2035) | 12.8% |
Table 2 shows that even reputable research firms disagree on base-year market size by roughly $3.5 billion for the same "lab automation" category, a spread that likely reflects differing definitions of what counts as automation versus adjacent categories like informatics or contract services. Notably, only Astute Analytica publishes a report using the "self-driving lab" label directly ([96]); larger, more established firms still fold the technology into the broader lab automation category, suggesting the term has not yet become standard market-research nomenclature even as the underlying technology matures. The AI-in-drug-discovery segment, by contrast, shows consistent high-double-digit growth rates across both firms that size it, reflecting stronger analyst conviction about near-term adoption of AI methods specifically, separate from the physical robotics layer.
This growth sits against a stable, sobering baseline on cost and risk. Grand View Research's own drug-discovery market report cites external research finding that "traditional drug discovery costs frequently exceed USD 2 billion per drug" with decade-plus timelines ([97]), consistent with the Tufts and Deloitte figures cited in the Introduction. McKinsey notes that clinical trial success rates "for the past two decades" have hovered "around 10 to 12 percent" ([98]), a figure roughly consistent with the Biotechnology Innovation Organization (BIO)'s largest-ever clinical success rate study, which found an overall Phase I likelihood of approval of 7.9% across 2011-2020 ([11]), though the same BIO dataset found some modalities perform far better, with CAR-T cell therapies showing "a Phase I LOA more than twice the 7.9% average across all diseases" ([99]). McKinsey also estimates typical "per-patient costs for pivotal clinical trials often exceed $40,000" ([100]), and separately projects that generative AI alone could generate "$60 billion to $110 billion a year in economic value for pharma and medical-product industries" ([101]).
A 2025 preprint benchmarking review of SDL performance is more cautious than vendor marketing. It reported that "acceleration factor" values across the studies reviewed have "a wide range... with a median of 6" ([17]), and that a related quality metric, the "enhancement factor," varies "by over two orders of magnitude" between studies ([102]). Because this manuscript is a preprint, its findings should not be characterized as peer-reviewed; the reported spread nonetheless warrants scrutiny of any headline multiplier without its underlying study design. Against that caution, one specific peer-reviewed head-to-head result from UCL School of Pharmacy stands out for its rigor: a self-driving pharmaceutical formulation robot was shown to "test 7 times as many formulations as a representative skilled formulator, whilst requiring only 25% of the human time" ([103]), reaching an optimal high-solubility formulation while "sampling only 256 out of 7776 potential formulations" ([104]), a concrete demonstration of the sample-efficiency argument for Bayesian optimization described earlier in this report.
Investment activity confirms the sector's growth even where market-sizing methodology is inconsistent. Beyond the vendor-specific rounds in Table 1, Crunchbase News reports that robotics startups broadly, spanning surgical, manufacturing-automation, and humanoid robotics categories, "pulled in just over $6 billion in 2025" through roughly July of that year ([105]), a category that increasingly overlaps with lab-automation robotics as AI-native entrants like Medra and Lila Sciences raise capital under both banners simultaneously.
“original claims of material novelty were subject to misinterpretation, their intention was to indicate that the materials were new to the prediction platform, not necessarily new to science
Case Studies and Real-World Examples
UC Berkeley's A-Lab: A Landmark Case Study, and a Cautionary One
The A-Lab, developed at Lawrence Berkeley National Laboratory and UC Berkeley, is described in its original 2023 Nature paper as "an autonomous laboratory for the solid-state synthesis of inorganic powders" ([106]). Over 17 days of continuous, unattended operation, it synthesized 36 of 57 targeted compounds ([107]), processing "50 to 100 times as many samples as a human every day" according to Berkeley Lab's own account ([46]). The case is important for pharma-adjacent SDL discussions not because it involved drugs, but because it is the most rigorously scrutinized autonomous-discovery result in the literature to date, and that scrutiny is instructive. In January 2026, Nature published a formal Author Correction after outside chemists challenged the original novelty claims; the authors acknowledged that their "original claims of material novelty were subject to misinterpretation, their intention was to indicate that the materials were new to the prediction platform, not necessarily new to science" ([18]), and after re-analysis confirmed the underlying prediction platform was correct in "36 of its 40 reported successes, with 4 compounds being inconclusive" ([108]). Chemistry & Engineering News reported that critics, including chemists at University College London and Princeton, remained unsatisfied even after the correction, with one telling the publication that "the advancement that it did for humanity is very incremental" despite the paper's outsized press coverage and citation count ([109]). The episode is a useful corrective for pharma buyers: headline throughput numbers from a self-driving lab and independently verified scientific novelty are not the same claim, and the two should be evaluated separately.
Carnegie Mellon's Coscientist: An LLM Agent Running Real Chemistry
Published in Nature in December 2023, Coscientist is an AI system driven by GPT-4 that autonomously designed, planned, and executed chemistry experiments, "including the successful reaction optimization of palladium-catalysed cross-couplings" ([44]), a reaction class central to pharmaceutical synthesis. Carnegie Mellon University described the achievement as the first time "a non-organic intelligent system has... designed, planned and executed a chemistry experiment" on its own ([110]). Notably, Coscientist ran one of its demonstrations remotely on Emerald Cloud Lab's infrastructure, with the paper reporting that "generated code was successfully executed at ECL... the sample was a caffeine standard sample" analyzed by liquid chromatography-mass spectrometry ([111]), a concrete demonstration of an LLM agent controlling a cloud lab it had never physically touched. Following the paper, Carnegie Mellon partnered with Emerald Cloud Lab to open what the university describes as the first university-based cloud lab, giving researchers "access to more than 200 pieces of equipment" ([112]) starting in early 2024.
University of Amsterdam's RoboChem: Compressing Months of Reaction Optimization Into a Week
Published in Science in January 2024, RoboChem is a closed-loop, Bayesian-optimization-driven platform for photocatalytic flow chemistry, directly relevant to pharmaceutical synthesis routes. Lead researcher Timothy Noël described the practical impact concretely: "in a week, we can optimise the synthesis of about ten to twenty molecules. This would take a PhD student several months" ([113]). A 2026 follow-up platform, RoboChem-Flex, published in Nature Synthesis, is a low-cost (roughly $5,000), modular self-driving lab validated across six diverse chemistry case studies, "including photocatalysis, biocatalysis, thermal cross-couplings and enantioselective catalysis" ([53]), explicitly aimed at making self-driving lab capability accessible to labs without a well-funded infrastructure budget.
Insilico Medicine: A Pharma-Native Robotic Discovery Pipeline
Insilico Medicine has built one of the most extensively documented pharma-native closed-loop programs. Its sixth-generation "Life Star" facility, launched in Suzhou BioBAY in December 2022, was built to "form a closed loop" between fully automated robotics and the company's PandaOmics AI discovery platform ([114]). In 2025, the company deployed a bipedal humanoid robot named "Supervisor" inside that facility to help train embodied AI on laboratory tasks ([84]), and by 2026 had launched LabClaw, its "Agent-Guard" architecture combining 5 AI agents and 28 skill modules to run the full loop from target discovery through wet-lab execution ([83]). The company reports that its wholly-owned pipeline, built on this platform since 2021, has grown to 30 assets, of which 10 have received IND clearance ([14]), among the most concrete disclosed pipeline outcomes tied to a self-driving lab investment anywhere in the industry.
Big Pharma In-House Programs: Recursion's REC-1245 and the Broader Trend
Recursion Pharmaceuticals' REC-1245 program illustrates what a large-scale, AI-guided closed loop can look like inside an established biotech: the company reports the program "moved from target ID to IND enabling studies in under 18 months with approximately 200 compounds synthesized" ([12]), supported by the company's FDA clearance to proceed to human trials ([115]). Recursion has also partnered with Enamine to curate "10 enriched screening libraries from over 15,000 newly synthesized compounds" across roughly 100 clinically relevant drug targets ([116]). Set alongside AstraZeneca's iLab and Eli Lilly's Studio Lab and NVIDIA collaboration described earlier, these cases together show a consistent pattern across large, well-capitalized organizations: pharma companies are not waiting for a single, standardized self-driving lab product to mature, but are instead building bespoke closed loops around their own existing chemistry and biology workflows, each disclosing a different slice of quantitative outcome (compound counts, timeline compression, or capital committed) rather than a common benchmark.
Implications and Future Directions
Several trends will likely shape how self-driving lab technology develops in pharma over the next several years. First, capital is flowing toward pure-play autonomy startups rather than only established automation vendors: Flagship Pioneering's Lila Sciences launched with $200 million in committed seed capital explicitly to build what it calls scientific "superintelligence" ([8]), and Medra raised $52 million to build what it calls "Physical AI Scientists," explicitly against the baseline that a new medicine still "takes 10-15 years and over $2B" to reach market ([9]). Second, large language model agents are moving from research demonstrations, such as Coscientist, toward production deployments, illustrated by Insilico's Agent-Guard architecture and by Bristol Myers Squibb research leadership's description of moving toward "an integrated learning ecosystem, where AI helps every experiment, clinical readout, and partnership compound into higher-conviction scientific decisions" ([117]). Third, cost democratization is a live research theme in its own right, exemplified by RoboChem-Flex's roughly $5,000 bill of materials, engineered specifically because "widespread adoption remains limited by high costs, complex infrastructure and limited accessibility" ([52]).
Regulatory clarity remains an open question that will materially affect adoption speed. The FDA's Emerging Technology Program and its FRAME initiative are actively working through "standards for AI models used for process control and release testing" ([91]), but no FDA guidance document as of August 2026 explicitly names "self-driving laboratories" or "autonomous experimentation platforms," addressing AI and machine learning in drug development and manufacturing only in broader terms. Industry standards bodies such as ISPE are separately pushing interoperability protocols like SiLA ([19]), which will need to mature before closed-loop systems can be validated and audited to the standard GMP manufacturing already expects. Finally, the academic research community itself is becoming more self-critical, as the A-Lab correction demonstrates; expect more rigorous, adversarial peer review of headline autonomous-discovery claims going forward, which should ultimately benefit pharma buyers by separating durable capability from one-off demonstrations.
For life-sciences organizations without in-house AI and automation expertise, this is precisely the environment in which an experienced, technology-agnostic advisor adds the most value: helping translate vendor and academic claims into a realistic pilot scope, a defensible validation plan, and a regulatory engagement strategy, the kind of "digital strategy development" and "process optimization" work intuitionlabs.ai describes as central to its advisory practice ([118]) ([119]). IntuitionLabs itself cites Deloitte research suggesting "AI-enhanced drug discovery and development can accelerate timelines by up to 60%" and McKinsey estimates that AI "could generate $100B+ in annual value for the pharmaceutical industry" ([120]) ([121]), figures broadly consistent with the McKinsey research cited independently earlier in this report, and a useful reminder that the underlying economic case for closed-loop R&D investment is corroborated across multiple independent analysts, not asserted by any single vendor alone.
Frequently Asked Questions (FAQs)
What is a self-driving laboratory? A self-driving laboratory (SDL) is a robotic experimental platform in which a machine-learning algorithm, rather than a human, selects each successive experiment based on the results of prior experiments, forming a closed loop. Nature Synthesis defines it as "a machine-learning-assisted modular experimental platform that iteratively operates a series of experiments selected by the machine learning algorithm to achieve a user-defined objective" ([1]).
How do self-driving labs differ from traditional lab automation? Traditional automation, such as a liquid-handling robot, executes a fixed, human-designed protocol without changing its plan based on results. A self-driving lab closes that loop, using an algorithm to interpret each result and choose the next experiment autonomously, as described in the taxonomy section above ([37]).
Are self-driving labs actually used in pharma today, or is this mostly a research concept? Both, but the evidence should be classified carefully. Academic demonstrations such as A-Lab and Coscientist document closed-loop experimentation. In pharma, AstraZeneca's iLab and Lilly and Strateos's Studio Lab are disclosed automation or cloud-lab programs, Recursion's REC-1245 is an AI-enabled drug-discovery program, and Insilico describes LabClaw as a laboratory-autonomy system. These disclosures do not provide equivalent public evidence that each system autonomously selects every next physical experiment.
What AI methods power self-driving labs? Commonly used approaches include Bayesian optimization and active learning, which are used for sample-efficient experiment selection; Bayesian optimization can identify near-optimal solutions with fewer experiments than random search ([4]). Reinforcement learning can support multi-step protocol optimization, while large language model agents are an emerging category ([41]).
What are the main obstacles to wider pharma adoption? Interoperable data standards, high infrastructure cost and complexity ([52]), and a structural tension between continuously learning AI models and the fixed, validated procedures GMP regulation expects ([20]).
How much faster are self-driving labs than conventional experimentation? It depends heavily on the study. A 2025 preprint benchmarking review reported a median "acceleration factor" of 6 across the SDL studies it reviewed, with results varying by orders of magnitude ([17]). As a preprint, it is not peer-reviewed, so buyers should treat any single vendor-quoted multiplier with caution and examine the underlying study design.
Conclusion
Self-driving laboratories represent a genuine, well-defined technical category, closed-loop systems in which an AI algorithm selects each next experiment, distinct from decades-old automated liquid handling and screening. The concept has moved from academic proof of concept toward pharma-relevant automation and AI-enabled discovery. Public disclosures describe AstraZeneca's iLab and Lilly's Studio Lab as automation or cloud-lab systems, Recursion's REC-1245 as an AI-enabled discovery program, and Insilico's LabClaw as a company-described laboratory-autonomy system; those examples should not be treated as equivalent evidence of verified autonomous next-experiment selection. The commercial vendor landscape spans mature cloud-lab operators such as Emerald Cloud Lab and Strateos, established robotics and orchestration companies such as Opentrons, Chemspeed, Automata, and Synthace, and a new wave of autonomy-native entrants such as Lila Sciences and Medra, collectively drawing on a broader pool of more than $50 billion invested in AI-driven R&D companies since 2015.
The evidence base, however, is genuinely mixed, and organizations evaluating this space should hold vendor claims and peer-reviewed measurement to different standards. Market-sizing estimates vary by billions of dollars depending on methodology, and the most scrutinized academic case study to date, the A-Lab, required a formal correction after independent chemists challenged its original claims. Meanwhile, rigorous head-to-head studies, such as the UCL formulation-robot result showing a sevenfold throughput gain at a quarter of the human time cost, demonstrate that the underlying technology does deliver measurable gains when properly validated. For manufacturing, CMC, process-control, and release-testing uses, FDA's Emerging Technology Program is relevant only to quality and facility-related information in applications; AI used in drug discovery and development should be discussed through the FDA engagement route applicable to its intended use ([122]) ([123]). Interoperability standards such as SiLA and robust validation practices will also influence whether self-driving labs move beyond disclosed pilots to broader pharmaceutical use.
Sources / 123
Need Expert Guidance on This Topic?
Let's discuss how IntuitionLabs can help you navigate the challenges covered in this article.
I'm Adrien Laurent, Founder & CEO of IntuitionLabs. With 25+ years of experience in enterprise software development, I specialize in creating custom AI solutions for the pharmaceutical and life science industries.
The information contained in this document is provided for educational and informational purposes only. We make no representations or warranties of any kind, express or implied, about the completeness, accuracy, reliability, suitability, or availability of the information contained herein. Any reliance you place on such information is strictly at your own risk. In no event will IntuitionLabs.ai or its representatives be liable for any loss or damage including without limitation, indirect or consequential loss or damage, or any loss or damage whatsoever arising from the use of information presented in this document. This document may contain content generated with the assistance of artificial intelligence technologies. AI-generated content may contain errors, omissions, or inaccuracies. Readers are advised to independently verify any critical information before acting upon it. All product names, logos, brands, trademarks, and registered trademarks mentioned in this document are the property of their respective owners. All company, product, and service names used in this document are for identification purposes only. Use of these names, logos, trademarks, and brands does not imply endorsement by the respective trademark holders. IntuitionLabs.ai is an AI software development company specializing in helping life-science companies implement and leverage artificial intelligence solutions. Founded in 2023 by Adrien Laurent and based in San Jose, California. This document does not constitute professional or legal advice. For specific guidance related to your business needs, please consult with appropriate qualified professionals.
Related Articles
How Many FDA Submissions Use AI? CDER's Review Findings
Analysis of CDER's disclosed AI submission counts from 2016 to 2026, the January 2025 AI credibility guidance, the Elsa review tool, and named case studies as of August 2026.
AI-Discovered Drugs in Clinical Trials 2026: Full Pipeline
A data-driven 2026 survey of AI-discovered drugs in clinical trials, mapping Insilico Medicine, Recursion Pharmaceuticals, Isomorphic Labs, success rates, funding, and named case studies.
Claude Science vs GPT-Rosalind vs Isomorphic Labs Compared
A July 2026 comparison of Claude Science, GPT-Rosalind, and Isomorphic Labs covering pricing, access models, published benchmarks, named enterprise adoption, and which platform fits which life-sciences use case.