AI labs go autonomous: pharma, cloud, and science unite
The gist
AI-powered autonomous labs are breaking out of the lab and into the fast lane, fusing robotics, cloud, and cutting-edge AI to turbocharge scientific discovery and drug development.
What to know
- Medra is building a $52M fully autonomous AI-driven lab aiming to be the 'TSMC for drug discovery,' blending AI reasoning with robotics for nonstop experimentation.
- The American Science Cloud launched in 2025, uniting 17 DOE national labs and 28 user facilities into an AI-ready platform that doubles U.S. research productivity.
- Next-gen AI agents like SAGA and Step-DeepResearch are now autonomously designing experiments and accelerating breakthroughs in antibiotics, materials, and more.
AI Labs: From Automation to Reasoning
A new era of scientific discovery is emerging as AI-powered labs blend advanced reasoning with robotics, enabling autonomous experiment planning and execution across fields from pharma to materials science.
The foundation of autonomous AI-driven scientific discovery labs emerged from the critical integration of advanced AI reasoning with robotic lab automation, enabling collaborative experiment planning and execution across diverse scientific domains. As Oliver Hsu highlighted in late 2025, this fusion of AI's interpretability and physical automation marked a transformative shift, particularly in mature markets such as life sciences, pharmaceuticals, chemistry, and materials science. Early pioneers like Periodic Labs, Medra, Chemify, and Yoneda Labs, alongside government-industry collaborations such as the Department of Energy's Genesis mission and DeepMind's UK partnership, exemplified this nascent ecosystem where AI not only automated tasks but began to reason about experiments, setting the stage for more autonomous scientific workflows.
Medra stands out as a trailblazer in the pharmaceutical sector, pioneering the deployment of physical AI scientists capable of autonomous experimentation by integrating AI reasoning with robotics. Founder Michelle emphasized that their systems go beyond mere automation, embodying intelligent robotics that continuously sense, correct, and reason scientifically. By early 2026, Medra was poised to launch one of the largest fully autonomous labs in the United States, backed by a $52 million Series A funding round led by Human Capital and partnerships with major pharma players like Genentech. This ambitious infrastructure aims to serve as the 'TSMC for drug discovery,' signaling a strategic move to embed autonomous AI labs deeply within the drug development pipeline.
National laboratories have been instrumental in advancing autonomous AI labs beyond proof-of-concept stages, with Argonne National Laboratory's RAPID platforms and Lawrence Livermore National Laboratory's ARMOR and APEX projects demonstrating cross-disciplinary applications in chemistry, materials science, and bioactive research. These labs integrate AI-driven closed-loop experimentation where robots conduct experiments, AI analyzes data, and systems autonomously decide subsequent steps, accelerating discovery by orders of magnitude. As Ian Foster of Argonne analogized in mid-2026, these developments are akin to the dawn of computing decades ago, underscoring the transformative yet early-stage nature of autonomous scientific discovery.
Complementing these domain-specific advances, companies like Intology have championed domain-agnostic autonomous AI systems that automate the iterative experiment-feedback cycle across fields from drug discovery to AI model development. Despite challenges such as costly, noisy, and slow experiments—exemplified by the complexities of post-training large language models—Intology has transitioned from heavy venture capital burn to establishing efficient infrastructure that optimizes compute resources. This broader vision aligns with the biopharmaceutical industry's push toward self-driving laboratories that integrate robotics, advanced sensing, and machine learning to enable continuous, data-driven discovery, although hurdles like sensor compatibility and data integration remain active areas of innovation.
AI Agents Rewrite Scientific Method
Reinforcement learning and generative AI agents now autonomously generate hypotheses and design experiments, rapidly accelerating breakthroughs and shifting the research paradigm toward closed-loop, self-improving science.
The year(https://podcasts.apple.com/podcast/id1528594034) marked a pivotal breakthrough for autonomous AI agents in scientific discovery, with generative design techniques—particularly de novo antibody design—revolutionizing drug development by drastically reducing timelines. By 2026-2027, experts predict that AI agents will generate the majority of high-quality scientific hypotheses, signaling a transformative shift in research paradigms where AI-driven closed-loop experimentation becomes the norm rather than the exception.
Innovative AI architectures such as SAGA and Step-DeepResearch have advanced closed-loop adaptive experimentation by enabling dynamic reformulation of research objectives and decomposing complex scientific tasks into planning, verification, and reporting stages. SAGA’s bi-level loop avoids reward hacking by iteratively refining objectives, while Step-DeepResearch’s 32B parameter agent achieves superior performance cost-effectively, illustrating how reinforcement learning and large language models synergize to create self-improving experimental workflows across domains like antibiotic design and materials science.
Integrating autonomous laboratories with AI models that have direct access to experimental data is proving crucial for overcoming challenges such as sparse rewards, experimental cost, and data reliability. Companies like Periodic Labs and DeepMind are constructing large-scale physical labs to generate experimentally verified reward signals, enabling reinforcement learning systems where subagents specialize in tasks and general LLMs orchestrate workflows. This closed-loop approach, which mimics a graduate student’s iterative hypothesis testing through simulations before physical experiments, addresses the limitations of training solely on textual data and published literature.
Multi-agent AI systems and generative AI platforms such as AutoLabs are pushing the boundaries of closed-loop experimentation by managing complex, large-scale autonomous research campaigns and translating natural language goals into multi-step robotic workflows. These systems tackle challenges like interpreting incomplete instructions and capturing implicit engineering knowledge, enabling researchers to perform up to ten times more experiments with minimal human intervention. Meanwhile, autonomous labs like Argonne’s RAPID demonstrate the power of parallel experimentation, conducting thousands of experiments simultaneously and accelerating discovery in fields from energy storage to bio-active factors.
American Science Cloud Unifies Research
The American Science Cloud is transforming U.S. research by federating national labs and user facilities into an interoperable, AI-ready platform that doubles productivity and standardizes access to cutting-edge tools.
Central to the national push for AI-driven scientific discovery is the American Science Cloud (AmSC), launched in October 2025 as the foundational infrastructure for the White House's Genesis Mission. As Chin Guok, CTO at ESnet, and Berkeley Lab’s Monga emphasize, AmSC unifies fragmented resources across 17 DOE national labs and 28 user facilities into a cohesive, AI-ready platform by leveraging ESnet’s high-performance network and Infrastructure Resource Orchestration middleware. This federated cloud infrastructure not only enables seamless interoperability without dictating scientific methods but also extends access to commercial cloud services and frontier AI models, effectively doubling U.S. research productivity through coordinated, AI-driven workflows.
The Department of Energy’s commitment of over $600 million to the Genesis Mission underscores a strategic investment in interdisciplinary teams tackling grand challenges such as fusion energy, quantum science, and autonomous experimentation. Complementing this, the DOE’s Atmospheric Radiation Measurement (ARM) User Facility is modernizing its infrastructure with AI-ready storage, GPU integration, and innovative AI agents like the ARM Data Advisor (ADA), which simplifies data discovery through natural-language interaction. These upgrades, alongside the Agentic Tooling and LLM Augmentation Stack (ATLAS) framework, exemplify efforts to standardize AI tool integration and promote responsible, trustworthy AI use, thereby enhancing the national research ecosystem’s efficiency and user experience.
Beyond DOE initiatives, the National Science Foundation’s 2026 investments totaling $83 million, including the $24.5 million FabAID project led by the Morgridge Institute for Research, signal a coordinated national effort to build AI-ready data infrastructure that integrates computing, data systems, and AI resources. These efforts align with the Genesis Mission’s goal to strengthen U.S. leadership in AI-driven science by fostering large-scale cloud platforms and interoperable ecosystems that accelerate innovation across disciplines.
The $20 million NSF grant awarded to the University of Chicago and Argonne National Laboratory for the PoLARIS project marks a significant leap in scaling AI-powered scientific infrastructure through a cloud laboratory that integrates robotics, AI, and data platforms. Designed to democratize access to advanced polymer science and bolster U.S. competitiveness in a $200 billion industry, PoLARIS will enable autonomous experiments running 100 to 300 times daily across multiple labs, with a full user-service launch planned within two years. This initiative exemplifies the transformative potential of coordinated, AI-driven national research ecosystems to accelerate discovery and industrial innovation at unprecedented scales.
Democratizing Discovery: Lab-as-a-Service
Startups like Medra and Lila Sciences are making fully autonomous labs accessible through cloud platforms and modular AI scientists, enabling even zero-employee biotech startups to run full R&D pipelines remotely.
By early 2026, Medra emerged as a pioneer in democratizing autonomous scientific discovery by offering physical AI scientists that can integrate into existing labs or operate within their own fully autonomous facility, slated to become one of the largest in the U.S. Their vision to become the 'TSMC for drug discovery' is backed by a robust $52 million Series A funding round led by Human Capital and supported by repeat investors like Lux and Melo Ventures, signaling strong commercial confidence in AI-driven lab infrastructure.
Lila Sciences revolutionized access to autonomous research through a cloud-based AI-driven lab platform that compresses years of biotech development into mere months with minimal staffing. Their model enables 'zero FTE startups' to conduct entire R&D pipelines remotely, charging platform fees and sharing milestone gains, exemplified by their rapid assembly of a CAR-T therapy that matched a $2.1 billion biotech acquisition. This approach is underpinned by a closed-loop system integrating AI with robotic instruments, generating experimentally verified reasoning traces that vastly exceed traditional data corpora, and supported by plans for a 'lights-out' data center to maximize throughput.
Addressing the complexity of instrument integration, Lila employs innovative solutions like vision-language models to operate legacy lab equipment lacking modern APIs, slashing onboarding times from 30 days to 30 minutes. This technical ingenuity complements broader efforts by Core Automation and Intology to maximize human agency and automate domain-agnostic scientific discovery cycles, respectively, highlighting a trend toward sustainable, scalable autonomous labs that blend human expertise with AI efficiency.
The democratization of autonomous science is further exemplified by initiatives like the NSF-funded PoLARIS project, which provides cloud-based AI-powered robotic labs accessible remotely to external researchers, running hundreds of autonomous experiments daily. This infrastructure not only accelerates reproducible research with minimal human intervention but also envisions broad accessibility—from universities integrating these labs to high school students remotely conducting experiments—thereby dismantling traditional institutional barriers and fostering inclusive innovation in critical sectors such as clean energy and healthcare.
Human Oversight Anchors Autonomy
Multi-agent AI systems and digital sandboxes are ensuring trustworthy, reproducible science by embedding human-in-the-loop checks that keep autonomous labs interpretable and verifiable.
By mid-2026, the integration of multi-agent AI systems with digital and physical sandboxes emerged as a cornerstone for trustworthy autonomous lab management, enabling dynamic task allocation and rigorous strategy validation. This approach, highlighted in the July 9 analysis on autonomous materials labs, underscores the necessity of human-in-the-loop oversight to optimize planning, decision-making, and reproducibility, ensuring that AI-driven research campaigns remain interpretable and verifiable.
Oak Ridge National Laboratory’s support for the DOE ARM User Facility’s AI overhaul exemplifies the push toward responsible AI governance and secure tool integration through frameworks like the Agentic Tooling and LLM Augmentation Stack (ATLAS). This infrastructure not only standardizes AI workflows but also facilitates scalable automation while embedding human validation checkpoints, aligning with DOE’s Genesis Mission to foster reliable and trustworthy scientific discovery environments.
Lawrence Livermore National Laboratory’s advancements in autonomous labs emphasize the critical distinction between automation and autonomy, where AI systems do more than execute pre-programmed tasks—they actively analyze data and determine experimental directions. As Aldair Gongora and Chris Spadaccini note, this intelligent guidance frees scientists to focus on experiment design and interpretation, reinforcing human oversight as essential for maintaining trust, reproducibility, and interpretability in complex scientific workflows.
Pacific Northwest National Laboratory’s ARCADIA platform represents a paradigm shift toward shared, AI-driven discovery ecosystems that integrate robotics, AI agents, and curated data to scale autonomous science responsibly. By embedding validation checkpoints where human scientists confirm AI actions before proceeding, ARCADIA ensures reproducibility and reliability as autonomy expands. Bob Runkle’s emphasis on autonomy complementing rather than replacing human judgment—‘We need people to imagine the science we should be doing’—captures the ethos that human insight remains central amid accelerating AI-driven experimentation.
HPC Meets Autonomous Science
Embedding AI inference directly into high-performance computing enables near real-time, closed-loop experimentation, multiplying research speed while preserving critical human decision-making at scale.
By mid-2026, Argonne National Laboratory's vision for fully autonomous scientific ecosystems centers on embedding AI inference services directly into high-performance computing (HPC) infrastructures, enabling AI agents to orchestrate complex workflows across multiple computing systems. This integration facilitates near real-time scientific discovery by automating data transfer, analysis, and feedback loops during experiments, as demonstrated in advanced applications like tomography and fusion research where AI processes results within minutes, dramatically accelerating research cycles.
The evolution toward fully autonomous labs is marked by AI agents that not only analyze data but also independently plan and execute experiments, particularly in drug discovery where AI-designed targets have outperformed human counterparts. This closed-loop experimentation exemplifies a paradigm shift where AI continuously runs simulations and adapts strategies until desired outcomes are achieved, signaling a transformative impact on the scientific method and innovation landscape.
Contrary to fears of human obsolescence, leading voices like those from Core Automation emphasize that the future of AI-driven scientific ecosystems aims to maximize human agency rather than eliminate human involvement. By automating routine tasks, researchers can conduct experiments at unprecedented speeds and scales—moving from one experiment per day to potentially hundreds—thereby fundamentally transforming research workflows while preserving critical human oversight and decision-making.
Despite ambitious progress, achieving artificial general intelligence (AGI) that can self-improve without human intervention remains elusive, as current AI systems still rely heavily on human-in-the-loop models. Ongoing research through 2024 and early 2025, particularly within small, highly efficient teams focused on agent-based automation, continues to push the boundaries of autonomy, highlighting both the technical challenges and the incremental nature of advancing toward fully autonomous scientific ecosystems.









