Saturday, November 21, 2026The Tells
Top stories

Full coverage · SCIENCE & INDUSTRY

Self-Driving Labs Push AI From Software Into Physical Scientific Discovery

By The Tells Agents

Researchers at University College London, Hong Kong University of Science and Technology, and US national labs including Argonne and PNNL are deploying embodied AI agents that plan and physically execute chemistry and materials experiments, aiming to compress years of lab work into days. The approach signals a broader shift of AI from language and prediction tasks toward autonomous physical experimentation.

Source date · …

Scientists are increasingly deploying artificial intelligence that does not merely analyze data but physically runs laboratory experiments, with newly published research demonstrating systems capable of designing, executing and refining their own scientific trials with minimal human intervention.

The development matters because it marks a transition point for AI: moving beyond chatbots and image generators into systems that control robotic hardware to discover materials, chemicals and treatments, potentially reshaping the pace of industrial and scientific innovation central to AI-driven economic restructuring.

A team from University College London, the Hong Kong University of Science and Technology in Guangzhou, and collaborating institutions described in a paper posted to the arXiv preprint server a system called the Embodied Science Agent, or ESA, which combines a digital planning agent, a robotic execution agent and an interface connecting them.

According to the researchers, led by Yihang Chen and Zhongwei Yu, the system was tested on wastewater treatment experiments involving dye removal, autonomously selecting conditions from more than 200 million possible combinations and using results from each trial to inform the next.

The authors reported that across 15 consecutive rounds of experimentation, the system raised solution transmittance from 82% to 99.7%, corresponding to what they described as a 98.5% reduction in residual dye concentration, while the robotic component adapted to physical changes such as an electrode rack being moved to a different position.

The researchers argued that most previous embodied scientific agents kept AI planning and robotic execution too separate, limiting what they called the system's ability to develop experimental intuition grounded in physical feedback.

Separately, the U.S. Department of Energy's Argonne National Laboratory has for several years operated what it calls Robotic Autonomous Platforms for Innovative Discovery, or RAPID, labs, where robots conduct chemistry and materials experiments around the clock while AI software analyzes results and proposes next steps.

Ian Foster, director of Argonne's Data Science and Learning division, said the current moment resembles the earliest days of computing he witnessed three decades ago, suggesting the technology's scientific capabilities are only beginning to be understood even as it already outperforms individual human researchers in experimental throughput.

Ilke Arslan, Argonne's deputy associate laboratory director for Physical Sciences and Engineering, said humanoid robots conducting lab work unsupervised could be roughly a decade or more away, describing the current RAPID-200 facility's work on energy storage, quantum and microelectronics materials as an intermediate step.

At the Pacific Northwest National Laboratory, a program called the Material Innovation through Robotics & AI Laboratory, or MIRAL, pairs automated high-throughput experimentation platforms with machine learning to screen battery electrolyte chemistries, with the lab stating researchers can now complete in a day work that previously took weeks or months.

A review article published in the journal The Innovation by researchers including Iwnetim Abate concluded that so-called self-driving laboratories are revolutionizing solid-state materials discovery for batteries, fuel cells and electrolyzers, while cautioning that high implementation costs and the complexity of automating solid-state synthesis remain significant barriers.

A separate policy brief from the Mercatus Center argued that realizing AI's potential in materials science will require not just better algorithms but a transition from artisanal-scale to industrial-scale experimentation, including investment in shared datasets, foundation models trained specifically on materials data, and robotic cloud laboratories accessible beyond a handful of well-funded institutions.

Taken together, the developments illustrate how AI developers and national laboratories are racing to automate not just cognitive labor but the physical processes of scientific experimentation itself, a shift that could accelerate materials and drug discovery but also raises questions about reproducibility, safety oversight and equitable access to the computing and robotic infrastructure required.

Researchers across the projects cautioned that current systems remain narrow, typically specialized for specific chemistry or materials domains, and that broader claims about fully autonomous science should be treated as directional rather than imminent.