Inside PNNL’s Lab-Wide Push for Trustworthy Autonomous Science

RICHLAND, Wash. — A quiet revolution is taking place in the halls and laboratories at the Department of Energy’s Pacific Northwest National Laboratory. Labs that once contained handheld equipment are being rapidly replaced with robotic handlers and data collection interfaces that integrate artificial intelligence and next-generation data collection systems.

At the heart of it all, PNNL’s Center for Robotics and Autonomy is laser-focused on integrating automation and assured AI into the Laboratory’s scientific experiments. 

CRA’s director, Bob Runkle, is forthright in stating the Center’s ambitious goal: “At PNNL, autonomy will touch every experiment in five years, and the Center for Robotics and Autonomy will make that happen.”

“The science of tomorrow won’t look like the science of today,” he said.

Why so fast?

“We are in a race to maintain a competitive advantage in scientific discovery in the service of our national interests,” said Runkle. “We can’t afford not to move as fast as we can to put the best tools in the hands of our scientists.” 

After all, he added, “the goal is for our researchers to be able to explore questions and envision solutions that were previously out of reach.”

Autonomous systems simply remove bottlenecks, allowing researchers with expertise in scientific disciplines to focus on what matters: identifying the highest priority research questions, understanding how to frame them, and recognizing when a research outcome makes meaningful progress. The goal is for autonomous systems to make the pace of discovery limited only by the rate of the physical or biological process being studied. 

“We aim to further scientific knowledge at the speed of nature,” he added, “For example, imagine if we could understand novel materials in real time as they form for the first time?”

How do we trust robots?

While speed is essential to discover solutions in science, energy and national security, the CRA is being built on a foundation of validation and verification — two principles that have always underwritten the work done at PNNL since its early days in nuclear treaty verification. The CRA is assuring that as machines take on more of the experimental workload, scientists can focus even more on confirming that the scientific output remains trustworthy and reliable. 

“Everything in science must be reproducible,” Runkle said. 

This means building in checkpoints where human scientists confirm that an AI agent’s actions or conclusions are sound before work proceeds.

ARCADIA: a discovery platform built with human needs in mind

At the center of PNNL’s approach is ARCADIA — Agentic Robotics and Curated AI Data for Intelligent Autonomy — a software platform that integrates AI agents, robotics and data systems into a single discovery ecosystem. Rather than building isolated autonomous labs for individual projects, PNNL is investing in ARCADIA as shared infrastructure meant to support autonomous science across the Lab.

Runkle is careful to describe the role of autonomy as complementary to, rather than a replacement for, human scientific judgment. 

“We need people to imagine the science we should be doing and to define the interesting questions,” he said. “That's a judgment call. That’s subjective and it should stay that way.” 

PNNL’s Malachi Schram, who is leading development of the platform, described the intent behind ARCADIA. 

“There is a lot of AI agentic work that’s going on across the Lab, and we would like to learn what everybody is doing so everyone benefits from this revolution,” he said. Schram emphasized that the goal is not to duplicate effort across the Laboratory but to create a collaborative and common platform that integrates prior and ongoing work to provide researchers the ability to go from ideas to physical execution.

ARCADIA takes advantage of a multi-agent tool called SciLink that helps automate the scientific process by assisting with everything from hypothesis generation and experimental design to simulation and analysis. But the real power comes from integration with an AI-ready data store that keeps information flowing seamlessly between instruments, researchers and autonomous systems. 

Why are autonomous labs realistic in 2026 and beyond?

People have been using robots for decades. That’s not new. But as recently as five years ago, programming a robot to move lab equipment required painstaking computer programming for every movement, every task, every time. It would take a software engineer months or even years to get a whole system in place. Now, large language models that understand written and spoken direction can program precise movements in minutes. That’s the difference. 

“I’m not a coder,” said Runkle. “But I can go into ARCADIA and, just using natural language, control the experiment. That’s the breakthrough. That’s what makes LLMs so powerful. You and I can interact with them in a natural way.”

At PNNL’s Autonomy Studio, researchers are designing digital twins, virtual labs that allow them to visualize and interact with robotic lab systems through an interface, rather than only through the physical hardware itself. This allows researchers to carefully design and calibrate experiments before deploying them in a real physical laboratory. 

Investment, industry partners and infrastructure

PNNL is backing CRA with sustained internal investment. The laboratory is committing more than $5 million annually through its Foundational Autonomy Investment to build the platforms, infrastructure and expertise required to scale autonomous science across the Laboratory. But no single entity could possibly do it all in isolation. 

That’s why PNNL is collaborating with industry partners to help strengthen its technical foundation. Through a collaboration with Amazon Web Services, PNNL is deploying a data lakehouse containing AI-ready data from instruments across the Laboratory, accessible to AI agents and autonomous systems. A partnership with Microsoft, using Microsoft Discovery, is supporting the development of machine learning models that predict and optimize chemical processes, including identifying materials for energy storage and the design of microbes with specific functions. PNNL is also working with Edison Scientific to embed specialized agents directly into research workflows.

Similarly, the Environmental Molecular Sciences Laboratory, a DOE user facility located on the PNNL-Richland campus, is applying similar design-and-build principles by partnering with Ginkgo Bioworks to build autonomous experimental platforms for microbial research. The two organizations have already built the Anaerobic Microbial Phenotyping Platform, which automates microbiology experiments. Together, PNNL and Ginkgo Bioworks are designing a custom Microbial Molecular Phenotyping Capability that will benefit a broad community of biology researchers with automation and AI-ready datasets.

Through ARCADIA, its partner network, and its emphasis on verification and reliability, the CRA represents an ambitious approach to autonomy, but one that is tied to the trustworthiness of the systems producing it.

“In five years, you’ll see autonomy deployed across PNNL,” said Runkle. “It may be a very small touch or a firm grasp, but either way we spread the efficiency across the Laboratory.”