Agentic AI Bot Helps Scientists Speak to Robots, Speed up Experiments

RICHLAND, Wash.—Researchers at the Department of Energy’s Pacific Northwest National Laboratory use a slew of autonomous robots for designing and implementing experiments. However, setting up an experiment on an autonomous lab robot is surprisingly slow. The effort requires a lengthy back-and-forth between a scientist and an engineer to design the experimental steps—a process that can take weeks.

To help researchers work more efficiently, a PNNL team developed a generative agentic AI that can quickly translate experimental goals into instructions for a laboratory robot. The translation agent, called AutoLabs, is currently designed to operate with Big Kahuna, an automated robot built by Unchained Labs that researchers use to study new and existing battery materials. The system can carry out multi-step experimental workflows, including mixing, heating, stirring and filtering samples with minimal human intervention. By automating these processes, researchers can perform 5 to 10 times more experiments than would be practical by hand.

The team published a paper in Scientific Reports about AutoLabs on June 25, and the software is also available for other researchers to download on GitHub.

“AutoLabs helps pave the way for a new generation of AI-driven automatic assistants for chemistry research,” said Gihan Panapitiya, a data scientist at PNNL and lead author on the paper. “Agents like AutoLabs can act as automated assistants, as well as reliable, self-correcting partners in the complex and creative process of scientific discovery.”

From design to experimentation

Autonomous science is not new, but it can get complicated. To design an experiment for an instrument like Big Kahuna, the scientist and instrument engineer must have an equal understanding of the experimental goals. The engineer has specialized knowledge about the instrument, while the scientist has specialized knowledge about the experiment—but this knowledge doesn’t necessarily overlap. 

The collaborative effort to design an experiment that Big Kahuna can conduct can take weeks while the scientist and engineer work together to refine the steps.

This is where paper coauthor Heather Job comes in. As a systems engineer at PNNL, Job’s role is to operate and program Big Kahuna and train others how to use it. She would work with a scientist to design an experiment for Big Kahuna.

“We're lucky in the fact that we have software that goes along with our robots. But unless you have a lot of training and you know what the robot is capable of, it can take a really long time to translate what you’d like to do into the robot’s specific operations,” Job said.

So Panapitiya and colleagues Emily Saldanha and Olivia Hess set out to explore whether they could create AI agents using readily available large language models to make Big Kahuna more accessible to more scientists. The result, AutoLabs, is an agentic AI system built on an OpenAI model. Agentic AI systems operate with several specialized “sub-agents” that are programmed with unique sets of expertise. Once given a task, they all act under one “supervisor” agent. 

Think of a computer program that could book an entire vacation for you: it might have sub-agents that check plane ticket prices, hotel deals, even the weather over the dates of travel. For AutoLabs, the sub-agents work together to analyze a user’s experimental request and translate that request into specific instructions for Big Kahuna. 

Putting AutoLabs to the test

Once Panapitiya and the team built AutoLabs, it was time to test whether it could successfully translate a desired experiment into instructions compatible with Big Kahuna.

The team developed five different experiments for Big Kahuna, with each experiment increasing in complexity. They then told AutoLabs what tasks they’d like Big Kahuna to complete. All the tasks involved mixing, heating or stirring different chemicals together, but the more complex experiments involved more chemicals, more constraints and more math. 

For instance, the simplest experimental task involved creating and mixing combinations of naphthalene and methanol in a single set of vials, with each sample containing a different amount of each chemical. A more complicated task involved multiple sets of vials and performing a chemical reaction using more than two chemicals, with other steps like heating and cooling at specific temperatures, stirring at specific rotations per minute and transferring chemicals from one vial to another.

In every instance, AutoLabs successfully translated the descriptions of the experiments into steps for Big Kahuna, Panapitiya said. That means it’s performing nearly as well as a scientist who is well trained to use the laboratory robot. 

“With AutoLabs, human experts can learn to use Big Kahuna quickly and guide the overall experimental strategy while the AI agent manages the granular implementation and validation,” Job said. “The collaboration creates a partnership that’s more robust than either a human or robot could achieve alone.”

She added that AutoLabs is not meant to replace a scientist, but to enhance the scientist’s process with speed and efficiency.

The team is now building out AutoLabs to conduct literature reviews and to learn over time, which means adding memory, Panapitiya said. AutoLabs is also flexible in its design and can be adapted to any kind of autonomous laboratory system.

“This iteration of AutoLabs was developed to generate hardware files for Big Kahuna, but its experiment design capabilities extend beyond this one robot,” Panapitiya said.

The work was supported by PNNL’s Generative AI for Science, Energy, and Security Science & Technology Investment under the Laboratory Directed Research and Development Program.