Teaching machines to engineer machines

Can a machine design, adapt, and improve other machines—safely, reliably, and without constant human oversight? Answering this question is at the heart of Yorie Nakahira’s research on next-generation autonomous systems.

Nakahira, assistant professor of electrical and computer engineering, has received a Young Investigator Award from the Japan Science and Technology Agency (JST) for developing a machine that builds autonomous systems.

Adaptation is essential because real-world environments are constantly changing. Robots often operate with limited data, limited memory, and limited computing power, making it unrealistic to train them once and expect safe performance forever.

To address this, Nakahira explores sequential fine-tuning, where systems learn gradually as new information becomes available. This enables continual adaptation rather than reliance on a single, massive training dataset.

However, adaptation introduces its own risks: the model can forget past skills or become overly confident. Nakahira’s work also emphasizes uncertainty awareness—teaching systems to recognize when they are unsure and to act cautiously in those moments.

If we succeed, anyone, not just engineers, could configure a robot to perform tasks safely and customize its behavior to match personal preferences or workplace norms.

Yorie Nakahira, Assistant Professor, Electrical and Computer Engineering

“A system needs to know what it doesn’t know,” she explains. “That awareness is critical for avoiding dangerous decisions in unfamiliar or ambiguous situations.”

Today, language models, learning algorithms, and control systems are often developed independently. This fragmentation makes it difficult to translate human intent into safe physical actions, especially in complex or changing environments.

Nakahira’s long-term vision is a unified framework that integrates language understanding, learning, uncertainty estimation, and control. Such systems would not only adapt more robustly and respect physical constraints but could also automatically discover new strategies for adaptation.

Ultimately, this foundation could enable AI systems to assume roles traditionally reserved for human experts in control and systems engineering—designing, tuning, and validating autonomous behavior on their own.

“If we succeed,” Nakahira says, “anyone, not just engineers, could configure a robot to perform tasks safely and customize its behavior to match personal preferences or workplace norms.”