Researchers at Penn Engineering recently launched a project designed to safely bring artificial intelligence agents into the physical world.
The new research, based in Penn Engineering’s xLab, is led by Electrical and Systems Engineering professor Rahul Mangharam. The lab explores how large groups of AI agents — systems that act on behalf of users — can operate together without causing harm in a high-stakes environment.
The team’s current goal is to design algorithms that integrate AI and machine learning into “life-critical systems,” Mangharam told The Daily Pennsylvanian. The research, he added, is especially important as the field transitions towards “embodied intelligence,” or physical AI systems.
While virtual agents can be scaled relatively easily, physical agents require engineers to consider spatial and reactive safety to ensure they avoid collisions — limiting ability to scale beyond small groups of agents.
According to Mangharam, most algorithms “don’t scale beyond a single digit number of agents.”
Haimin Hu, a postdoctoral researcher in the xLab, said that their team is working with a scale far beyond that, from hundreds to thousands of agents.
He told the DP that the scaling process also introduces real-world constraints such as communication delays and challenges with “human-AI interactions.”
Researchers must also ensure their designs allow thousands of independent agents to make decisions simultaneously without conflict.
RELATED:
‘Holes’ in federal AI healthcare regulation should be patched, Penn Med faculty say
Penn lab uses AI models to track political biases across news publications
“We are interested in understanding how to scale these systems up and how to have them develop their own skills and form teams,” Mangharam said. “Rather than assigning teams, [agents] can observe others and decide how to partition the problem and work together.”
One of the biggest barriers, Hu explained, is bridging the gap between theory and real-world deployment.
“The theory has been developed in the past couple of decades, but they haven’t been able to apply to real-world settings and make them work for physical systems like drones and vehicles,” he said.
In the future, the research has potential in aerial transportation and air traffic control systems, Hu stated.
Mangharam also noted the challenge of integrating AI into real-world systems without compromising safety.
“We want to make sure that as AI goes from a simple ChatGPT helping you write a document to working in more critical aspects,” he said. “We are developing the technologies that ensure they are safe.”
RELATED:
‘Holes’ in federal AI healthcare regulation should be patched, Penn Med faculty say
Penn lab uses AI models to track political biases across news publications
Senior reporter Saanvi Ram covers undergraduate sciences and can be reached at ram@thedp.com. At Penn, she studies health and societies. Follow her on X @Saanvi_vivi.
