Physical AI Fuels the Machines of Tomorrow

Carnegie Mellon University has long been at the forefront of robotics and autonomy. Now, as artificial intelligence increasingly moves into the physical world, the university is shaping what that future will look like. 

Ahead of the AI Horizons Summit, Martial Hebert, dean of the School of Computer Science and a speaker at the conference, shared his perspective on physical AI — what it is, why it matters now and how it’s set to transform industries and create new opportunities. 

How do you define physical AI? Why is it important now?

Physical AI refers to any system that interacts with and adapts to the physical world. Unlike digital AI agents, physical AI involves sensing, reacting and making decisions in real-world environments. If a physical system does not have that intelligence, it's just a machine. 

Physical AI has always been important, but now we are quickly advancing in areas such as vision and decision-making that support applications not possible before. I am confident we are going to see more exciting developments in the quality of precision in machines like autonomous vehicles, humanoid robots or robots in health care that all rely heavily on physical AI to complete tasks safely and accurately. 

CMU has been a leader in robotics and autonomy for decades. What role is the university playing now in advancing physical AI? 

At CMU, we focused on intelligence and autonomy long before we had the capabilities we have now. For decades, we embraced physical systems as much as digital. 

For example, we created the Robotics Institute 45 years ago within the School of Computer Science to embrace a view of robotics centered around intelligent physical systems, grounding the work in physical AI long before it was a common term. This philosophy gave us our standing at the forefront of robotics and robotics research, a position that continues today. 

We still prepare students to lead in the field because we teach flexibility and adaptability. Our researchers incorporate intelligence into real-world systems that have applications in an incredible variety of industries. With our advancements in AI, we are creating levels of decision-making that did not exist before, seen in projects like BrickGPT from a team in SCS that was supported by our Manufacturing Futures Institute(opens in new window). 

How do you see physical AI transforming industries like transportation, health care or manufacturing? 

All of those industries involve some sort of physical interaction, like moving materials in manufacturing and self-driving cars for transportation. Having AI injected into these applications gives the machines a more competent level of decision-making, enhancing efficiency and safety at the same time. In all these industries, we must be able to trust the machines to work alongside humans. Physical AI is the answer to that. 

What are some of the toughest technical challenges researchers are still working to solve? 

One of the biggest challenges for anything physical is gathering data. With generative AI, there are large amounts of data like images, videos and text that are readily available. But what is the equivalent for physical systems? Acquiring data at even a fraction of the volume that is used in a typical AI system is extremely difficult. 

Researchers are also tackling how to characterize and validate performance in physical AI. When a robot operates in an environment, there are no guarantees of what that robot will do. The next frontier for physical AI will be developing guarantees of performance to ensure we are building systems we can trust. 

Pittsburgh has become known for AI startups. What opportunities does physical AI create for new companies, and what excites you about the local ecosystem?

Pittsburgh has a history of leading research in robotics and AI due to the large number of talented researchers that find their place here. Physical AI presents opportunities not only for new startups, but also for established companies and with industry partners working with CMU. 

Physical AI allows companies to approach problems from a broad perspective — one that is not siloed into one particular view. With tech companies in Pittsburgh hosting so many experts in so many different areas, it is really the place to be for physical AI innovation. 

At CMU, we teach students how to be entrepreneurs and practitioners in their field, and they continue to be the type of talent the industry is looking for, which creates an exciting local ecosystem. 

What skills or training will tomorrow’s engineers and researchers need to succeed in this emerging field? 

The important thing is that we not just teach the right tools, platforms and techniques that are relevant today, but that we help researchers understand how these techniques will undoubtedly evolve. Engineers and researchers must be able to develop and try new approaches to keep up with these evolutions. Whatever people learn today, there's going to be different approaches a few years from now. It is essential to not only learn the foundations of computing and AI, but to also be prepared for the changing landscape in the future.