People of ACM - Peter Stone

June 3, 2025

Why is this an exciting time to be working on developing intelligent learning agents?

Well, in my opinion, it’s always been an exciting time to be working on developing intelligent learning agents. That’s what I’ve been doing non-stop for the last three decades! I see the creation of robust, intelligent agents as the persistent long-term goal of my research career. And while not everyone in the field of AI shares this goal, many people do.

I consider an autonomous agent to be a computer program (or robot) that can sense, decide, and act in a closed loop, with no need for human intervention. It may have a need, or an opportunity , to interact with people along the way. But it should also be able to persist on its own over an extended period of time.

As highlighted on my personal webpage, I consider adaptation, interaction, and embodiment to be essential capabilities of such agents. Thus, my research focuses mainly on machine learning, multiagent systems, and robotics.

Throughout my career, there has been steady advances in some of the underlying, fundamental capabilities of robots and AI technologies, each of which has presented new opportunities for autonomous agents. Most recently, generative AI tools have opened up a trove of new possibilities for intelligent learning agents, either by incorporating large pre-trained models into existing agent architectures, or by extending these models with sensing and acting capabilities. These tools have made it a particularly exciting time for the field. But I only expect things to get more and more exciting in the coming years as new algorithms and tools continue to be developed!

You have led the technology development for teams in the international RoboCup Competitions (your team, UT Austin Villa, has won many trophies). To score goals, the robots must sense, decide, and act, both with robots (their teammates) and against robots (members of the opposing team). Aside from the physical mechanics of the robots, what has led to their improvement in soccer since 1997?

The physical mechanics have indeed improved a lot. But many other things have improved dramatically as well. The amount of computing power that is feasible to put onboard a robot has increased many times over. And algorithms for computer vision, robot localization, and agile locomotion have also improved a lot. There are also many new algorithms for learning and decision-making. Really, looking back to 1997, all aspects of the robotics stack have improved considerably.

Will you provide a few examples of how these innovations might be applied in other areas?

Within RoboCup, there are several leagues beyond soccer that involve physical, interactive autonomous robots. For example, there is a league focusing on disaster rescue, one on general-purpose service robots in a home environment, and one on industrial applications. We’ve often seen cross-pollination of ideas and algorithms across these leagues.

Meanwhile, there have been several companies that have gotten their start from RoboCup. Perhaps most notably, the founders of Kiva Systems, which was purchased by Amazon, credit RoboCup as helping inspire their designs for fleets of warehouse robots.

Your research community has set an ambitious goal of developing a robot team that can beat a human team by 2050. What makes you think this goal is achievable?

To be honest, I’m not sure if it is. I’ve always thought of it as an aspirational goal. 2050 is still a long way away. But there is still a large gap between the physical agility of professional soccer players and humanoid robots. I’ve often joked that we chose 2050 because by then most of us who were around at the start of RoboCup will have retired: if the target is missed, it won’t be our fault! In the 1950’s, there were predictions that computers would be better than people at chess within 10 years. It did happen eventually, but not for closer to 50 years. On the other hand, computers bested people at the game of go much earlier than many people expected. I hope I live to see whether 2050 ended up being too optimistic or too pessimistic for super-human soccer!

In one of your most cited papers “A Multiagent Approach to Autonomous Intersection Management,” you (along with co-author Kurt Dresner), contend that a new paradigm is needed to handle traffic flow at intersections. What is a key insight of this paper?

Kurt had the inspiration for that paper when he was sitting in his car at a red light while late for a meeting with me. When he eventually burst into my office, he told me excitedly that he had a new idea for an application of multiagent systems. The main idea was that if autonomous vehicles could communicate with an intelligent agent at an intersection, then we could use the space and time within the intersection much more efficiently than stopping all the traffic from going in one direction for minutes at a time—in which case he wouldn’t have been late for our meeting!

Instead, each car could call ahead for a reservation in space-time indicating exactly what parts of the intersection it wanted to occupy at what times. As long as no previous reservation claimed any of those points in space and time, the reservation would be granted. We found that using this protocol, we could get many more cars through intersections in any given period of time than when using traffic signals or stop signs.

You have been actively involved with the AI100 Study, and you (and your co-authors) shared some perspectives about this initiative in a recent Communications of the ACM article. There has been much discussion lately on how AI might impact jobs and the economy. Have your perspectives on this topic changed since you led a study panel for AI100 in 2016?

Perhaps in some of the details. After all, AI technologies have advanced dramatically in the past 10 years. But qualitatively, I think most of what we wrote in that report still holds up. Two quotes from the report are as follows:

“AI will likely replace tasks rather than jobs in the near term and will also create new kinds of jobs. But the new jobs that will emerge are harder to imagine in advance than the existing jobs that will likely be lost.”

“There is even fear in some quarters that advances in AI will be so rapid as to replace all human jobs—including those that are largely cognitive or involve judgment—within a single generation. This sudden scenario is highly unlikely, but AI will gradually invade almost all employment sectors, requiring a shift away from human labor that computers are able to take over.”

To me, these perspectives still seem relevant today.

Given current trends, what might be a novel application of AI in 2035?

If you mean a novel commercial application, I think that most applications that will make it into peoples’ homes by 2035 are already in the pipeline in research labs. I hope that there will be more commercially viable general-purpose robot platforms available for homes and offices. But history has taught us that it can be very difficult to field cost-effective, robust robots.

Meanwhile, I think it will take most, if not all, of the time between now and 2035 for each sector of the economy to fully realize the potential of current generative AI models within their industries. From manufacturing, to healthcare, to entertainment, there will need to be close collaborations between AI experts and domain experts to craft the applications that are most suitable for practitioners in each domain. Prototypes have certainly started to emerge, but I think there’s still a long way to go.

If you mean novel prototype application, if I knew, I’d already be working on it. Or maybe I already am and will let you know when I’m ready to publish!

-->{C}

 

Peter Stone is a Professor at The University of Texas at Austin, where he is the founder and director of the Learning Agents Research Group (LARG). He is also the Chief Scientist of Sony AI. His research interests include machine learning, multiagent systems, and robotics. Applications of Stone’s work include robot soccer, autonomous vehicles, general-purpose service robots, and human-interactive agents.

Among his honors, he is an ACM Fellow and has received the IJCAI Computers and Thought Award, which is presented every two years by the International Joint Conference on Artificial Intelligence. Stone recently received the ACM-AAAI Allen Newell Award for significant contributions to the theory and practice of artificial intelligence (AI), especially in reinforcement learning, multiagent systems, transfer learning, and intelligent robotics.