Peter Stone Leads the Way at the University of Texas
FRANKLY, IT DOESN’T look like much. Eight robots you might have once found on Toys ‘R’ Us shelves occupy a small field. They collectively move with the grace of a toddler, just at the pace of someone much, much older. (If one bot goes down, it might be a minute before it gets back up.) A small ball you could mistake for a dog toy seems to be the object of everyone’s attention. But if you squint, it’s soccer. Soccer-ish, at least.
Even if this game doesn’t move with similar speed, the flashes of brilliance and creativity on display here are every bit as impressive as what Messi or Ronaldo do. That’s because in RoboCup—the now annual engineering competition where the long-term goal is to have a team of humanoid robots that can beat the World Cup champs by 2050—all of the action happens solely at the discretion of these devices. It started small in 1997, but in 2019 more than 170 teams from across 30 countries competed head-to-head. If you’re into artificial intelligence, this may be its premier competition. And there may be no better evidence of RoboCup’s significance than the fact that the newly minted executive director of Sony AI America credits his 20-plus years of competition with helping him land this industry dream job.
“I can’t say enough about how it’s influenced my career and really the careers and research of thousands of people around the world,” says Dr. Peter Stone, Sony’s choice to lead the company’s American artificial intelligence initiatives. “RoboCup brings together many, many research challenges—there’s computer vision involved, locomotion and movement, low-level control. Soccer is the central motivating theme, but it’s expanded beyond that.”
Stone has long been drawn to RoboCup. As a student pursuing his Ph.D. in computer science at Carnegie Mellon in the ’90s, he was researching the idea of robot soccer as an application domain for artificial intelligence before the competition even officially started. He was also admiring the then-newly released Sony AIBO, a revolutionary consumer AI product that would ultimately become the original main platform for RoboCup teams. (“It could sense through cameras, it had enough processing to do real complex decision making, it had well-engineered actuators and motors so it could walk quickly, go in multiple directions, and manipulate a ball,” Stone says. “It was way ahead of its time in the robotics industry.”) When the time came to organize the first RoboCup in 1997, Stone was the only person from the U.S. involved with the planning because his adviser at the time, AI heavyweight Manuela Veloso, was involved, too.
Stone remained part of that community from then on as he advanced within his own career and the larger AI community. Since becoming a faculty member at the University of Texas at Austin in 2003, he’s led a team of graduate students to participate in RoboCup every year. They’ve even become a bit of a dynasty in a newer side competition where robo-soccer is played out through 3D simulations. (The squad has won eight of the last nine years.) And despite continuing to run Cogitai, the AI startup he founded in 2015, Stone happily found time to accept the role of president for the RoboCup competition in 2017.
This new role also reunites Stone with a longtime RoboColleague: Hiroaki Kitano, the president and CEO of Sony Computer Science Laboratories Inc., who will oversee Sony’s AI initiative globally.
How does one AI professional go from robot soccer to helming an industry leader like Sony AI? The path centers on Stone’s two areas of expertise, which happen to mesh beautifully with the beautifully robotic game and with Sony’s potential vision for AI: reinforcement learning and multi-agent systems.
At the most basic level, reinforcement learning is a type of machine learning focused on sequential decision making, how to take actions over time, essentially. If a robot on the UT team identifies the ball is at its feet early in the game, it will likely try to advance that ball toward the opponent’s goal with the intent to score. But say UT is up late in the game, and this bot has been trained on thousands of previous matches and situations. Upon receiving the ball, instead of simply automatically moving it forward again, perhaps the bot will aim to maintain possession, understanding that this is the more valuable move in this late-match situation.
In RoboCup, this kind of AI/ML work intersects directly with Stone’s other specialty, multi-agent systems. This concept is exactly what it sounds like: how do individual autonomous agents learn to collectively work together and make evolving decisions? These robot soccer matches, after all, are not one-on-one competitions. Rather, they feature three field players and a goalkeeper on each side. Just like in real life, teamwork becomes pivotal. And the team needs to understand scenarios just like an individual does.
When Sony made its surprise announcement about a dedicated AI division with a U.S. branch based in Austin last fall, the company also shared three initial focus areas: imaging and sensing; gastronomy; and gaming. That last field is specifically what Stone and his U.S. team will dig in on, and suddenly the pairing of task and talent makes perfect sense.
Sony has a long history of industry-leading gaming offerings—from the PlayStation to developing games like The Last of Us to dabbling in handheld and VR formats. And gaming looks like a particularly good area to apply the reinforcement learning and multi-agent systems expertise of Stone and his team. Again, reinforcement learning is about autonomous systems improving with experience and changing tactics over time, and multi-agent systems involve improving interactions between autonomous entities. So when a gamer decides to take on a computer team in a game like Overwatch—a team-based multiplayer first-person shooter where the objective is to maneuver around compounds and take down your opponents first—it’s easy to see how the work of Sony A.I.’s US office can be immediately applicable.
“Well, not ‘easy’ in the sense it’ll be easy to do,” Stone jokes. “But ‘easy’ in that we’re an obvious match.”
When Stone talked with Cognitive Times in late 2019, he wasn’t quite ready to discuss specifics, though. He couldn’t say whether work at Sony AI U.S. was building toward particular games or focused on specific hardware. He did, however, confirm that work is well underway and that ultimately the effort from his team will include some public-facing products. So expect Sony AI to maintain a presence at the big AI industry conferences (say NeurIPS or AAAI) and for some of the work Stone oversees to eventually reach Sony’s gaming audience.
Being right for a job doesn’t mean a job is right for you, of course. And last year Stone had no shortage of commitments between his startup, his research and teaching at UT, and his role in industry events like RoboCup. He had no shortage of likely suitors, either. Back in 2015, Stone founded Cogitai to focus on creating a “world-class, general-purpose reinforcement learning platform.” Stone wanted the company to become an industry leader in this particular application of AI/ML, and having that kind of prowess (along with the associated staff expertise and proprietary software) means major industry players will quickly start knocking on the door for partnerships or acquisitions.
But even setting RoboCup aside, Stone has long been intertwined with Sony throughout his career. On the small end, he’s presented with or sat on panels alongside Sony AI experts at many conferences, including discussing the future of AI and creativity with Kitano and chess legend Gary Kasparov at SXSW 2019. On the larger end, Stone in fact professionally partnered with Sony long before ever becoming a direct employee of the company. Sony was among the earliest investors in Cogitai… following conversations Stone had with Kitano and other Sony-based RoboCup colleagues. These interactions and relationships built over two decades later made Sony the logical place for Stone.
“My relationship with Sony, it goes back to the 1990s. And now my team at Cogitai ended up becoming the Sony A.I. US branch as everything launched,” Stone says. “Any time there’s a bringing together of a new team and a new organization, there’s a startup phase and a lot of relationship building. So we’re still transitioning, but absolutely there’s a long history of work we’ve done that we expect to bring immediate value.”
For Stone, ultimately Sony’s focus area pitch may have been as enticing as his history with the company.
At a time when AI is often portrayed as a concept to remain uneasy about—Democratic presidential candidate Andrew Yang has prominently called automation a threat to the future of the country and China’s AI advancements a threat—Sony AI is taking a decidedly different approach. Its initial focus areas aren’t framing AI as a human replacement, rather as a human enhancer.
“With many AI research organizations, there’s a lot of envisioning of what the long-term future may be and how to have a long-term impact. And a lot of those AI discussions are about replacing people and automating jobs that people are doing,” Stone says. “But Sony is a leader in the spaces of creativity and entertainment, so there’s a lot of untapped opportunity there [to apply AI]. I think there are opportunities within products, making them more intelligent, engaging, and adaptive to individuals to just bring out the best in people. One of the exciting aspects of this for me is, using AI for creativity is much more about magnifying and bringing out the full potential of people—not replacing them. There’s a feel-good sense to these types of applications.”
Feeling good about artificial intelligence may feel like a far-off goal, but it’s not impossible. Stone recalls a period before he entered the field where the promise and hype around AI rose to a fever pitch and public sentiment clustered around excitement. But when the industry failed to deliver, only public skepticism remained. “In the field we called it the ‘AI Winter,’” Stone says. “People had become disillusioned through overhype and failed promises.”
But that down period in public conversations around AI might be pivotal in retrospect—it allowed a new generation of thinkers and engineers to develop an interest without a ton of professional competition, and when these future AI industry leaders started their careers they found lots of room for innovation (for ideas like starting a robot soccer league, for instance). Stone considers himself lucky to have “grown up at the right time and been at the right place” in that sense, and now he’s trying to cultivate those same opportunities for the next generation while simultaneously helping Sony achieve something special.
So for the time being, Stone will continue to teach at UT and push his grad students for more. He’ll finish out his three-year term as RoboCup president and push the competition forward as the premier environment for AI experimentation (and thus AI professional growth). And if the new executive director has his way, the industry will avoid another winter as Sony both delivers exciting applications and shows a way forward for those leery of AI.
“Over the years we see cycles in AI, and we’re at a very absolutely high point now within the field. But it’s important we learn the lessons from the past, and the tendency has been to over-romanticize and overhype what the tech can do,” Stone says. “There has been a whole lot of progress, but there’s a long way to go for the full dreams of people in AI. So I strive to keep an even keel in regards to the successes and challenges yet to come, but it’s been a privilege to be a part of this journey of the whole field. And now it’s great to be a part of a company that has played a central role in the history yet still has a lot to contribute.”