AiThority Interview with Pete Wurman, Director at Sony AI

Please tell us a little bit about your journey in the AI and ML technology space. How did you start at Sony AI?

I got my undergraduate degree from MIT in mechanical engineering. I didn’t feel ready to be an engineer, so I went to the University of Michigan to get a master’s degree in mechanical engineering. Along the way, I got a job programming at the university, and eventually decided to go back to school and get a Ph.D. in computer science. From there, I became a professor in the Computer Science Department at North Carolina State. In 2004, as I went up for tenure, my roommate from my undergraduate days at MIT came up with an idea for a robotic warehouse system and convinced me to help him start what became Kiva Systems. In 2012, Amazon bought Kiva and started deploying our robots to their warehouses. The division is now called Amazon Robotics and has deployed approximately 500,000 robots. In 2015 I briefly retired, but soon joined Cogitai, a small AI startup focused on reinforcement learning. Sony acquired Cogitai in 2019 and founded Sony AI.

Please tell us more about the first global release of Gran Turismo Sophy for Gran Turismo 7 and your role in this achievement?

We first unveiled Gran Turismo Sophy, Sony AI’s breakthrough AI racing agent, in February 2022, in conjunction with the publication of the research on the cover of the February 10^th, 2022  issue of Nature, The aim of GT Sophy was to deliver new AI-powered gaming experiences to players around the world.

Since that time, Sony AI, in partnership with PDI and Sony Interactive Entertainment, has continued to evolve and refine GT Sophy’s capabilities and explore ways to make it more accessible to the greater Gran Turismo driver community via a software update to Gran Turismo 7. While we continued to train the agent as we had for the Race Together events, to deploy GT Sophy to the community we had to integrate it directly into Gran Turismo 7. We worked closely with PDI to incorporate the neural network into the game engine.

With the launch of Gran Turismo Sophy ‘Race Together’ mode we have accomplished that goal of advancing our game AI breakthrough from research project to commercialization and making it accessible to players of all levels.

The special mode is currently available as a limited-time event until the end of March, at which point we will assess player feedback and use that input to continually improve the GT Sophy Race Together mode feature for future releases.

I served as the lead author on the research and the and lead the team of more than 25

researchers and engineers that worked on the project from the U.S., Canada, Japan, Switzerland, Sweden, and Italy.

Recommended: AiThority Interview with Keith Hartley, CEO of LevaData

What is the difference between previous in-game AI and GT Sophy?

The built-in AI is rule based AI that tries to follow a specified driving line and has only a rudimentary notion of sportsmanship. GT Sophy is trained using reinforcement learning and makes decisions that are more contextually nuanced.

In comparison to the regular built-in AI, GT Sophy creates more fun and interactive wheel-to-wheel racing. This drastically improves the quality of the races, making them less predictable and more human-like, which is a first in racing games.

In the era of Metaverse and NFTs, how does Sony AI continue to lead the market with its breakthrough AI?

Sony AI was founded with the mission to “unleash human imagination and creativity with AI” and our work builds on Sony’s long history of creative and technological innovation in entertainment on a global scale.

As part of Sony Group, we have access to Sony Group’s diverse business portfolio and technical assets, including imaging and sensing technology, robotics technology, and entertainment assets such as movies, music, and games to support that mission. We also have on our team some of the world’s top AI researchers and engineers from around the world, including some of the best reinforcement learning experts.

For the launch of GT Sophy specifically, the accomplishment is the result of a unique collaboration between Sony AI Inc., a cutting-edge AI research organization, Polyphony Digital Inc. (PDI) the creative studio behind the world-famous Gran Turismo games, and Sony Interactive Entertainment (SIE), the PlayStation company.

Recommended: AiThority Interview with Shaun McGirr, Field CDO at Dataiku

Tell us more about the AI engine at the backdrop of this gaming revolution?

GT Sophy is an autonomous AI agent trained utilizing a novel deep reinforcement learning technique developed by Sony AI.

You can read more about the technical specifications and accomplishments that made GT Sophy possible in our Sony AI blog series on the making of GT Sophy.

What are the major milestones in gaming AI technology that you have witnessed in the last 5 years? How does Sony AI lead the pack and race ahead of its competitors?

There have been other game AI breakthroughs to date – starting in the 1990’s with Kasparov vs. IBM’s Deep Blue and continuing into the last decade with AI agents that have mastered the complex decision-making games of chess, shogi and go as well as other real-time, multiplayer strategy games such as StarCraft and DOTA.

While GT Sophy shares much with these previous milestones, especially the use of neural networks and reinforcement learning, it differs from them in important ways.

First, Gran Turismo is a physically realistic simulation of motorsports. Racing is a very unforgiving sport; to be competitive a driver must be pushing the car to its absolute traction limits and one small mistake leads to disaster. So it must learn not only how to drive fast, but how to master tactical driving skills like slipstream passing — all while making the decisions in real time.

At the same time, because competitors can physically interact with each other, there is an element of sportsmanship that is not present in other video games that is very hard to teach to the agent. The agent had to learn how to play in the same physical space as the humans without being able to practice against actual humans (there are too few humans at these elite levels). Unlike other games, playing against yourself does not fully prepare you to play against humans; slight errors/differences by the human can lead to penalties for the agent if it isn’t prepared.

Recommended: AiThority Interview with Tal Shaked, Moloco’s Chief Machine Learning Fellow

Thank you, Pete ! That was fun and we hope to see you back on AiThority.com soon.