An artificial intelligence program has beat a human at chess, checkers, the game of Go, and Jeopardy.
Now GT Sophy, an AI agent has defeated some of the world’s best drivers of the PlayStation racing simulation game, Gran Turismo Sport.
Sony AI with Polyphony Digital and Sony Interactive Entertainment announced the breakthrough in artificial intelligence recently. And the achievement has ties to the University of Texas at Austin.
It’s a huge milestone for the advancement of robotics, said Peter Stone, professor of computer science at UT Austin and director of Texas Robotics.
Gran Turismo involves real-time strategic, tactical decision making and split-second reactions and GT Sophy mastered those capabilities, said Stone. He is also executive director of Sony AI America and is a core member of the team that worked on the project. He is one of the authors of a paper published in Nature titled “Outracing Champion Gran Turismo Drivers with Deep Reinforcement Learning.” Four of the authors are affiliated with UT Austin.
Gran Turismo is a very real simulation in the sense that it models the friction of real race cars and resistance, Stone said. In fact, real race car drivers use it for training, he said.
The robot has learned social navigation, Stone said. People do that smoothly and seamlessly. If there is a crowd that lets out after a football game at UT, people leave the stadium and cross the street without running into each other because they can use their senses, experience, and knowledge to navigate through the crowd. That is what the researchers have taught GT Sophy.
“What we’ve done is created a computer controller,” Stone said.
Right now, the AI technology is in the laboratory stage, but researchers are working closely to integrate it into a future version of Gran Turismo to enhance the gameplay experience, he said.
The focus of Sony’s work in AI is to enhance human creativity, Stone said.
“This is an example of that,” Stone said.
In 1997, the IBM supercomputer called Deep Blue beat Garry Kasparov, the world’s best chess player, but that didn’t mean people quit playing chess, Stone said. It’s the same with AI in video games, he said. AI will enhance the game playing experience, but it will not replace humans, he said.
At UT Austin, Stone and his team are training robots to play soccer in a RoboCup, but those robots even if they defeat professional soccer players will not end the sport, Stone said.
“The same thing here,” Stone said. “Even if computers are better than people at driving race cars, the F1 will still remain a human sport.”
Austin has become one of the premier centers in the world for robotics research. In 2002, when Stone joined the faculty, there were only two roboticists on campus and today there are 16 faculty members. Texas Robotics has several industry affiliates including Amazon, Sandia, and Apptronix. It is also in talks about working with Tesla, Stone said.
The AI technology developed for GT Sophy does present new opportunities that may be used in areas such as autonomous racing, autonomous driving, high-speed robotics, and control, Hiroaki Kitano, CEO of Sony AI, said in a news release.
But right now, there are no plans to use the technology behind GT Sophy in autonomous cars, Stone said.
“Training a video game is different than training in real cars,” Stone said.
To perfect GT Sophy, Sony AI tested the AI agent by racing against four of the world’s best Gran Turismo drivers in two events on July 2, 2021, and Oct. 21, 2021. GT Sophy learns continuously by playing the game. The AI agent made improvements based on the results of the first race and then outperformed the human drivers in the second event in October.
GT Sophy was trained to master race car control, racing tactics, and racing etiquette to avoid at-fault collisions and respect opponent driving lines.
“One of the most exciting challenges of our time is how can we use AI to enhance what people do,” Stone said. “We are using AI to augment human capabilities rather than to replace them.”
The full press conference announcing the GT Sophy breakthrough is posted below.
