Losing traction while driving at high speed is often very bad news. Scientists at the Toyota Research Institute and Stanford University have developed a pair of self-driving cars that use artificial intelligence to do so in a controlled manner (a trick better known as “drifting”) in order to push the limits of autonomous driving.
The two autonomous vehicles performed the daring tandem drift maneuver at Thunderhill Raceway Park in Willows, California, in May. promotional videoThe two cars roar around the track within a few feet of each other after the human drivers relinquish control.
Chris GerdesGerdes, a Stanford University professor who led the project’s involvement, tells WIRED that the techniques developed for the feat could eventually aid future driver-assistance systems. “One of the things we’re studying is whether we can do it as well as the best human drivers,” Gerdes says.
Future driver-assistance systems could use the algorithms tested on the California track to intervene when a driver loses control, thereby getting the vehicle out of trouble like a stunt driver would. “What we’ve done here can be scaled up to address larger problems, such as automated driving in urban scenarios,” Gerdes says.
The project is a clear demonstration of high-speed autonomy, though self-driving vehicles are still far from perfect. After a decade of promises and hype, taxis now operate without drivers in some limited situations. However, the vehicles are still prone to getting stuck and may require remote assistance.
Researchers at Toyota and Stanford University modified two GR Supra sports cars with computers and sensors that track the road and other vehicles, as well as the cars’ suspension and other properties. They also developed algorithms that combine advanced mathematical models of tire and track properties with machine learning that helps the cars teach themselves to master the art of drifting.
Lin MingA professor at the University of Maryland who studies autonomous driving, she says the work is an exciting advance toward helping self-driving cars operate in extreme conditions. “One of the biggest challenges for autonomous vehicles is operating safely on rainy, snowy or foggy days, or in poor lighting at night,” she says.
Lin adds that the Toyota-Stanford project demonstrates the importance of combining machine learning with physical models of the world. “While it is only a preliminary demonstration, it is clearly going in the right direction,” he says.
In 2022, Toyota and Stanford first demonstrated algorithms that allowed self-driving cars to drift. Getting two vehicles to perform that trick in unison requires even better control and involves the vehicles communicating with each other. The cars received data from laps completed by professional drivers. Their respective computers calculated an optimization problem up to 50 times per second to decide how to balance steering, accelerator and braking.
“What we’re really studying here is how to control the car at the extremes of performance, when the tires are spinning, the kind of situation you would encounter when driving on snow or ice,” said Avinash Balachandran, vice president of TRI’s Human Interactive Driving division. “When it comes to safety, being an average driver isn’t enough, so we’re really looking to learn from the best experts.”
The world has seen remarkable advances in artificial intelligence lately thanks to the great language models that power programs like ChatGPT. However, as the dual-drift demonstration highlights, mastering the messy and unpredictable physical world remains an entirely different proposition.
“In an LLM, a hallucination may not be the end of the world,” Balachandran says, referring to the way large language models can get data wrong. “Obviously, that could be very different with a car.”
