A.I. robotaxis already running in China could be coming soon to the U.S.

June 14, 2022, 7:00 PM UTC

Forget driver’s ed. Soon teaching a car to drive will be more effective than teaching a human.

Artificial intelligence may be invisible to the unsuspecting eye, but it’s already at work on city streets in Shanghai and Suzhou, where cars are teaching each other how to drive themselves, thanks to a partnership between SAIC Motor, China’s largest automaker, and Suzhou-based artificial intelligence company Momenta.

Together the companies deployed 60 robotaxis across the two megacities from December 2021 through March 2022 to validate the technology for use in the cars of the future. Using a smartphone app, thousands of city dwellers in China completed their errands by summoning the robotaxis staffed with human operators who could take control in a potential emergency. Momenta says that 80% of riders used the app at least twice after their initial trial, demonstrating the “vast potential” of the service.

Now the company is working toward scaling its technology for General Motors, Mercedes-Benz, and dozens of other automakers with the goal of launching inside passenger vehicles within the next several years. But first, Momenta CEO Cao Xudong says, it must amass 100 billion miles of driving data.

Momenta headquarters in Beijing, China.
Courtesy of Momenta

Dozens of companies in the U.S. and around the world, including Cruise and Argo AI, are developing next-generation systems for autonomous cars. But Momenta’s strategy may be positioned to bring it to fruition faster. Like others, its machine learning observes and clocks millions of maneuvers made by human drivers, but instead of using those actions as the basis for hardcoded driving rules, Momenta lets the data instruct the other cars directly.

The company also has two key advantages over competitors: access to China’s high-definition road maps and significant partnerships with mass-market carmakers who are already collecting data from tens of thousands of cars on the road. General Motors, Mercedes-Benz, Toyota, auto parts supplier Bosch, and others have invested more than $1 billion in Momenta to accelerate the development of autonomous driving.

“The challenge of solving scalable autonomy is massive,” Xudong tells Fortune. “When customers see we are able to develop algorithms for mass production programs, they understand our capability.” 

The traditional rules-based approach to machine learning programs the cars to respond to a stimulus with a specific action. “If X happens, you do Y,” says Ram Vasudevan, associate professor of mechanical engineering at the University of Michigan. “If the car’s camera reads a stop sign, you stop.” Developers hardcode the software with rules for various maneuvers such as what to do when changing lanes or how to move around a stationary vehicle. When making a lefthand turn, for instance, the car is programmed to precompute its trajectory and slow down accordingly.

But anyone who has driven a car knows that traffic situations aren’t always so straightforward. Take the classic example: A child runs out into the street in pursuit of a ball, leaving the driver to decide whether to brake or swerve. Most companies pre-program a lane-change maneuver, but Momenta’s data teaches the car to consider other factors, such as the weather and time of day.

That’s why Momenta forgoes rules-based algorithms to train cars from scratch. “What Momenta is doing is truly artificial intelligence,” Vasudevan says.

Riders can summon a robotaxi through the SAIC Mobility App with one-click.
Courtesy of Momenta

The “data-driven,” end-to-end learning approach is much harder and more expensive to pull off. The millions of miles its partners log on the road help cars learn to drive in a range of environments, and the only way to generate such volumes of data is to collaborate with automakers who are already investing billions of dollars toward mass producing their own autonomous vehicles.

The advantage of Momenta’s approach is that it can handle more complex situations and nuances. “How do you know that you’ve preprogrammed all the rules out there?” Vasudevan says. “The challenge is that you’re going to need lots of labeled data in order to create a rule for every driving scenario. Even after all of that, there may be a situation that you haven’t programmed into your autonomous stack.”

The market for autonomous driving is growing rapidly. By 2025, Momenta expects that nearly 60 million vehicles will come equipped with Level 2 or 3 advanced driver assistance systems, the stepping stone to fully self-driving cars.  

In the U.S., federal regulators ruled in March that manufacturers no longer need to equip fully autonomous vehicles with steering wheels, brake pedals, or certain other manual driving controls. This paves the way for American automakers to begin testing their models at scale on public roads.

GM-owned Cruise and Alphabet’s Waymo received permits from the California Department of Motor Vehicles to operate robotaxis on public roads under limited conditions, and Tesla CEO Elon Musk said during the company’s quarterly earnings call in April that the automaker plans to develop a robotaxi without a steering wheel or pedals by 2024. “I think that that really will be a massive driver of Tesla’s growth,” Musk said.

Sign up for the Fortune Features email list so you don’t miss our biggest features, exclusive interviews, and investigations.

Read More

Artificial IntelligenceCryptocurrencyMetaverseCybersecurityTech Forward