How China’s First Self-Driving Unicorn Searches For Data

Cao Xudong turned on the side of the road, in jeans and a black T-shirt printed with the word “Momenta”, the name of his startup.

Before founding the company, which surpassed $ 1 billion in valuation last year to become China’s first self-driving “unicorn,” he had already led an enviable life, but was convinced autonomous driving would be the real thing. Big Deal.

Cao isn’t just going on the moon of fully autonomous vehicles, which he says could be 20 years away. Instead, it is taking a two-part approach to selling semi-automatic software while investing in research for next-generation self-driving technology.

Cao, pronounced “tsao”, was pursuing his Ph.D. in mechanical engineering when the opportunity arose to work in Microsoft’s Asia-based research arm, supposedly the “West Point” for China’s first generation of artificial intelligence experts. He stayed there for more than four years before quitting to get his hands on something more practical: a start-up.

“The academic research for AI was maturing at the time,” Cao, 33, said in an interview with TechCrunch, reflecting on his decision to leave Microsoft. “But the industry that implements AI had just started. I believed that the industrial wave would be even more extensive and intense than the academic wave that lasted from 2012 to 2015.”

In 2015, Cao joined SenseTime, now the world’s highest-value AI startup, thanks in part to the lucrative facial recognition technology it sells to the government. During his 17-month tenure, Cao built the company’s zero-employee research division into a strong team of 100 people.

Before long, Cao found himself craving a new adventure again. The founder said he doesn’t care about the result as much as the possibility of “doing something.” That trend was already evident during his stay at the prestigious Tsinghua University, where he was a member of the outdoor club. He wasn’t particularly drawn to hiking, he said, but the opportunity to accept challenges and be with daring people with the same endurance capacity was attractive enough.

And if building self-driving cars would allow you to make a mark on the world, you’re ready for that.

Make the computer, not the car.

Cao led me to a car equipped with the cameras and radars you could detect in an autonomous vehicle, with hidden computer codes installed in the trunk. We jump inside. Our driver chose a route from the high definition map that Moments It was built, and as soon as we got closer to the highway, the autonomous mode kicked in on its own. Then the sensors began to feed real-time data about the surroundings on the map, with which the computer could make decisions on the road.

Personal moment installing sensors to a test car. / Photo: Momenta

Momenta will not make cars or hardware, Cao said. Rather, it gives cars autonomous characteristics by making their brains, or deep learning capabilities. In effect, it is a so-called Tier 2 supplier, similar to Intel’s Mobileye, which is sold to Tier 1 suppliers who actually produce the automotive parts. It is also sold directly to original equipment manufacturers (OMEs) who design cars, order parts from suppliers, and assemble the final product. In both circumstances, Momenta works with clients to specify the final piece of software.

Momenta believes that this approach to asset safety would allow it to develop cutting edge driving technology. By selling software to automakers and parts manufacturers, you not only generate revenue but also generate mountains of data, including how humans intervene, to train your codes at relatively low costs.

The company declined to share who its customers are, but said they include major Tier 1 automakers and suppliers in China and abroad. There will not be many of them because a “partnership” in the automotive sector requires deep and resource intensive collaboration, so it is believed that less is more. What we do know is that Momenta has Daimler AG as a sponsor. It is also the first Chinese company in which the father of Mercedes-Benz has invested, although Cao did not reveal if Daimler is a customer.

Let’s say you operate 10,000 autonomous cars to harvest data. That could easily cost you $ 1 billion a year. “100,000 cars would cost $ 10 billion, which is a scary number for any tech giant,” Cao said. “If you want to acquire seas of data that are significant in scope, you have to build a mass-market product.”

Highway Pilot, the semi-autonomous solution that controlled our car, is Momenta’s first mass-produced software. More will be launched in the coming seasons, including a fully autonomous parking solution and a robotaxi self-driving package for urban use.

In the long term, the startup said it aims to address inefficiencies in China’s $ 44 billion logistics market. People hear about warehousing robots built by Alibaba and, but overall, China is still on the lower end of logistics efficiency. In 2018, logistics costs accounted for nearly 15 percent of the national gross domestic product. In the same year, the World Bank ranked China 26th on its logistics performance index, a global benchmark for efficiency in the industry.

the moments

Cao Xudong, co-founder and CEO of Momenta / Photo: Momenta

Cao, an unassuming CEO, raised his voice and explained the company’s two-step strategy. The two-pronged approach forms a “closed loop,” a term Cao repeatedly called upon to talk about the company’s competitive advantage. Instead of choosing between presence and the future, like Waymo. meets Level 4, a designation given to cars that can operate in basic situations without human intervention, and Tesla with semi-autonomous driving, Momenta works on both. It uses revenue-generating businesses like Highway Pilot to fund robotaxis research, and sensor data collected in real-life scenarios to power models in the lab. The lab’s results, in turn, could improve what is implemented on public roads.

Human or machine

During the 40 minute drive in midday traffic, our car was able to change lanes, merge into traffic, create distance from reckless drivers on its own, except for a brief moment. Towards the end of the trip, our driver decided to grab the steering wheel to change lanes as we approached a car parked dangerously in the middle of the exit ramp. Momenta calls this an “interactive lane change,” which it claims is designed to be part of its automated system and, by its strict definition, is not a human “intervention.”

“The interaction between humans and cars will continue to dominate for a long time, perhaps another 20 years,” Cao noted, adding that the setup takes safety to the next level because the car knows exactly what the driver is doing through the cameras. the interior cabin.

“For example, if the driver is looking down at his cell phone, the [Momenta] The system will alert them to pay attention, ”he said.

I was not allowed to film during the trip, so here are some images from Momenta to take a look at their highway solution.

Human beings are already further on the autonomous spectrum than many of us think. Cao, like many other AI scientists, believes that robots will eventually take the wheel. Waymo property of the alphabet has been running robotaxis in Arizona for several months now, and startups like also offer a similar service in Texas.

Despite all the hype and the boom in the industry, thorny doubts remain about passenger safety, the regulatory scheme, and a host of other issues for fast-moving technology. Uber’s fatal car accident last year delayed the company’s future projects and sparked a public reaction. As a Shanghai-based venture capitalist recently suggested to me, “I don’t think humanity is ready to drive itself.”

The biggest problem in the industry, he argued, is not related to technology, but social. “Self-management poses challenges for the legal system, culture, ethics and justice of society.”

Cao is very aware of containment. He acknowledged that as a company with the power to drive future cars, Momenta has to “take a great responsibility for safety.” As such, it required all company executives to travel a certain number of autonomous miles, so if there is a loophole. In the system, managers are likely to stumble upon it before customers do.

“With this policy in place, the administration will pay great attention to the security of the system,” Cao said.


New Momenta headquarters in Suzhou, China / Photo: Momenta

In terms of actually designing the software to be reliable and tracking liability, Momenta appoints a “systems research and development architect,” essentially tasked with analyzing the black box of autonomous driving algorithms. A deep learning model has to be “explainable,” said Cao, which is key to figuring out what didn’t work: is it the sensor, the computer, or the navigation app that isn’t working?

Going forward, Cao said the company is in no rush to make a profit as it is still spending a lot on R&D, but said the margins on the software it sells “are high.” The startup is also heavily funded, which Cao Curriculum certainly helped attract, as did his other co-founders Ren Shaoqing and Xia Yan, who were also Microsoft Research alumni. Asia.

As of last October, Momenta had raised at least $ 200 million from big-name investors, including GGV Capital, Sequoia Capital, Hillhouse Capital, Kai-Fu Lee’s Sinovation Ventures, Lei Jun’s Shunwei Capital, the investment arm of the electric vehicle maker. NIO, WeChat operator Tencent, and the Suzhou government, which will host Momenta’s new 4,000-square-meter headquarters right next to the city’s high-speed station.

When a bullet train passes through Suzhou, passengers can see from their windows the recognizable M-shaped building of Momenta, which in the next few years could become a new landmark of the historic eastern Chinese city.