[Corrected 5/21/19, 11:04 am. See below.] Sebastian Thrun first made his mark on autonomous vehicle development at the dawn of that industry, when he led a Stanford team whose robot car Stanley won the $2 million DARPA Grand Challenge in 2005 by racing driverless through the Mojave Desert for 132 miles.
These days, Thrun seems just as excited by another imprint he’s made on the still-forming AV industry—not as a researcher, but as an entrepreneur in online education. Thrun (pictured) says many software engineers working at the raft of companies now developing self-driving vehicles have trained in the technology via Udacity, the educational technology company he founded in 2012.
Mountain View, CA-based Udacity, which offers online courses in basic programming and higher tech-related skills, launched an eight-course Nanodegree program focused on self-driving technology in 2016, in a partnership with Mercedes-Benz Research & Development North America, the Silicon Valley R&D center of Mercedes-Benz. This week, Thrun and Udacity are announcing a second AV course sequence, the Sensor Fusion Engineer Nanodegree program, in an expansion of the Mercedez-Benz partnership.
Since those partners first offered the original Self-Driving Car Engineer Nanodegree program, more than 21,000 students have signed up from more than 120 countries. Mercedes-Benz has hired more than 40 of the students who have earned that Nanodegree, Udacity says, and others work at companies including Audi, BMW, Bosch, Jaguar Land Rover, Lyft, and chipmaker Nvidia. [An earlier version of this story gave an inaccurate figure for student enrollments in the Self-Driving Car Engineer Nanodegree program. We regret the error.]
“I know of no other car company without one of our grads,” Thrun told Xconomy in an interview.
By broadening the ranks of AV-trained engineers, Thrun says, Udacity is a catalyst for a self-driving car industry that now has deployed cars on public roads in a growing array of pilot projects.
Udacity’s new four-course Sensor Fusion program covers the art of blending together data from an autonomous vehicle’s cameras, LIDAR, and other sensors—with the help of machine learning software—to allow the car to interpret the environment it’s navigating through. For example, an AV has to tell the difference between a highway center line and a raised roadside curb.
Thrun says 90 percent of what a self-driving car must master is “perception.’’
“Driving itself is simple compared with understanding the world,’’ Thrun says.
Udacity’s turning point
The original Self-Driving Car Engineer Nanodegree program was one of Udacity’s most popular offerings, boosting revenue and the company’s “branding’’ as a source of learning opportunities closely aligned with skills needed in the booming tech economy, Thrun says.
The boost was well-timed for a period when Udacity was off-balance—growing fast, but also growing staff and expenses. The company, backed by investors including Andreessen Horowitz, Drive Capital, and GV, the venture capital arm of Google parent company Alphabet, was valued at $1 billion in 2015, at the time of its last funding round. Its revenue grew 25 percent to approximately $90 million in 2018. But Udacity was not heading apace toward profitability when, late last year, the company began a restructuring and a series of substantial layoffs, reducing the current workforce to about 300 full-time staffers worldwide.
CEO Vishal Makhijani left Udacity in October, and Thrun, as president and executive chairman, resumed a more active role in the company’s management. He had served as Udacity’s CEO until Makhijani assumed that role in April 2016, when Thrun began focusing on his work as chief executive of the flying-car startup Kitty Hawk. Udacity hired Lalit Singh as COO this month, and the search for a new CEO is continuing.
Thrun sees the new Sensor Fusion Nanodegree program as part of Udacity’s planned resurgence—not only because of its content, but also because it’s beginning under a new regime of support services for students.
A big part of Udacity’s revival plan is to tackle an ongoing challenge for online education: raising course completion rates. Back in 2012, the company was among the pioneers offering MOOCs, or Massively Open Online Courses. Its early, free courses, open to any student anywhere, offered global access to training in computer science and programming languages. But completion rates ranged in the low single digits. The question was, how much support from coaches and other staffers was needed to help students along?
Udacity experimented by forming partnerships with universities, and collaborating with companies such as AT&T and Nvidia on course design. It developed an enterprise services business line, providing companies with tailored skills training for their employees. In 2014, Udacity began supplementing its open menu of single courses with more than 30 Nanodegrees on topics including data analytics, machine learning, robotics, and autonomous flight engineering.
Currently, the average completion rate across all of Udacity’s Nanodegree programs is 34 percent, the company says. Thrun is aiming to drive that figure higher.
“We want to stand for success,’’ Thrun says. As part of that effort, Udacity fielded a group that Thrun calls “Team Ninja’’ to ask hundreds of students why they’d bailed on a company course or program. The answer for 40 percent of the dropouts: “I got stuck,’’ Thrun says.
Udacity adopted a Team Ninja recommendation: Match all students with a “technical mentor” who can explain things they can’t figure out for themselves. As of May 1, all students in any Nanodegree program have been assigned such mentors, and also two other kinds of support staffers: first, expert reviewers who check over student work on projects—even to the extent of giving line-by-line feedback on a student’s computer code; and second, career counselors who help students with things like brushing up their LinkedIn profiles and getting ready for job interviews, Thrun says.
“My ambition is to make Udacity how a life-long university would look like, 50 years from now,’’ Thrun says.
Enrollment in the Sensor Fusion Nanodegree program begins May 21, and there’s no limit on the number of students. The program is designed to take students about four months to complete; the fee is $399 per month. Udacity allows enrollees to take longer—though they’ll pay for each additional month.
AVs’ still-undefined future
Udacity is expanding its educational options in autonomous navigation at a time when the future commercial use of self-driving vehicles is still hazy. Will these vehicles take over long-haul trucking, become part of public transit systems, and relieve unprofitable ride-hailing companies like Uber and Lyft of the expense of paying drivers?
Should developers of autonomous vehicles design the technology not only for high performance, but also for low cost? If they don’t keep the price within bounds, might fleets of self-driving cars make a profit only as luxury car services, but not as providers of affordable everyday rides for masses of people? Could some real-world applications be closed off because the tab is too high?
“That’s possibly the most important question,’’ Thrun says.
Some AV innovators are concerned enough about cost that it dictates their design decisions, Thrun says. For example, Tesla developed its driver assistance system Autopilot without using expensive LIDAR technology, relying instead on less-costly sensors such as cameras.
Other companies simply push toward making AV systems that work, without worrying too much about costs today, Thrun says. It’s hard to predict the outcomes of either approach. “We’re in a very early stage of research,’’ Thrun says.
The cost problem may solve itself—at least in part—due to factors such as economies of scale, he says. Some elements of autonomous navigation systems, such as radar, are already dropping in price because they’ve found mass markets for current-day uses, such as vehicle cruise control. A radar installation that would have cost $50,000 about 20 years ago now sells for something like $60 to $80, Thrun says.
In addition, the cost of each self-driving car is likely to dip lower once the cars are manufactured en masse, such as in production runs of 100,000, he says.
Still, he says, costs have to come down. “I’m sure it’s on people’s minds,” Thrun says.
Photo of Sebastian Thrun in 2014, courtesy of Udacity