- Argo AI is a startup that appeared seemingly out of nowhere six months ago, with $1 billion in backing from Ford.
- Argo is developing self-driving technology that Ford can use to deploy vehicles for commercial on-demand service. In other words: something like a self-driving taxi service.
- There are the technical challenges that must be solved by 2021 if Ford is to meet its own publicly declared deadline.
Somewhere between the 14th and 15th floors in a concrete stairwell, Bryan Salesky pauses, searching for the right words to explain his mission for the foreseeable future. He wants to give cars the eyes, ears, and brains they need to operate without humans. And he wants to do it for Ford Motor Company by 2021.
The CEO of Argo AI — a startup that appeared seemingly out of nowhere six months ago, with $1 billion in backing from Ford — is hardly alone in the pursuit to transform the automobile into a vehicle controlled by artificial intelligence. Though a fire alarm interrupted an interview in a San Francisco conference room, Salesky stays focused and collected. And if he is feeling the pressure to develop and deliver this system so Ford — its sole customer, backer, and majority shareholder — can deploy fully autonomous vehicles in just four years' time, it doesn't show.
Instead, he comes off as optimistic about the company he founded with Peter Rander, who, as former engineering lead at the Uber Advanced Technologies Group, helped bring the ride-hail company's first-generation self-driving prototypes to public roads.
Yet, Argo stands out from the hundreds of companies pursuing self-driving technology due to its unique deal with Ford that has invested big in this little-known startup that is now primed to compete with Google's Waymo, Uber, GM's Cruise Automation, Tesla, and Aurora — a short list of heavyweights all working on a so-called "full stack solution" of self-driving cars.
Argo was literally yanked out of obscurity by the Ford deal, but there are dozens (if not more) AI / auto startups still wallowing in obscurity. "Artificial intelligence will be an essential player in autonomous vehicles of the future," said Michelle Krebs, executive analyst with Cox Automotive's Autotrader. "That's why automakers like General Motors, Toyota, and Ford are snapping up companies with AI expertise."
Argo won the lottery, essentially. (Its website is still laughably sparse.) What remains to be seen is if Ford made a wise roll of the dice. Much of exactly what Argo is doing remains unknown.
In broad terms, Argo is developing self-driving technology that Ford can use to deploy fully autonomous Level 4-capable vehicles for commercial on-demand service. In other words: something like a self-driving taxi service. Level 4 is a designation by SAE International that means the car takes over all of the driving in certain conditions.
More from The Verge:
ScreenDrive: Tesla Model S is the epitome of a tablet on wheels
Fiat Chrysler joins BMW in Intel's self-driving super group
Google spinoff Waymo patents softer self-driving cars to protect pedestrians
Argo is tasked with developing the entire "virtual driver system," which means all of the sensors like cameras, radar, light detection, and ranging radar known as LIDAR, as well as the software and compute platform. Ford has also charged Argo with how to create high-definition maps, keep them "fresh," and sustain that over time, Ford's CTO and vice president of research and advanced engineering Ken Washington said during a presentation at The Information's autonomous vehicle summit in June.
The end result is a full-stack system designed just for Ford's self-driving vehicles so they know where they are in the world, can detect and understand objects in their environment, and then make the right decisions.
The road ahead for Argo AI — and Ford, for the matter — is dotted with obstacles. Argo is now battling with the competition to recruit roboticists and machine learning experts. It's a fight that is pushing salaries, including bonuses and equity offerings, for self-driving car engineers well beyond $250,000 a year.
And then there are the technical challenges, which must be solved by 2021 if Ford is to meet its own publicly declared deadline.
"I think that Ford had a start," Salesky said, noting the company's decade of autonomous research and development. "There was nothing that Ford was doing that was inherently wrong or busted. They were pursuing one method and we were like, 'Hey there are a bunch of methods we can pursue and all combine to solve this problem robustly.'"
The two big technical challenges are with perception, or the ability for the car to see and understand objects around it, and decision-making.
Salesky, along with many others in the field, say perception is the stickier problem, because it's important for the autonomous vehicle to not only detect relevant objects, but to predict what those objects like a car, pedestrian, or bicyclist are going to do. Once it has the correct and robust information, making a decision is relatively easy.
There are two schools of thought on how to solve these problems. Startups like Drive.ai and tech giant Nvidia argue that deep neural networks — a sophisticated form of artificial intelligence algorithms that allow a computer to learn by using a series of connected networks to identify patterns in data — can be applied to everything the self-driving car does from recognizing objects to making decisions.
Proponents of deep learning say these algorithms most closely mimic how the human brain learns. But Salesky contends deep learning requires more research and computing power for it to be used in autonomous vehicles in the near term.
The other approach — and the one Argo is taking — is to train these deep nets to solve very specific problems and then surround that with machine learning algorithms that tie into the broader system.
Machine learning is a form of artificial intelligence that uses algorithms to identify and analyze patterns in data, learn from it, and then make predictions. For instance, machine learning algorithms are used to take data from cameras or LIDAR to teach the vehicle how to recognize a stop sign or moving car.
To the casual observer, it might seem like the race to deploy autonomous vehicles began less than two years ago. And by some measure, it did. High-profile acquisitions — like GM's purchase of Cruise Automation — and a string of public declarations by tech companies and automakers to bring self-driving cars to roadways have dominated the headlines.
But in reality, the road to developing autonomous vehicles began in earnest more than a decade ago.
Scan the directories of every company working on self-driving cars today, and you'll likely discover the names of people who participated in at least one of the three autonomous vehicle challenges in 2004, 2005, and 2007 that were funded by the Defense Advanced Research Projects Agency.
Since those early days of research and experimentation, a group of robotics and machine learning experts have migrated between academia to the world of startups and large corporations — often times moving from co-worker to competitor.
Those early projects forged the people and relationships that are now showing up in the some of the leading self-driving car programs globally. Argo AI is one example.
When Salesky left the Google self-driving project in September, the race to deploy autonomous vehicles had already shifted into overdrive, with automakers, tech companies, and startups jockeying for a leading position — a battle that would have seemed ridiculous just three years before. Salesky was part of that small circle that companies coveted.
When Salesky arrived at the National Robotics Engineering Center, a unit within Carnegie Mellon University's Robotics Institute, he found his true love. He was senior software engineer on the winning team in the 2007 autonomous vehicle challenge funded by DARPA. Chris Urmson, who would later head up Google's self-driving project and recently co-founded self-driving car startup Aurora, was director of technology of that team.
Salesky made the leap over to the Google self-driving car project, where he eventually became director of hardware development. During his time at Google, he headed up discussions with Fiat Chrysler Automobiles that led to a partnership to produce self-driving Chrysler Pacifica minivans.
In short, Salesky held a golden ticket to a well-funded startup or a high-profile position at a major automaker or tech company. He chose option number three.
"I felt like the next step should be something I'd want to commit to for a very long time, and something that would give us the best opportunity possible to get a product out the door," Salesky said. "I thought the best way to do that was to start a company and find an OEM that was like-minded with us and have a deep partnership."
Around the same time, Ford was ramping up its own self-driving car plans: in August announced it would deploy a commercial self-driving car service by 2021. The automaker said it would achieve that target date by expanding its research lab in Palo Alto, California, and investing or buying autonomous vehicle technology startups. All of this is now ultimately the responsibility of Jim Hackett, who in May took over as president and CEO of Ford. Hackett reports directly to Bill Ford, executive chairman, great-grandson of the founder.
Not long after leaving Google, Salesky and Rander started their new company with a small investment from an undisclosed source; neither Salesky or Ford will identify the source of the seed money. Negotiations with Ford, which included some involvement from Bill Ford, began in earnest before the end of 2016. The deal was sealed and announced in February. While the former CEO Mark Fields was privy to negotiations with Argo AI, the primary players were Washington; Raj Nair, the former CTO who now leads Ford North America; and John Casesa, group vice president of global strategy.
Argo was not acquired by Ford, stressed Salesky, the only time he exhibits even a smidge of prickliness in a long and winding interview. Ford is the majority stakeholder and has two seats on a five-seat board. Salesky and Rander have two seats as well. Still, Ford's influence is notable. Since Argo AI's public debut in February, it has amassed more than 100 employees, many of them Ford engineers who were working in the R&D department on a virtual driver system.
"Most of our software engineers, the architects who know how to write well-designed and scalable software, they went to Argo," said Washington.
An examination of LinkedIn profiles shows the largest percentage of Argo AI engineers and data scientists at this point have come from Ford, followed by more than a dozen from Uber Advanced Technologies Group, and a smattering of other tech companies and organizations, including the NREC, Apple, Microsoft, and Google.
The plan is to hit 200 employees by the end of the year. Employees will be spread out across Argo's three locations in Pittsburgh, Pennsylvania; Mountain View, California; and Dearborn, Michigan, where the autonomous vehicles Ford uses for testing collect data. Argo will refresh the virtual driver system in those cars toward the end of the year. Argo has also taken over, and is now evolving, a simulation system that Ford was developing, according to Washington.
As young as Argo is, it's not quite accurate to call these "early days" anymore. Videos of fully autonomous test vehicles navigating streets successfully (with a human behind the wheel, just in case) no longer hold the same cachet as they did a year ago. The long technical and regulatory slog is on. In Salesky's view, it's not a winner-takes-all game. "Three trillion miles are driven each year in the US. That's a huge amount of opportunity," Salesky said. "There's room for more than one player in this space."
Tesla, GM, and Volvo are some of the other car companies actively pursuing autonomous development. But still, other automakers are pooling resources and working with suppliers, skeptical about the risks of investing in autonomy. In contrast, Ford and Argo are holding their cards close to the vest during this R&D phase. It's a partnership that appears to rest on being the player that comes in first.