For many of those engineers that left, there was growing concern around Uber prioritizing demonstrating its self-driving car progress in public, rather than more quietly developing and perfecting that technology until it is completely ready for debut.
The shift to show off Uber's progress to the world has come from Levandowski, insiders say. Rather than dedicating the unit's resources to meeting long-term goals, many said he prefers putting on public demonstrations to convince people that Uber is ahead of its competitors.
An Uber spokesperson said that real world tests are critical to the success of the technology and the company has always had plans to launch pilots.
But sources say that laser-focus on public demonstrations led to the rushed launch of Uber's self-driving pilot in San Francisco, as well as a haphazard demo of Otto's autonomous trucks performing a beer delivery in Colorado. Both are examples of the problems plaguing the unit.
After the September launch of semi-autonomous vehicles in Pittsburgh — which was already part of the original plan — Levandowski rushed another demonstration a month later.
One of Otto's trucks drove itself 120 miles down a public highway in Colorado to deliver 45,000 beers as part of a deal with Anheuser-Busch. This took place only eight months after Otto was founded, so it came as a surprise to many that the company was able to show off its technology so early in its existence.
According to an Uber spokesperson the Otto team worked closely with Colorado officials and ran extensive tests with the trucks before meeting an internally set safety criteria and launching the demo. Indeed, as BuzzFeed News reported, the company completed five tests without human intervention.
However, the team was able to put on a show because they cobbled together unfinished systems that were "thrown together," according to a source.
To be fair, it's not exactly uncommon for an early-stage startup to demo technology when it's not yet fully formed. When investors looked at virtual reality headset Oculus, for instance, it was still duct-taped together. Even Cruise — a self-driving startup acquired by General Motors — demoed unfinished technology to potential investors.
But the stakes are higher when it comes to self-driving vehicles, which have the potential to cause fatal accidents. Even though Otto's truck was driving itself on a public highway alongside other human drivers, the computers were crashing every few hours, these people said.
"A huge amount of luck went into that demo," one said.
Fortunately, the truck was accompanied by a motorcade of state troopers during its demo. But for some of Uber's top engineers, luck was simply not good enough.
Sources say that after the Otto demo, some ATG staffers became concerned about the ethics of fielding unsafe systems and rolling them out onto public roads. Shortly after the October demo, Rander, Browning and Bagnell left the company.
Those departures came at a critical time for the company.
Just two weeks after the engineers left, Uber began its second commercial test of its autonomous cars — this time in San Francisco.
Even to the public, it was clear the rollout of these Volvos was a logistical disaster. On December 14, or the first day of Uber's commercial test, one of the cars ran a red light. While Uber maintains it was due to human error, a damning New York Times reportrevealed the internal logs showed the system failed to recognize the traffic light, a significant technological issue to say the least.
But with nascent semi-autonomous technology — which means it's not supposed to completely drive itself — safety drivers have to be at the ready to take over at a moment's notice. A skilled and trained safety driver would've prevented this, several people contend.
But, "we don't have those [types of] drivers in San Francisco," one source familiar with the matter said.
That wasn't the only hole in the company's plans to roll out its cars in its hometown. The scene ahead of the launch was just as messy, sources say.
As early as September, Levandowski began emailing with the California Department of Motor Vehicles, arguing that the company didn't need to obtain the $150 permit required of all autonomous vehicle companies to operate semi-autonomous cars on public roads. The DMV argued that it did, but Levandowski informed the department Uber would continue with its plans anyway.
By late October, it became clear the operations team in San Francisco was struggling to prepare for the public launch of the company's on-demand autonomous cars, according to insiders.
It was then that a number of software engineers and safety drivers from the Pittsburgh arm of the ATG were sent to California to help the San Francisco team, which had problems performing certain critical tasks, according to sources.
It's not unusual for cross-functional teams to help out on getting pilots up and running, an Uber spokesperson contended. This was also the case when the company demonstrated its technology in Arizona recently.
For the San Francisco pilot, the engineers were sent to help with tasks such as determining what data was missing — like where traffic lights are located — and what the vehicle was not seeing, as well as things like how to onboard new code.
Part of the reason the California launch wasn't as seamless as the Pittsburgh launch was because Uber had originally charted out parts of Arizona with plans to begin its second self-driving test in the state by the winter of 2016, according to several sources.
After the red-light disaster, the California DMV put its foot down and revoked the company's licenses, forcing Uber to ship its vehicles to Arizona, the site of Uber's original testing plan. Levandowski had insisted on moving it to San Francisco since Arizona wasn't flashy enough, sources said.
Uber, no stranger to local regulatory battles, has since applied for and received the proper permits to operate in California. In the coming weeks, the company expects to roll out a number of its self-driving Volvos in San Francisco, according to internal documents Recode obtained.