“Slow,” “steady” and “deliberate” are rarely used when describing Uber. But months after an Uber vehicle operating in autonomous mode fatally crashed into a pedestrian, the company has had no choice but to move carefully.
After pulling all of its self-driving cars off public roads, laying off all the vehicle operators in Pittsburgh and San Francisco and shutting down its Arizona self-driving operation, Uber is very gradually beginning to prepare to start testing self-driving cars again. That starts with manually driving its fleet of Volvo XC90s in Pittsburgh to map out the streets.
Mapping out city streets is essentially square one for testing self-driving cars — that’s a tough spot to be in as a company with serious ambitions for developing autonomous technology, especially in the face of massive competition. But Uber has to move slowly to regain the trust of not just the public, but also city and local governments.
This gradual start in Pittsburgh also gives Uber time to implement some of the safety features and changes recommended by an internal team of experts that conducted a review of the company’s self-driving development and operation.
Those changes may seem small individually, but together they point to a significant change in how Uber is thinking about self-driving. While much of the inception and some of the way in which the company later operated its self-driving efforts was fueled by competition — specifically with Google parent company Alphabet — the company is now introducing features and technology that prioritize safety.
Specifically, Uber is going back to having two people in each car, and will be monitoring its safety drivers — the people who are designated to take over control of the car as needed — in real time.
It’s not common for autonomous test cars to be driven with just one person behind the wheel so early in the technology’s development. But around November 2017, Uber switched two safety drivers to one for many of its cars, in part to increase the number of miles each car was driven in autonomous mode, sources told Recode. Miles driven autonomously is just one of many barometers for how advanced a company’s technology is and is how the software driving the cars learns and gains more experience.
An Uber spokesperson said that the company decided to make the transition from two drivers to one because the company felt that the role of the second operator — which was to gather information about how well the car was driving itself — could be done after the fact and didn’t have to be done in real time.
A byproduct of that, however, is drivers were left alone to determine when to take over control over the car as needed, and there was no one there to ensure they were not distracted.
Add to that, the company’s in-car camera system did not detect when a driver was distracted — it simply recorded the driver. That means in order to see whether a driver was distracted at any point in time, a manager had to manually go through the footage after the fact.
While sources say that was inefficient in and of itself, it was also exacerbated by how quickly Uber was hiring safety drivers. In Arizona, some managers were in charge of about 25 people, one source said, making it difficult to spot-check whether anyone had unsafe driving behaviors. That said, if a driver was caught doing something unsafe, the company typically had a one-strike policy, sources said.
This new off-the-shelf camera system Uber is integrating will instead detect distracted driving in real time, log it into the system automatically, and then make a noise to alert the driver to pay attention. (Similar to a system that startup Nauto has built.) So managers will simply have to look at the software or platform to see all the instances of a driver being distracted.
While an Uber spokesperson said the company was being more deliberate about the ratio of managers to vehicle operators — the company is filling about 55 roles for what it calls “mission specialists” — this new real-time alert system will help alleviate some of the difficulty of auditing driving behavior manually.
That’s hugely important. A closer investigation of the fatal crash that took the life of Elaine Herzberg in Arizona revealed that Uber relied heavily on the humans sitting behind the wheel to react immediately in an emergency. And in this case the driver was distracted.
That’s the risk with depending on a human driver — they become too reliant on the automated technology to drive itself safely and, in turn, get distracted. To that end, Uber is also ensuring that the automatic emergency braking system that is already built into the car will be engaged no matter when it is being driven.
From a pure miles-driven standpoint, Alphabet’s Waymo announced it had autonomously driven more than eight million miles on public roads just last week. At last count, Uber had only reached three million miles before it halted all of its self-driving testing in March.
While the company doesn’t have to start entirely from scratch once it does start testing its autonomous technology again, Uber was struggling to meet internal goals in the lead up to the crash.
As we first reported, Uber’s fleet of vehicles could only drive less than a mile before a safety driver had to take over in March 2017. As of March 2018, the company was still struggling to meet its goal of driving 13 miles without a driver taking over. While it was a meager increase from the year prior, it was still far lower than its competitors like Waymo, which said it drove an average of 5,600 miles before a driver had to take over control in 2017.
At the core of this issue is safety. Companies building self-driving cars pitch it as a potentially life-saving technology, but those companies still have a lot to prove. That’s especially true as the U.S. Congress actively works to regulate autonomous technology.
This article originally appeared on Recode.net.