Uber has more than 20 self-driving car partners, but they all want one thing: data. So the company says it will make it available through a new division called Uber AV Labs.
Despite its name, Uber has no intention of returning to developing its own robotaxis, halting development in 2018 after one of its test vehicles killed a pedestrian. (Uber ultimately sold the division in a complicated deal with Aurora in 2020). But the company plans to send its cars equipped with sensors into cities to collect data for partners like Waymo, Waabi and Lucid Motors, but no deals have been signed yet.
Broadly speaking, self-driving cars are in the midst of a transition from rule-based driving to greater reliance on reinforcement learning. Real-world driving data then becomes extremely valuable in training these systems.
Uber told TechCrunch that the self-driving car companies that need this data the most are those that already collect large amounts of data themselves. This is a sign that, like many frontier AI labs, we recognize that “solving” the most extreme edge cases is a game of volume.
physical limits
Currently, the size of self-driving car companies’ fleets imposes physical limits on the amount of data they can collect. And while many of these companies create simulations of real-world environments for edge cases, there’s nothing like driving on real roads, and driving them over and over again, to discover all the strange, difficult, and completely unexpected scenarios that cars end up in.
Waymo provides an example of this gap. The company has been operating or testing self-driving cars for a decade, but its current robotaxis were recently caught illegally passing parked school buses.
Access to a larger pool of driving data could help robotaxi companies solve some of these problems before or when they become serious, Praveen Nepali Naga, Uber’s chief technology officer, told TechCrunch in an exclusive interview.
tech crunch event
san francisco
|
October 13-15, 2026
And Uber doesn’t charge you. At least not yet.
“Our main goal is to democratize this data, right? So the value of this data and the advancement of AV technology for our partners is far greater than the money we can make from it,” he said.
Danny Guo, Uber’s vice president of engineering, said the lab first needs to build a basic data foundation before determining a product’s market suitability. “Because if we don’t do this, I can’t believe anyone else can,” Guo said. “Therefore, we believe we must now take on this responsibility as those who have the potential to unlock the potential of the entire industry and accelerate the entire ecosystem.”
screw and sensor
The new AV Labs division started small. So far, the company has just one car (a Hyundai Ioniq 5, but Uber says it’s not partnering with a single model), and Guo told TechCrunch that his team is still literally screwing in sensors like lidar, radar and cameras.
“I don’t know if the sensor kit will fall off or not, but that’s the crap we have,” he said with a laugh. “I think it’s going to take a while to say we’re going to put 100 cars on the road and start collecting data. But prototypes are already out there.”
Partners do not receive raw data. Once the Uber AV Labs fleet is up and running, the department will “need to align and work with our partners on data,” Naga said. This “semantic understanding” layer is what driving software from companies like Waymo uses to improve real-time route planning for robotaxis.
Still, Guo said it’s likely Uber will take an intrusive step, essentially putting its partner’s driving software into AV Labs’ vehicles and making them run in “shadow mode.” If an Uber AV Labs driver behaves differently than the self-driving car software’s shadow mode behavior, Uber will notify our partners.
This will not only help find flaws in the driving software, but also help train the model to drive more like a human rather than a robot, Guo said.
Tesla’s approach
If this approach sounds familiar, that’s because it’s essentially what Tesla has been doing for the past decade to train its own self-driving car software. However, Tesla has millions of customer cars on roads around the world every day, and Uber’s approach lacks the same scale.
That doesn’t matter to Uber. Guo said he expects more targeted data collection based on the needs of self-driving car companies.
“There are 600 cities to choose from. [from]. If a partner tells us a specific city of interest, we simply [cars]” he said.
Naga said the company plans to expand the new division to hundreds of employees within a year, and Uber wants to move quickly. And while he sees a future where he can leverage Uber’s entire fleet to collect even more training data, he knows a new division has to start somewhere.
“Based on our conversations with partners, they’re just saying, ‘Please give us anything that helps,’ because the amount of data that Uber can collect exceeds all of the data that Uber can collect on its own,” Guo said.
