"The bottleneck is data," Uber's chief technology officer said. Uber wants to outfit a slice of its millions of drivers with sensor kits to gather street-level data for companies building self-driving cars. The idea, set out by Praveen Neppalli Naga at TechCrunch's StrictlyVC event in San Francisco, grows out of a project called AV Labs that Uber launched in late January. Today AV Labs runs a small, company-operated fleet with sensors. Uber says it hopes a distributed network of driver cars could feed a central, labelled "AV cloud" that partners can query and test in shadow mode.

From rides to real-world sensors

Uber has long stopped short of building its own self-driving cars. The company now wants to play a different role. It plans to turn participating drivers’ vehicles into moving data collectors. That would feed manufacturers and software teams training autonomous driving systems.

Praveen Neppalli Naga, Uber’s chief technology officer, set out the plan in an interview at TechCrunch’s StrictlyVC event. He said the company hopes to scale beyond the small, sensor-equipped fleet AV Labs currently operates. The long-term aim is to outfit members of the driver network with sensor kits so their daily trips add to a huge pool of street-level data.

Uber already runs AV Labs, which it launched publicly in late January. For now the project uses a dedicated fleet run by Uber itself and kept separate from the platform’s drivers. Naga said that's intentional. Uber wants to learn how the sensor kits work before bringing drivers into the programme. He also noted legal questions. There needs to be clarity, he said, on what sensors mean and on what sharing their data entails in each state.

Scale as competitive advantage

The strategic logic is simple. Uber has access to millions of vehicles on the road every day. If even a fraction of those cars carried sensors, the amount of real-world footage and telemetry would outstrip what most single AV companies can collect on their own. Naga argued that the main limit for autonomous technology today isn't the core software.

The bottleneck is data, he said.

Uber already has relationships with other players in the field. The company counts 25 autonomous vehicle partners, including Wayve, which operates in London. That network gives Uber potential customers for whatever dataset it builds. Partners would be able to query a central library of labelled sensor data. They could also run their trained models in shadow mode during real Uber trips to simulate how an AV would have behaved without putting one on the road.

Shadow mode testing lets a company compare its model’s decisions with what a human driver actually did. That offers a cheaper way to validate systems at scale. Uber says the AV cloud would let partners request targeted data sets. For example, a partner could ask for footage from a particular intersection at a specific time of day to help train on that scenario.

Mastheads

Where regulation and privacy fit in

Uber is clear that technical work is only one part of the challenge. Naga flagged regulatory uncertainty as a constraint. He said the company needs to ensure states have clarity on sensors and on data sharing before widening the programme to independent drivers. That suggests legal and privacy questions will shape the rollout.

Turning drivers’ cars into sensors would touch on several sensitive areas. It would capture street scenes, passenger movements and other personal data. Uber will therefore need to build ways to label, anonymise and control access to that information if it wants partners to use the AV cloud. Naga’s remarks indicate Uber recognises these points and plans to proceed cautiously.

Commercial and strategic consequences

If Uber can scale a sensor network, the company would offer something many AV builders lack: a vast, geographically distributed source of labelled real-world data. That could lower the capital barrier for small AV firms that can’t afford to field many test vehicles. It would also change the economics of model training, making targeted data requests possible instead of broad, expensive collection drives.

Uber has previously moved away from building self-driving hardware and software itself. A co-founder has publicly called that decision a mistake. The new plan reframes Uber’s role in autonomy. It moves the company from trying to own the driving stack to supplying the raw material other companies need to build it. Naga said Uber intends to invest more directly in its partner companies as well.

That shift could keep Uber central to the transport ecosystem even as autonomous vehicles proliferate. If partners use Uber’s labelled datasets and shadow testing, they will become integrated with the platform without Uber having to field its own robotaxis.

Scaling from a small company fleet to a distributed network presents engineering questions. The sensor kits Uber tests must be robust enough for daily use in a range of car models and climates. They must also capture data in formats that partner teams can label and feed into training pipelines. Uber intends to learn those technical and logistical lessons within AV Labs before approaching drivers.

There are also operational questions. Uber will need systems to tag data by location, time and scenario. It will need labelled examples to make the AV cloud useful. And it must put controls in place so partners can query and test models without exposing raw personal data. Naga’s comments point to plans for a library of labelled sensor data, but the details of data handling will determine how broadly partners can access it.

Making driver cars into data sources could reshape the competitive map for autonomy. Large AV companies have built fleets to collect specialised datasets.

A scalable sensor grid could offer alternative access to similar or richer sets of real-world scenarios. For smaller AV outfits, being able to query a third-party dataset would lower the upfront costs of model development.

Uber’s partner list already spans 25 companies. That built-in customer base gives the plan an early route to utility. If partners run models in shadow mode on real trips, they get a continuous feedback loop from actual urban driving environments. Uber gets a steady stream of validation that makes its data library more valuable.

Uber hasn't laid out a timetable for when drivers might start hosting sensor kits.

Related Articles

Uber hasn't given a timetable for when drivers might begin hosting sensor kits.

This article was created with AI assistance.