Uber’s AV Labs program aims to equip human drivers’ cars with sensors to sell real-world training data to AV companies at a scale no single AV firm could match.
Key Takeaways
CTO Praveen Neppalli Naga says the AV data bottleneck is access and capital, not underlying technology – Waymo-style deployment is too expensive for broad coverage.
Current AV Labs uses a small dedicated Uber-operated fleet; outfitting millions of contractor-owned vehicles is the stated long-term goal, pending state-by-state sensor regulations.
Uber is building an “AV cloud”: a labeled sensor data library that 25 partner AV companies can query for specific scenarios (intersection, time of day, geography).
Shadow mode lets partners simulate how their trained models would have performed against real Uber trips without deploying an actual vehicle on-road.
Uber holds equity stakes in multiple AV players and positions the data as “democratized,” but commercial leverage over partners who depend on its ride marketplace is structurally unavoidable.
Hacker News Comment Review
Commenters broadly dispute Naga’s “bottleneck is data” framing – Tesla’s fleet and dashcam/delivery data already exist at scale, making generic sensor data a nice-to-have rather than a differentiator in 2025.
The shadow mode feature drew more technical interest than the sensor grid itself; running AV models against live Uber trips at scale is the near-term product, not hardware rollout.
Driver compensation and consent are the obvious unresolved problems if Uber monetizes data collected through contractor-owned vehicles, and no regulatory or payment framework exists yet.
Notable Comments
@cheriot: frames this as Uber deliberately fostering AV competition – a scaled single AV company is an existential threat to Uber’s marketplace.
@bastawhiz: “This is them trying to sheepishly reenter the race they were dramatically ejected from” – referencing Uber’s 2018 fatal AV crash.
