Autonomous Driving Data Flywheel

Autonomous driving data flywheel is Cao Xudong’s strategic trunk for Momenta in Momenta IPO后再访曹旭东：就是想做没有尽头的AI. The claim is that autonomous driving cannot be solved only by clever module design: the algorithm architecture has to be data-driven, and the company has to obtain huge physical-world driving data through mass-production deployment.

In Cao’s telling, this produced Momenta’s “two legs” strategy. One leg is mass-produced assisted driving, which creates customer pressure, revenue, and real-world data. The other leg is more fully driverless Robot applications such as Robo One, Robotruck, and Robotaxi. The flywheel therefore links product delivery, data collection, training, safety improvement, and commercial expansion rather than treating them as separate roadmaps.

The source also makes the concept a bridge to robotics. Cao argues that World Models and physical-world data from driving can support a shared model base for different robot businesses, although the source does not prove that home robots already have the same data scale as cars.

Key Claims

• Autonomous-driving progress depends on data scale, data quality, and architecture that can keep absorbing fleet experience.
• Mass production is not only revenue; it is the mechanism that makes the data loop real.
• Supplier concentration can follow from the flywheel because early scale, data, customer programs, and R&D spending reinforce one another.
• The flywheel complements Physical World Data Flywheel but is narrower: cars can create repeated, sensor-rich, regulated physical-world data before general robots have comparable household deployment.
• Robotaxi, Robotruck, and delivery robots are downstream business routes when the same perception, prediction, planning, validation, and safety infrastructure can be reused.

Connections

• Momenta and Cao Xudong - source company and speaker.
• Huawei and Tesla - competitive and benchmark cases in the source’s autonomous-driving discussion.
• Physical World Data Flywheel - broader embodied-AI data-loop concept.
• World Models and Physical AI - model and field context for transferring physical-world learning beyond cars.
• Robotaxi Economics and Autonomous Vehicle Safety Benchmark - commercialization and safety-evaluation context.