China’s urban self-driving shift: from rules to one-stage AI
China’s urban self-driving systems are trading thick rulebooks for leaner brains. After conquering highways, Navigate on Autopilot (NOA) is now tackling city streets where e-bikes weave, scooters run reds, and drivers play urban chicken. The first wave of urban NOA relied on thousands of hand-coded if-then rules—an “exam-cram” approach that excelled on pre-mapped routes but froze on the long tail of unscripted behavior. Penetration stalled at 15.1% in 2025 and stayed concentrated in premium models above RMB 200,000–300,000 because the system couldn’t reliably handle the chaos.
The bottleneck in two-stage systems
To escape the rule trap, developers turned to end-to-end learning: let a deep neural network ingest raw sensor data and spit out driving commands. Early production systems split the job into two stages—perception then planning—so engineers could debug each block separately. But the boundary between modules discarded subtle cues, creating a bottleneck that limited real-world robustness.
One-stage end-to-end: cutting the middleman
The next leap is a single-stage architecture that feeds raw camera, radar and lidar inputs straight into one neural net, which outputs trajectories or controls without intermediate abstractions. By removing the perception-planning divide, the model learns directly from raw sensor noise and learns to trust its own imperfect understanding of the scene. Early deployments show smoother handling of erratic road users and fewer driver interventions on congested routes.
Why it matters
A one-stage, end-to-end stack matters because it can scale affordably. Without brittle rule sets or information-leaking interfaces, the technology can eventually ship on lower-priced cars, expanding urban NOA beyond luxury vehicles. The real stakes are whether China’s chaotic streets can be tamed not by more rules, but by AI that learns the rules by itself—and stays calm when some road users refuse to follow any.
Source: DEV Community. AI-assisted editorial synthesis — TechnoExpress.