China's urban pilot-assist stacks give up on enumerating the intersection

A rule library is written by a fixed number of engineers. The situations it has to cover are produced by everybody else on the road, and in the Chinese city those authors are not working from the same document: scooters going the wrong way, pedestrians crossing mid-block, delivery riders weaving between cars, congested-intersection chicken that nobody is taught ^[5]. Each if-then-else statement buys coverage of one situation ^[4]. The population generating situations is not bounded by the rule set, so a bigger team writes more rules and still does not close the set. That is the difference between a hard problem and a structurally wrong method.

The failure mode described by one early test team is the diagnostic: the system scores beautifully on pre-mapped routes, then hesitates and asks the driver to take over the moment it meets something unrecorded ^[6]. A takeover request is not a slightly worse drive. It is the feature switching itself off in exactly the conditions the buyer was sold on.

Which is why the penetration figure carries more weight than the architecture argument. Urban NOA reached about 15.1 percent in China in 2025 and stayed concentrated in cars above RMB 200,000 to 300,000 ^[7]. The other 84.9 percent of the market did not have it ^[16], and the source's own reading is that the one-trick behaviour is the reason, with no route down-market until the long tail is handled ^[8]. Rule enumeration priced itself into a segment where the driver tolerates the handback.

The first production answer, two-stage end-to-end, moved the defect rather than removing it. Stage one compresses sensor data into a semantic map, drivable regions and detected objects; stage two turns that into a trajectory ^[10]. The module choosing what to keep does not know what the planner will need, and the boundary discards signal by construction ^[12]. The result was a car that stopped being rigid and started being clumsy: hard braking, hesitant lane changes, jerky longitudinal control ^[13].

One-stage removes the boundary, and with it the thing that made two-stage shippable. The clean intermediate products were what let engineers monitor and debug perception and planning independently ^[11]. A single network from sensor input to trajectory, with no manually defined intermediate module ^[14], has no such outputs to inspect. The validation burden moves off the unit test and onto fleet behaviour.

On what one-stage buys, the source is qualitative. It describes smooth following, early throttle lift and coordinated steering, the internalised vehicle sense of a human driver ^[15]. There is no takeover rate, no intervention distance, no comparison on the same routes ^[17]. So the consensus among leading players over the past year ^[14] currently rests on how the car feels, argued against a rules baseline whose weakness was measured in the same currency. That is a reasonable trade to make and an unreasonable one to verify, and the vendors who go first will be the ones explaining a bad intervention without a module to point at.