China's intelligent driving is moving fast from highway Navigate on Autopilot (NOA) into the far harder world of urban NOA. The first leap moved hands-free driving out of the closed expressway and into real city streets. The second leap, the one now underway, is rewriting how the car actually thinks.
China's urban pilot-assist evolution at a glance. (Image: EV & Auto Club)
1. The Rules Era: An "Exam-Cramming" Trap for City NOA
Highway NOA was relatively simple to crack. The road is closed, the geometry is consistent, the actors are mostly cars, and a mature rule-based stack can deliver a comfortable product. Urban NOA is a different beast. The system has to handle traffic lights, unprotected turns, pedestrians, e-bikes, food-delivery scooters running red lights, and a hundred flavors of "I-don't-care-about-the-rules" intersection behavior. The complexity grows exponentially.
The earliest urban NOA architectures followed one mantra: cover every possible scenario with hand-written rules. Engineers enumerated traffic situations and wrote thousands of if-then-else statements: when to start moving after a light turns green, how much to slow when cut off, how to plan a trajectory for an unprotected left turn. On the highway this approach can pass a test. In the city it falls apart for a single structural reason.
China's urban road users, almost by definition, do not follow the rules. Electric scooters drive the wrong way. Pedestrians cross mid-block. Food-delivery riders weave between cars. Drivers in congested intersections play chicken in the kind of "zipper merge" etiquette nobody teaches. These are the long-tail scenarios that no rule library can fully enumerate. As one early test team admitted about their own city NOA: "It feels like exam cramming β it scores beautifully on the routes we pre-mapped, and the moment it hits an unrecorded scenario, it hesitates, behaves awkwardly, and then asks the driver to take over."
That "εη§" (one-trick) experience is precisely why urban NOA penetration in China only reached about 15.1% in 2025, and remained concentrated in vehicles priced above RMB 200,000β300,000. Until the system can handle the long tail, it cannot move down-market.
2. Two-Stage End-to-End: The Clumsy Half-Step
To break out of hand-written rules, the industry pivoted to end-to-end (E2E) learning: let a deep neural network replace the rule stack, take sensor input, and directly output driving commands. Train it on massive amounts of human driving data, and the system should learn to drive like a human.
The first wave of E2E that hit production, however, was almost always two-stage. The pipeline was split in two: the first stage turned raw sensor data into an intermediate representation (a semantic map, a set of drivable regions, detected objects). The second stage took that intermediate representation and produced the trajectory or control command.
Two-stage E2E shipped faster, and it shipped more explainably. Because the perception and planning modules output clean intermediate products, engineers could monitor and debug each one independently. The fatal flaw was information loss. The intermediate representation chosen by the perception module is not guaranteed to contain every detail the planning module needs; the boundary between the two modules is, by construction, a bottleneck that throws away signal.
In one sentence: two-stage E2E is still "perception module first, planning module second." It can be trained jointly, but the boundary is hard. The user experience improves over pure rules, but the car still drives stiffly β hard braking, hesitant lane changes, jerky longitudinal control. It solved the "rigid" problem without solving the "clumsy" one.
3. One-Stage End-to-End: The Real "Muscle-Memory Driving"
A Velodyne LiDAR unit β one of the sensor types feeding modern driving stacks. (Image: Wikimedia Commons, CC BY 2.0, Steve Jurvetson)
Over the past year, one-stage end-to-end has ignited an arms race and rapidly become the technical consensus among the leading players. Its core idea is brutally simple: one model, all the way. A single neural network takes sensor input at one end and outputs a driving trajectory at the other, with no manually defined intermediate module, and the entire stack is optimized end-to-end against a safety objective.
The most visible feature of one-stage E2E is how human it drives.
A one-stage model has, in effect, developed "muscle memory" for driving. It does not need an engineer to write "when cut off, slow by X m/sΒ²". Through training on massive human driving data, the model internalizes a vehicle sense that mirrors a human driver's intuition: it follows smoothly, anticipates the lead car's movement, lifts off the throttle early, and coordinates steering and throttle with a fine touch. There is no jerky wheel input, no repeated micro-corrections, no rushed stops. A reviewer who drove a Chery Exeed ET running the Horizon HSD one-stage E2D stack summarized it: "Steering and throttle are coordinated with a very delicate touch. There is no abrupt steering or back-and-forth correction." That is the look of an end-to-end model that has truly learned to drive.
Today, one-stage E2E has received explicit backing from Tesla, Huawei, Horizon, Momenta, Bosch, QCraft (θ½»θζΊθͺ), and MOVING ACT (ιθ§ζΊθ½), with each investing heavily and shipping the architecture into real vehicles. The technology has gone from research demo to volume production in roughly twelve months.
| Architecture | Core Logic | Typical Experience |
|---|---|---|
| Rules era | Hand-written if-then-else rules |
"Exam cramming" β fine on pre-mapped routes, awkward on the long tail |
| Two-stage E2E | Perception + planning split, joint training | Ships fast, but information loss keeps the car stiff |
| One-stage E2E | Single model, sensor-in to trajectory-out | "Intuition driving" β smooth, human-like, robust on the long tail |
4. The New Race: Safety, Explainability, and Scale
One-stage E2E has a high ceiling, but it also brings a new problem: the black-box trap. Because the decision process is opaque, when the car does something odd it is hard to tell whether the perception failed or the planning failed. Root-causing field issues becomes genuinely difficult.
Regulators are already moving. On August 4, 2026, the Ministry of Industry and Information Technology (MIIT) formally published the country's first mandatory national standard for L3 and L4 autonomous driving β GB 44721-2026, "Safety Requirements for Automated Driving Systems of Intelligent and Connected Vehicles." It is scheduled to take effect on July 1, 2027. The standard requires vehicles to carry a tamper-proof Automated Driving Data Recording System (the "vehicle black box"), mandates full life-cycle safety responsibility on the OEM, requires hardware redundancy on steering, braking, and the intelligent driving controller, and bans misleading marketing terms such as "full autonomous," "L2.999," or "quasi-L3." As one regulator put it: "OEMs are now the safety principal for the entire life cycle β from design, through production, to OTA updates and recall."
Earlier, on September 17, 2025, MIIT had already opened the public consultation for another mandatory national standard β the Safety Requirements for Combined Driver Assistance Systems β which targets the L2 layer (today's adaptive cruise, lane centering, NOA) and sets a uniform safety baseline for hands-off and eyes-off detection, lane-change behavior, and the data-recording chain that feeds incident analysis.
With regulation tightening, the industry's competition logic is shifting from "whose ceiling is higher" to "who can make end-to-end both smart and trustworthy." Leading players are now investing in safety-explainable one-stage E2E: weaving safety mechanisms into the model architecture itself and providing auditable intermediate monitoring points that engineers and regulators can both inspect.
Scale is moving into the deep water. In 2026, the number of production models carrying one-stage E2E is expected to reach several dozen. In November 2025, Bosch pushed the industry's first volume-production one-stage end-to-end advanced driving solution to the Chery Exeed Sterra ES. In the first half of 2026, ZYT (formerly DJI Automotive) began rolling its "ι«ζζ§" end-to-end 4.0 model to existing FAW Hongqi owners via OTA, starting with the Hongqi HS6 PHEV and Tiangong 05.
Closing
The arc of China's urban pilot assist is a clean one. From the "exam cramming" of hand-written rules, through the clumsy two-stage end-to-end transition, to today's volume-shipped one-stage E2E, it has been a step-by-step escape from "how do we get the car to learn to drive" toward "how do we get the car to drive like a person."
The next station on that escape is harder than the last. It is not enough for the system to be smooth; it must be transparent, verifiable, and cheap enough to leave the RMB 200,000+ club. The real "democratization" of city NOA will arrive the day a one-stage model with a black box under the seat can be safely sold, maintained, and explained at half the price of today's flagship. That day, not the day of any single demo, is the one that will define the next decade of Chinese intelligent driving.
