AI Driving Assist vs Full Self Driving in China

H2: What’s Legally on the Road Today — Not What’s Promised

In Beijing’s Chaoyang District, a Xiaomi SU7 Max navigates the 4th Ring Road during rush hour — lane-keeping active, adaptive cruise engaged, automatic lane changes triggered by driver confirmation. But when the driver takes hands off the wheel for more than two seconds? The system flashes amber, vibrates the steering wheel, and disengages within 5 seconds. That’s not failure — it’s compliance.

China does not permit fully driverless operation on public roads *without remote supervision* or *geofenced operational design domains (ODDs)*. As of September 2026, no vehicle sold to consumers may operate with SAE Level 4 autonomy in unrestricted urban environments. What *is* legal — and widely deployed — is SAE Level 2+ ADAS with driver monitoring, continuous torque feedback, and mandatory fallback protocols. This distinction isn’t semantic. It’s the line between what you can buy today and what requires a government-issued pilot license.

H2: The Legal Framework — Three Layers of Control

China’s approach is layered: national policy (MIIT, MOJ), provincial enforcement (e.g., Shanghai’s Intelligent Connected Vehicle Pilot Regulations), and municipal implementation (Shenzhen’s 2024 V2X Roadside Unit mandate). Key instruments include:

• The Interim Measures for the Management of Intelligent Connected Vehicle Testing and Demonstration Application (revised March 2025): Requires all L3/L4 test vehicles to carry black-box logging, real-time telematics to local traffic authorities, and human remote operators for every 10 vehicles in pilot zones.

• The Automotive Driving Automation Classification and Evaluation Requirements (GB/T 40429–2026): Enforces strict driver engagement verification — eye-tracking + torque sensing + pedal position — for any system marketed as "intelligent driving". No exceptions for premium trims or subscription tiers.

• Local pilot policies: Shanghai allows unsupervised L3 operation *only* on designated expressways (G1503, G1501) and only for registered fleet vehicles (e.g., Baidu Apollo Robo-Taxis). Shenzhen permits L4 robotaxis *with safety drivers* in Nanshan District — but bans them from school zones and construction corridors after 8 p.m.

Crucially, there is *no national L3 consumer certification pathway*. Unlike Germany’s UN-R157 approval or Japan’s Type Approval for conditional automation, China treats L3 as a *fleet-only, geofenced, supervised capability*. That means even the Huawei ADS 3.0 system in the Aito M9 — marketed as "city NOA" — remains legally classified as Level 2+ because it lacks fail-operational redundancy and cannot sustain motion without driver readiness.

H2: Brand-by-Brand Reality Check (Updated: September 2026)

Let’s cut past marketing slides. Here’s how major EV brands implement AI driving assist *within current law* — and where they’re pushing boundaries:

• BYD: Blade Battery-equipped models (e.g., Yangwang U8, Seal U) run DiPilot 3.0 — a sensor-fused system (11 cameras, 5 radars, 12 ultrasonics) that handles highway merging, ramp navigation, and parking. But all functions require *continuous torque input*. No hands-free mode. DiPilot is explicitly labeled "driving assistance", not "autonomous driving", per MIIT registration documents.

• Xiaomi Auto: SU7’s XIAOMI PILOT uses NVIDIA Orin-X (508 TOPS) and a 128-line LiDAR. Its city NOA covers 18 cities — but only with driver supervision and a mandatory 30-second re-engagement reminder every 3 minutes. Xiaomi publicly confirmed in Q2 2026 earnings that its L3 application to MIIT remains under review, with no expected approval before Q1 2027.

• NIO: NIO NOP+ runs on the Adam supercomputer (4 chips, 1016 TOPS total). It supports urban navigation, valet parking, and intersection negotiation — but requires the driver to keep hands near the wheel at all times. NIO’s battery-swap network integrates with its ADAS map updates: each swap triggers an OTA check. Still, no model qualifies for L3 under GB/T 40429–2026 due to lack of redundant braking/steering actuation.

• ZEEKR: The 001 FR and 007 use Mobileye EyeQ6 + Qualcomm Ride Orin. Its “City Navigate on Pilot” activates only after 100 km of verified safe driving history (logged via NIO-style cloud scoring). Even then, it disables in rain >15 mm/hr or when roadside signage is occluded — per Shanghai Municipal Traffic Bureau testing protocol.

• Li Auto: AD Max 3.0 (on L9/Mega) uses dual Orin-X chips and 128-line LiDAR. It leads in cross-intersection turning and tunnel navigation — but its “hands-off” mode is *legally disabled by default*. Drivers must enable it manually per trip, and it deactivates if cabin temperature exceeds 38°C (per thermal fatigue risk modeling in MIIT’s 2025 Human Factors White Paper).

H2: The Hard Technical Boundaries — Why “Full Self-Driving” Isn’t Legal (Yet)

Three non-negotiable gaps separate today’s AI driving assist from lawful full self-driving in China:

1. Fall-back readiness: L3+ requires the vehicle to safely achieve minimal risk condition (MRC) *without driver intervention*. Current production systems — including Huawei ADS 3.0 and XPeng XNGP — rely on driver take-over within 10 seconds. That violates Article 7.2.3 of GB/T 40429–2026.

2. Redundancy architecture: Law mandates dual independent braking and steering control paths for L3+. No consumer EV meets this. BYD’s e-platform 3.0 has brake-by-wire redundancy; steering remains single-path. NIO’s NT 2.0 platform added dual EPS motors in 2026 prototypes — but none are certified or in production.

3. Data sovereignty & auditability: All ADAS logs must be stored locally (not just cloud-synced) and readable by provincial traffic bureaus using MIIT-approved decryption keys. Tesla’s Autopilot logs — encrypted end-to-end — failed Shanghai’s 2025 audit. Result: Tesla removed “Full Self-Driving” labeling from Chinese UIs in April 2026 and now displays only “FSD Beta (Supervised)”.

H2: What’s Coming Next — The 2027–2028 Regulatory Inflection

MIIT’s 2026–2030 Intelligent Mobility Roadmap targets L3 certification for *defined highway segments* by Q3 2027 — but only for vehicles with:

• Triple-redundant perception (camera + radar + LiDAR, all cross-validated in real time), • ISO 26262 ASIL-D rated steering/braking actuators, • Onboard V2X modems compliant with GB/T 31024–2026 (5G-V2X PC5 direct communication), • Real-time cyber-security intrusion detection (per CC EAL5+ certification).

Early candidates? The Geely-backed Zeekr 009 Grand (Q4 2026 launch) and BYD’s upcoming Denza Z9 GT — both confirmed to include triple-redundant chassis controllers and C-V2X roadside handshake capability. But even these will launch with L2+ firmware, awaiting final MIIT type-approval.

H2: Practical Advice for Buyers and Fleets

If you’re evaluating vehicles for personal or commercial use, ask these five questions — *before* signing:

1. Does the system require torque-on-wheel *at all times*, or only periodic confirmation? (Torque = legal Level 2; confirmation-only = gray zone — currently disallowed in Guangdong and Jiangsu.) 2. Is driver monitoring optical (camera-based) *and* physical (steering torque + pedal pressure)? If only one, it fails GB/T 40429–2026 Annex D. 3. Where is log data stored? Cloud-only = non-compliant. Local eMMC + encrypted cloud sync = compliant. 4. Does OTA update delivery require manual acceptance, or is it silent? Silent updates for ADAS logic are banned — MIIT requires explicit user consent per GB/T 42325–2026. 5. Is the vehicle registered under a provincial pilot program? Only Shenzhen, Shanghai, Beijing, Chongqing, and Hefei allow *any* form of unmonitored L3 testing — and only for approved fleet partners (e.g., Pony.ai, WeRide, Baidu).

For corporate fleets, note: MIIT’s 2026 Pilot Fleet Management Rules cap autonomous mileage per vehicle at 15,000 km/month unless paired with ≥1 remote operator per 5 vehicles — and those operators must hold Class A1 driving licenses plus 200 hours of teleoperation training.

H2: Comparative Snapshot — Capabilities vs. Compliance

H2: Beyond the Car — How Infrastructure Changes Everything

China’s edge isn’t just in vehicles — it’s in *orchestrated infrastructure*. By end-2026, 320,000 km of national highways and 2,100 urban intersections will have C-V2X roadside units (RSUs) broadcasting real-time signal phase, emergency vehicle alerts, and road friction data. This enables cooperative perception — where the car doesn’t need to "see" the ambulance coming around the blind curve; the intersection tells it.

That’s why brands like SAIC MG and Dongfeng’s Voyah are co-developing with provincial governments: their next-gen OTA updates won’t just refine neural nets — they’ll activate new V2X-triggered behaviors *only where infrastructure is verified*. A vehicle in Chengdu may auto-brake for jaywalkers at a V2X-equipped crosswalk; the same vehicle in Kunming (no RSUs) falls back to camera-only detection — and disables the feature entirely if confidence drops below 92%.

This infrastructure-first strategy reduces the burden on onboard AI — and accelerates regulatory trust. When the car’s decisions are corroborated by fixed, auditable infrastructure, the case for higher automation levels strengthens.

H2: Final Word — Don’t Confuse Capability With Permission

You can buy a car today that *technically* handles 95% of urban driving scenarios. But legality hinges on *how reliably it handles the other 5% — and who bears responsibility when it doesn’t*. In China, that responsibility remains unequivocally with the human driver — not the algorithm, not the brand, not the OTA update.

That won’t change overnight. But it *is* changing — methodically, transparently, and with teeth. Every LiDAR scan, every V2X handshake, every logged disengagement feeds MIIT’s national ADAS performance database. And when the data shows sustained <0.001 disengagements/km across 100,000 vehicles — that’s when L3 certification shifts from pilot exception to consumer standard.

Until then, treat every “AI driving” feature as what it is: a sophisticated co-pilot — not a replacement. Use it. Rely on it — within its documented ODDs. But keep your hands ready, your eyes scanning, and your understanding grounded in what the law permits *today*, not what the brochure promises tomorrow.

For a complete setup guide covering regional ADAS compliance, OTA update tracking, and V2X readiness maps, visit our full resource hub at /.