China ICV Standards
Objectives · Core Concepts · Implementation

Establish Unified AD Terminology

• Covers all road vehicles (Categories M/N/O)
• Defines 0–5 automation levels and their boundaries
• Standardises DDT, ODD and other core concepts
• Provides the terminological foundation for all upper-level standards

Dual-Dimension Classification by Function & Responsibility

Based on the SAE J3016 six-level framework, classification centres on "who executes the DDT" and "who bears responsibility for system failure." L0–L2: the driver is always the DDT executor; L3+: the system assumes the DDT but L3 still requires a driver ready to take over; L4: the system can autonomously execute an MRM; L5: no ODD limits. Key contribution: cleanly separating the responsibility boundary between "driver assistance" and "automated driving."

The L2/L3 Boundary Is Now a Legal Boundary, Not Just a Perception Gap

Until 2026 this was a marketing-versus-standard mismatch with no direct enforcement hook. It is now enforced from both sides: GB 47955-2026 defines what an L2 combined driving assistance product is permitted to do, and the draft ADS mandatory GB defines what an L3/L4 system must prove. A function that behaves like L3 while being sold as L2 now fails one standard or the other rather than merely misleading buyers.

E/E System Lifecycle Safety Development Framework

• Applies to all safety-related E/E systems in series-production road vehicles (except mopeds)
• Covers system, software and hardware development, production, operation and disposal
• Main products: ECUs, domain controllers, ADAS systems, brake/steer/powertrain controllers, BMS
• Part 12 adds motorcycle applicability; Part 11 adds semiconductor application guidelines

Risk-Based ASIL Classification + V-Model Development

Through Hazard Analysis and Risk Assessment (HARA), hazards are rated by Severity (S), Exposure (E) and Controllability (C) to determine ASIL (QM, A, B, C, D). Safety goals decompose into functional → technical → software/hardware safety requirements along the V-model. ASIL D requires diagnostic coverage ≥99% and PMHF ≤10⁻⁸/h. ASIL decomposition allows splitting high-ASIL requirements across independent redundant modules for engineering feasibility.

HARA Accuracy & Supply Chain ASIL Decomposition

• Challenge 1: ASIL allocation subjectivity; OEM–Tier1 rating discrepancies. Mitigation: HAZOP analysis for systematic hazard enumeration, validated via Functional Safety Assessment (FSA).
• Challenge 2: ASIL requirement opacity in the Tier2/3 supply chain. Mitigation: Per Part 8, establish Development Interface Agreements (DIA); suppliers must provide safety-element documentation.
• Challenge 3: Software ASIL D development cost. Mitigation: Proper ASIL decomposition via partitioning for coexistent multi-ASIL code.
• New in 2026: the MIIT 2026 work points call for dedicated functional-safety and SOTIF standards for driving-automation systems, by-wire chassis, battery management and drive-motor systems. The generic ISO 26262 transposition is being supplemented with domain-specific Chinese standards — expect subsystem-level requirements that the generic standard leaves to engineering judgement.

Addressing the "Correct but Insufficient" Gap Beyond Functional Safety

• Applies to systems where insufficient intended functionality or reasonably foreseeable misuse causes hazards
• Targets ADAS/ADS perception, decision and actuation modules; cameras, radar, LiDAR; AI/ML inference
• Not applicable to hardware random failures (governed by ISO 26262)
• Covers design, verification, validation and operation phases

Classified Governance of "Known Unknowns" and "Unknown Unknowns"

SOTIF sorts hazardous scenarios into four zones: Zone 1 (known hazardous, known trigger — eliminate); Zone 2 (known hazardous, unknown trigger — expose through testing); Zone 3 (unknown hazardous, unknown trigger — reduce to an acceptable level); Zone 4 (known safe). Core toolchain: insufficiency identification → trigger condition analysis → verification strategy → safety validation. Provides the framework for out-of-distribution inputs in AI perception systems.

Quantifying Acceptability of "Unknown Unknowns" — Now a Mandatory Problem

• Challenge 1: Zone 3 cannot be enumerated, making "sufficiently safe" hard to demonstrate. Mitigation: Large-scale simulation, scenario generation and mileage statistics to establish residual-risk baselines — now with a codified method in GB/T 47025-2026.
• Challenge 2: AI model opacity conflicts with explainable safety arguments. Mitigation: Performance-based verification supplemented by adversarial testing and coverage analysis.
• Challenge 3: Functional safety vs SOTIF boundary ambiguity. Mitigation: Use "triggered by hardware failure" as the criterion; run a joint Co-HARA process.
• Changed in 2026 — SOTIF is becoming mandatory by proxy. The draft ADS mandatory GB requires a lifecycle safety file built as claim → argument → evidence, refreshed on every hardware change and algorithm OTA. That is a SOTIF-style residual-risk argument with legal consequences attached. A recommended standard that enterprises could scope down at will now supplies the substance of a mandatory approval artefact.
• Frontier Challenge (🔥 still the hottest topic): LLM/VLM SOTIF risk. Generative models in the perception-decision chain expand Zone 3 dramatically: SOTIF assumes trigger conditions are identifiable, while probabilistic semantic generation explodes the trigger space from finite physical states toward near-infinite semantic combinations. Mitigations in practice: ① epistemic uncertainty as a backup criterion; ② a safety monitor over model output; ③ restricting foundation models from the safety-critical decision path. See the Automotive AI chapter — as of 2026 this is no longer a standards vacuum.

Standardise Full-Lifecycle Vehicle Data Protection

• Applies to all data generated and collected during R&D and production
• Covers complete vehicles, ICVs and NEV data processing
• Regulates general data requirements, personal information protection, important data protection and data classification
• Directly referenced by GB 44495-2024 — the data-compliance foundation under the cybersecurity mandate

Data Classification + Full-Lifecycle Personal Information Protection

Three tiers: ① classification — vehicle data split into personal information and important data, the latter including operation data, map data and road information, with enhanced protection where national security is implicated; ② full-lifecycle personal information protection — requirements across collection, storage, use, transmission and deletion, with anonymisation meeting an irreversibility criterion; ③ cross-border transfer control — important data shall not leave China in principle, with security assessment required for exceptions. A data-sovereignty clause with few parallels in international vehicle data standards.

Data Chain Risk Identification + An Impending Mandatory Upgrade

• Challenge 1: Complex data chains across vehicle, cloud, app and Tier1 blur risk-identification boundaries. Mitigation: maintain Data Flow Diagrams stating data category, processing purpose and retention period at each node.
• Challenge 2: JV OEMs must reconcile foreign-partner requirements with data sovereignty. Mitigation: domestic localisation nodes and de-identification before transmission.
• Challenge 3: Anonymisation criteria remain ambiguous. Mitigation: follow the annex test methods and engage third-party assessment.
• Challenge 4 (new in 2026): the recommended standard is on its way to becoming mandatory. ICV data security requirements and data interaction and management are being developed as mandatory national standards and were at consultation stage in the 2026 work points, alongside review-and-approval work on important-data identification, intrusion detection and information-security audit guidance. Mitigation: treat GB/T 44464 conformance as a floor rather than a target, and build the data inventory now — a mandatory version will require evidence, not intent.

The Technical Groundwork Beneath GB 47955

• Part 1: single-lane driving control (ACC + LKA — classic L2 combined longitudinal/lateral control)
• Part 2: multi-lane driving control, including lane change assist — the core of L2+ NOA functionality
• Scope: Categories M and N, all production vehicles with the function
• Main products: ADAS controllers with ACC + LCC/LKA, cockpit domain controllers, integrated driving-parking domain controllers

Driver-in-the-Loop + Clear Function Boundaries + Safe Exit

With driver-in-the-loop as the premise, the system must continuously monitor takeover capability (hands-on-wheel, gaze direction) while active. Core requirements: ① hands-off or eyes-off beyond threshold must trigger escalating warnings; ② the system must not create a misconception of vehicle control — no "autonomous driving" illusion; ③ lane change assist requires driver intent confirmation and may not autonomously execute lane changes. Key parameters: hands-off detection sensitivity, maximum response time, post-exit system behaviour. Note how directly the GB 47955 per-class envelope inherits this logic — the recommended standard supplied the concepts that the mandatory standard made enforceable.

Role Change: From Design Reference to Supporting Method

• Challenge 1: non-uniform compliance recognition for hands-on detection approaches (capacitive vs torque sensor vs camera). Mitigation: the standard is performance-based rather than prescriptive; demonstrate compliance through false-positive/false-negative performance validation.
• Challenge 2: multi-lane change scenario coverage is enormous. Mitigation: combine with GB/T 44719 road tests and — now published — GB/T 47025-2026 simulation, as a genuine three-pillar exercise rather than an aspiration.
• Challenge 3 (resolved in 2026): Before 2026, L2+ NOA straddled the L2/near-L3 boundary and was prone to design non-compliance, because no standard clearly owned it. GB 47955-2026 resolves this by construction — NOA is now a named product class with its own permitted envelope, rather than an unregulated space between two standards. The residual risk shifts from "which standard applies" to "which class does my function belong to."

From Parking Assist Toward Regulated Automatic Parking

• Scope: M/N passenger cars and light commercial vehicles with intelligent parking assist
• Covers parallel, perpendicular and angled parking
• Includes APA (driver retains pedal control) and fully automatic parking (no driver intervention)
• Speed range generally ≤10 km/h

Success Rate + Obstacle Detection + Dual Safety Exit

Evaluated on parking success rate (in both standard and tight spaces) and obstacle detection response. Key safety mechanisms: ① ultrasonic and visual obstacle detection front, rear and side, with a stop required within a specified distance; ② the driver may take over at any time (steering input or brake application exits immediately); ③ a safe default state when no action is preset (remains stationary after stopping). Note that the mandatory ADS standard explicitly excludes automatic parking from its scope — parking is being regulated on its own track, not folded into ADS.

Unstructured Environments — Now With a Mandatory Deadline Approaching

• Challenge 1: poor parking-lot marking quality (faded, incomplete, misaligned) drives high vision-based failure rates. Mitigation: performance assessment must include degraded-marking scenarios; ultrasonic is typically retained as a backup modality.
• Challenge 2: GNSS denial in underground garages degrades positioning accuracy. Mitigation: IMU + visual odometry fusion; memory-parking positioning accuracy is expected to be specified in the mandatory automatic-parking standard.
• Challenge 3 (new in 2026): the mandatory automatic parking system standard is at consultation stage, and parking combined driving assistance is also listed for consultation in the 2026 work points. Valet and summon functions that currently ship under a recommended standard will shortly face a mandatory baseline — including the unresolved question of who is responsible when no one is in the vehicle. Mitigation: do not assume the current recommended requirements represent the eventual floor, particularly for driverless parking manoeuvres.

The Technical Basis the First L3 Approvals Were Granted Against

• Scope: M/N vehicles equipped with ADS (L3 and above)
• Main products: passenger cars, commercial vehicles, AD trucks, robotaxi
• Regulates general requirements, DDT execution, DDT fallback, HMI requirements and verification methods
• Was the core technical basis for the first two conditionally licensed L3 products in December 2025 — this standard is what "L3 approved in China" currently means in technical terms

ODC Boundary + MRM + Multi-Pillar Validation

① ODC-constrained activation: ADS must define a clear ODC, activate only within it, and exit actively when exceeded.
② Driver takeover monitoring: ≥2 indicators (eye movement, head motion, etc.); TOR issued if the seat is left for >1 s or the seatbelt is unbuckled; MRM triggers automatically after 10 s without response.
③ MRM: hazard lights activated, system autonomously executes a safe stop.
④ Multi-pillar validation: audit assessment + simulation + proving ground + road testing — China explicitly elevates audit assessment alongside the UNECE NATM testing pillars — with DSSAD mandatory.

Every Soft Spot Here Is What the Mandatory Successor Hardens

• Challenge 1: MRM stop strategies vary enormously by scenario (highway/urban/tunnel) and the standard requires only "reasonable." Mitigation: define scenario-specific MRM defaults within the ODC and justify reasonableness through audit — in the mandatory version this justification becomes a safety-file obligation with quantified residual risk attached.
• Challenge 2: OEDR degradation at ODC boundaries (weather changes, road-type transitions) is hard to test exhaustively. Mitigation: simulate boundary-exit scenarios and accumulate road data near boundary conditions — now with GB/T 47025-2026 supplying the method and the mandatory standard demanding sim-to-real consistency.
• Challenge 3: trade-off between not disrupting traffic flow and conservative safety strategies. Mitigation: quantify conservative-behaviour impact through simulation and set acceptable thresholds.
• Transition planning: models approved against GB/T 44721 will need to demonstrate compliance with the mandatory successor within its transition window. Because the mandatory version adds a quantified benchmark and a lifecycle safety file rather than merely tightening parameters, this is not a delta-compliance exercise — the evidence base has to be constructed, and the recommended-standard approval package will not contain it.

Unified Foundation Element Set for ODC Definition

• Scope: M/N vehicles with ADS (other types may refer to it)
• Regulates the ODC foundation element set (Level 1: ODD + occupant status + vehicle status)
• Road type classification: 5 categories, 12 subcategories
• Dedicated digital-information chapter: V2X communication, HD map availability and other ODC elements

Binary "Permitted / Not Permitted" ODC Element Annotation

Every ODC element must be annotated "permitted" (does not affect activation) or "not permitted" (suppresses activation or forces exit), eliminating ambiguity. Hierarchy: Level 1 (road environment + vehicle status + occupant status) → Level 2 (specific parameters: max speed, min curve radius, minimum visibility) → Level 3 (sensor status, map update status). Key innovation: V2X availability and HD map coverage are ODC elements, reflecting China's vehicle-road-cloud integration approach — a more infrastructure-aware model than ISO 34503.

ODC Detection Becomes a Mandatory Real-Time Function

• Challenge 1: weather conditions as ODC elements are hard to sense accurately in real time. Mitigation: the standard permits cloud weather data fused with vehicle sensing; design conservative degradation (prefer a false exit over operating outside the ODC). The mandatory ADS standard raises the stakes by requiring the system to anticipate degradation and act before the boundary.
• Challenge 2: V2X coverage as an ODC element geographically limits availability while network deployment is uneven. Mitigation: a fallback mode outside coverage — ADS must not be fully dependent on connectivity.
• Challenge 3: the number of ODC elements makes definition work enormous. Mitigation: the foundation element set is non-exhaustive; select and extend as needed.
• New in 2026: with the UN ADS instruments scoping approval by declared ODD and the Chinese mandatory standard requiring continuous ODD self-assessment, ODC annotation moves from documentation exercise to runtime safety function. The binary permitted/not-permitted design turns out to be well-suited to this: it is machine-checkable, which a prose ODD description is not.

Real-Vehicle Validation Within the Multi-Pillar Framework

• Proving ground (41798): controlled scenarios in closed facilities — safe, repeatable
• Road (44719): real-world validation on open roads, covering naturalistic scenarios
• Complementary: proving ground validates typical hazardous scenarios, road testing accumulates naturalistic mileage and edge cases
• Covers L2–L4 AD function validation requirements

Scenario Standardisation + Test Data Traceability

Proving ground: standardised scenarios (cut-in, emergency braking, obstacle avoidance, intersection conflict) ensure cross-institutional comparability and reproducibility, with key items covering longitudinal/lateral control accuracy, perception range and takeover response time. Road testing: specifies route coverage (highway, urban, ramp), data recording (DSSAD required), test-operator safety procedures and incident reporting. With GB/T 47025-2026 these now form a genuinely complete three-pillar set rather than two pillars and a plan.

Coverage Completeness + Cross-Institutional Consistency

• Challenge 1: no standardised metric establishes whether test scenarios "sufficiently" cover the hazardous scenarios within an ODD. Mitigation: statistical coverage estimation across the joint simulation/proving-ground/road matrix — now materially easier to construct because the simulation pillar has defined pass criteria and a credibility regime, so its output can carry weight in the argument.
• Challenge 2: uncontrollable interference in road tests (sudden obstacles, pedestrian incursions) undermines repeatability. Mitigation: detailed logs and scenario classification, with natural events used as simulation replay datasets rather than direct compliance criteria.
• Challenge 3: differing proving-ground configurations make cross-institutional results incomparable. Mitigation: align with GB/T 43119 for unified hardware; a mutual-recognition mechanism remains under development.
• New in 2026: the 2026 work points place AD test scenarios and fusion positioning standards at review-and-approval. A national test-scenario standard would complete the picture by standardising the scenario library as well as the test methods — the piece that currently sits with ASAM internationally.

Post-Production Continuous Operational Safety Testing

• GB/T 43766: operational safety test technical requirements and overall framework
• GB/T 44850: companion test item list and test methods
• For operators and regulators requiring periodic operational safety assessment
• Covers continuous safety monitoring of in-service AD vehicles after production

Extending From Pre-Approval Validation to Full-Lifecycle Monitoring

The key distinction from GB/T 41798/44719: those target pre-approval R&D validation, this pair targets post-production operational monitoring. The logic is that AD software is continuously updated by OTA, so safety must be reconfirmed after each update. Operational safety testing establishes periodic inspection covering functional regression, critical sensor health assessment and ODD capability spot checks, linked with GB 44497 so that anomalous events in operational data trigger targeted testing. This is the domestic analogue of what the UN ADS regulation now calls In-Service Monitoring and Reporting — China built the test-method side first, the UN has now added the reporting obligation.

Post-OTA Regression at Fleet Scale

• Challenge 1: full-scale safety testing after every OTA is prohibitively expensive for large fleets. Mitigation: change-impact analysis and targeted regression on affected modules only, leaning heavily on rapid simulation validation — which now has a standardised method and a credibility regime behind it.
• Challenge 2: non-uniform safety-event classification in operational data impedes cross-regional aggregation. Mitigation: align with DSSAD data recording and establish unified event classification coding; the UN ADS ISMR requirements and US SGO experience both point toward a shared taxonomy, and this is where an internationally comparable dataset could realistically emerge first.
• New in 2026: once the mandatory ADS standard takes effect, operational safety testing stops being a recommended good practice and becomes the mechanism by which the safety file is kept true after each OTA. Its cost model moves from optional to structural.

Unify Facility Hardware Standards to Make Results Credible

• Scope: planning, construction and accreditation of AD closed test facilities nationwide
• Regulates functional zone division, road scenario configuration, test equipment and safety protection
• Provides the hardware basis for GB/T 41798 proving-ground testing
• Goal: resolve non-uniform construction across 40+ facilities and non-mutually-recognised results

Minimum Scenario Set + Modular Expansion

Specifies a foundation scenario set (straight high-speed section, intersections, ramps, parking areas, pedestrian crossing zones) plus optional expansions (tunnels, signal-controlled intersections, complex weather simulation zones). V2X RSU coverage of core test areas is required, reflecting vehicle-road-cloud integrated testing needs, with a unified data acquisition and transmission interface supporting remote monitoring and data sharing.

Coordinating Nationwide Facility Investment

• Challenge: existing facilities vary greatly and retrofit costs create weak compliance motivation.
• Mitigation: two levels ("basic pass" and "excellent") permit incremental investment toward basic compliance with progressive upgrade, supported by accreditation and subsidy policy.
• New in 2026: as of May 2026 more than 57,000 km of roads nationwide have been opened for ICV testing. Facility standardisation and open-road availability are no longer the constraint on validation volume — toolchain credibility and evidence assembly are.

C-V2X Direct Communication Vehicle System Specification

• Scope: vehicle systems based on LTE-V2X (PC5 interface) direct communication
• Not applicable to WiFi/DSRC (802.11p) based V2X
• Regulates protocol stack, message format, functional requirements, performance (latency, reliability) and security
• Application scenarios: FCW, EEBL, RLIA and 10+ others

C-V2X Strategic Route + Unified Cellular Evolution

China chose C-V2X on 3GPP cellular technology over DSRC/WAVE. Rationale: ① shared evolution path with 5G and future 6G — PC5 defined in Release 14, enhanced in Release 16 (5G NR-V2X); ② reuse of existing 4G/5G infrastructure via the Uu interface for beyond-line-of-sight cases; ③ supply chain maturity. This standard specifies the vehicle-side PC5 requirements — the compliance foundation for moving China's V2X from pilot to scale.

Interoperability + Security Credential Infrastructure

• Challenge 1: poor interoperability between modules from different suppliers; inter-vehicle messages not correctly parsed. Mitigation: interoperability test labs with test-method standards, plus the C-V2X certification system.
• Challenge 2: a unified national SCMS is not yet established. Mitigation: national V2X security credential management work continues; this standard references GB 44495 for the certificate management framework.
• Challenge 3: LTE-V2X to 5G NR-V2X upgrade compatibility. Mitigation: modular design, with an NR-V2X supplementary standard extending this framework.
• New in 2026: the 2026 work points advance direct-communication warning applications, platooning technical requirements and connectivity grading standards. Platooning is the notable one — it moves V2X from advisory warnings toward cooperative control, where a message loss becomes a vehicle-control safety issue rather than a missed notification, pulling V2X into functional-safety and SOTIF scope.

Technical Specifications for AD Perception Foundation Components

• GB/T 45086.1: vehicle satellite positioning (GNSS/BDS + RTK fusion), graded accuracy requirements
• GB/T 43249: vehicle passive infrared detection (occupant presence, animal/pedestrian detection)
• GB/T 43250: vehicle active infrared detection (short-range obstacle detection, blind spot monitoring)
• Applies to M/N series production vehicle sensing components

Graded Perception Performance + BDS Priority Adaptation

The positioning standard specifies BDS B1I/B2I/B3I band performance indicators and priority while remaining compatible with GPS/GLONASS/Galileo. Accuracy grades: standard (<5 m, navigation) → sub-metre (<0.5 m, L2 assistance) → centimetre (<0.1 m, L3+ precision). Infrared: passive IR addresses driver health and rear child presence (hot-car protection); active IR (NDIR/ToF) addresses short-range precise ranging, supplementing ultrasonic sensors.

Fusion Positioning Under Signal Obstruction

• Challenge 1: urban canyon multipath causes positioning drift, affecting L3 precision requirements. Mitigation: RTK + IMU + visual odometry multi-source fusion, with map matching correction where HD maps are used. Note that fusion positioning standards are at review-and-approval in the 2026 work points — this challenge is being addressed with a dedicated standard.
• Challenge 2 (active IR): strong direct sunlight reduces detector SNR. Mitigation: performance requirements specified across lighting conditions including direct sun; optical filters and adaptive exposure required.
• Challenge 3 (passive IR): high summer cabin temperatures approach body temperature, reducing thermal contrast. Mitigation: high-ambient testing (≥40 °C), driving differential thermal imaging or visible-light fusion solutions.
• New in 2026: positioning accuracy is no longer only a performance question. Because GB/T 45312 makes HD map availability and V2X coverage ODC elements, and the mandatory ADS standard requires continuous ODD self-assessment, positioning integrity — knowing when the position is untrustworthy — becomes a safety function. Accuracy grades alone do not express that; integrity monitoring and protection-level reporting do.