Chinese Independent OEMs' ADAS and Autonomous Driving Report, 2026

Research on OEMs' Intelligent Driving: Era of Physical AI, Standard Configuration of D2D, and Initial Exploration of L3 Commercial Pilot Projects

From 2023 to 2025, the intelligent driving installation structure of passenger cars in China has shown a clear trend of stepped upgrading and structural substitution. Non-intelligent driving level (NL) and low-level intelligent driving (L1) have seen declining installation volume, and have gradually withdrawn from mainstream market, confirming that intelligent driving is fully becoming a standard configuration for passenger cars; basic high-level intelligent driving (L2) remains the absolute foundation of the industry. Although its installation volume slightly declined in 2025, the basic market is stable, and the industry's growth focus has clearly shifted to higher-level assisted driving. Wherein, L2.5 highway NOA and L2.9 urban NOA have become the core growth engines, with their installation volumes achieving a substantial jump in 2025 respectively, and the penetration rate of high-level NOA functions rises rapidly. Yet the installation volume of L2+ has shrunk slightly, indicating that its functional value is being replaced by more complete high-level solutions such as L2.5/L2.9. The overall intelligent driving market presents a clear pattern of "low configuration clearance, stabile basic market, and high-level outbreak".

In terms of OEM structure, independent brands and joint venture/foreign brands show a distinct differentiation of "radical skipping" and "conservative progressive progress" in intelligent driving upgrading:

Independent brands have skipped low-level and vigorously developed high-level intelligent driving. The proportion of non-intelligent driving (NL) dropped sharply from 60.9% in 2022 to 36.0% in 2025, and low-level intelligent driving of L0-L1 was almost zero (1.3%). In the same period, the proportion of high-level intelligent driving such as L2 and above doubled to 62.7%. Among them, the installation rate of L2.9 urban NOA has risen from 2.1% in 2022 to 17.2% in 2025, realizing a leapfrog structural shift from "dominated by non-intelligent driving" to "dominated by high-level intelligent driving".

In contrast, joint venture/foreign brands have adopted a conservative route of steady substitution and hierarchical iteration. The proportion of non-intelligent driving has dropped from 40.0% to 14.5%, but low-level intelligent driving of L0-L1 still retains a considerable share of 16.4%. Although the proportion of high-level intelligent driving of L2 and above has risen to 69.0%, the overall market still maintains a decentralized pattern of coexistence of multiple levels including "non-intelligent driving + low-level + high-level", forming a sharp contrast with the radical route of independent OEM brands.

2026 will become a key inflection point for China's automotive industry to move from "quantitative change" to "qualitative change" in intelligent driving capabilities. By systematically sorting out intelligent driving strategies, strategic layout, technical routes and implementation progress of 15 Chinese independent OEMs from 2023 to 2026, the analyst has summarized four core insights.

Insight 1: the core of intelligent driving competition has shifted to generational innovation of underlying architectures. The industry is fully entering the era of physical AI driven by large models from traditional assisted driving relying on rule programming, realizing human-like decision.

The essence of physical AI is to deeply integrate physical laws, large models and world common sense into intelligent driving system, fundamentally solving the shortcoming of traditional AI's "physical blindness". Traditional artificially rule-driven intelligent driving can only identify targets such as vehicles, pedestrians and traffic cones, lacking an understanding of physical world and causal logic, and unable to predict behavioral intentions. It is prone to jamming, sudden braking, misjudgment and other problems in long-tail scenarios not covered by rules.

Physical AI intelligent driving, through multi-modal perception, not only identifies pixel information, but also understands 3D space, depth, motion state, object material and physical properties in an integrative way, accurately grasps spatial constraints, motion trends and causal relationships, realizes intention prediction and risk deduction, and completes low-latency execution of perception-inference-control through end-to-end closed loop, making decision closer to the fluency and robustness of human driving.

Intelligent driving algorithms in physical AI era can be roughly divided into the following routes:

VLA (Vision-Language-Action) route: integrates vision, language and action modalities to realize end-to-end intelligent decision from environmental perception and instruction understanding to behavior execution. Representative companies: Li Auto, XPeng, Deeproute.ai, Xiaomi, etc.

World Model route: Constructs an abstract representation of the virtual world by learning dynamic laws of the environment, and optimizes decision strategy of the Agent through low-cost simulation and deduction. Representative company: NIO

One-Model E2E + Reinforcement Learning + World Model: Directly maps raw input to action output, omits the link of manual feature design, and independently and iteratively optimizes decision strategy with the help of environmental reward signals. Representative companies: Momenta, SenseAuto

Hybrid Mode: For example, Geely launched the World Action Model, integrating VLA + End-to-End Safety Adversarial Model + World Model

Li Auto's intelligent driving technical route has undergone several switches: from HD map-dependent, rule-based solution to "end-to-end" → "dual-system solution (end-to-end + VLM) → VLA → MindVLA-01". Li Auto's core of intelligent driving in 2026 is to fully switch to the MindVLA-01 unified foundation model, taking end-to-end VLA + world model + closed-loop reinforcement learning + self-developed chips as the path, aiming to build a general Agent in the physical world, and it will be mass-produced and launched on all-new L9 in 2026 Q2.

MindVLA-01 takes the native multi-modal MoE-Transformer as the unified foundation, and fuses three modalities of vision, language and action at the bottom; realizes accurate environmental perception through 3D space understanding (3D ViT + feedforward 3DGS); has multi-modal prediction and in-depth thinking with the help of predictive latent world model; relies on unified action generation (Action Expert + parallel decoding + discrete diffusion) to output automotive-grade stable control; and realizes rapid model iteration and efficient on-vehicle deployment through large-scale closed-loop reinforcement learning (MindRL) and the software and hardware collaboration of self-developed Mach chip, and comprehensively builds a physical AI intelligent driving brain integrating "seeing - thinking - acting".

Insight 2: intelligent driving functions are evolving from "HD map-free urban NOA" to standard configuration of "Door-to-Door (D2D)".

Emerging OEMs such as XPeng and Li Auto launched parking space → city → highway uninterrupted D2D intelligent driving in January 2025. Major mainstream OEMs have accelerated the implementation of D2D from 2025 to 2026, and the industry has entered the era of full-process intelligent driving from "segmented assistance".

In terms of technical routes, emerging OEMs generally adhere to the strategy of independent R&D and tackling key problems to seize technological high ground. XPeng, NIO and Li Auto all adopt a combination of independent algorithms and high-compute chips, relying on about 1000TOPS-level computing power to support full-scenario map-free D2D, with significant technological leadership. Among them, Li Auto plans to launch the self-developed and mass-produced intelligent driving chip Mach 100 for the first time on All-new Li Auto L9 Livis in Q2 2026. The chip adopts a data stream native architecture, which can be deeply adapted to the MindVLA-01 VLA large model, with a single-chip computing power of up to 1280TOPS. In terms of the computing power of self-developed chips (Li Auto Mach 100 (1280TOPS) > NIO (1000TOPS) > XPeng (750TOPS)), the independent R&D camp overall enjoy a bigger lead than the cooperative camp.

Traditional OEMs are more inclined to a pragmatic path of strategic cooperation and rapid implementation. Huawei's partners such as SAIC and BAIC directly use Huawei ADS 4.0 + Ascend chips to realize rapid adoption of D2D functions, with high technology multiplexing rate; Chery and BYD adopt mature solutions such as Horizon and NVIDIA, focusing on cost-effective D2D to cover the mainstream vehicle market of RMB150,000-300,000. It is worth noting that D2D functions do not rely on extremely high computing power. Third-party chips of 200TOPS level (such as Huawei Ascend 610, single Orin-X) can support the full-scenario link, and higher computing power is mostly reserved for technical redundancy and subsequent high-level autonomous driving upgrades.

For chip pattern, the D2D intelligent driving market presents a tripod pattern among NVIDIA, Huawei and Horizon, and self-developed chips are becoming the core technological moat of leading emerging OEMs. NVIDIA is still the absolute mainstream of general-purpose chips in the industry, with Thor-U prevailing among cooperative models at 700TOPS level. Huawei has deeply bound with traditional OEMs with its "chip + ADS full-stack solution" and quickly cut into high-end intelligent driving market. Horizon focuses on mid-end and mainstream mass market, achieving extensive coverage with cost-effective solutions. Simultaneously, self-developed chips have become a key to the differentiation of emerging OEMs, realizing the maximization of model efficiency and intelligent driving experience through deep customized collaboration of computing power and algorithms.

In the future, driven by the maturation of end-to-end large model technology, the cost reduction of intelligent driving chips and scenario closed loop, D2D is expected to evolve from a high-end optional configuration to a mainstream standard configuration, and become the core competitiveness benchmark of intelligent vehicles from 2027 to 2028. With the standardization of D2D functions, competition will shift from "available or unavailable" to "good or bad". The adaptability to extreme scenarios (mountain cities, underground garages, rain and snow weather), zero disengagement and cost control will become core decisive factors.

Insight 3: L3 technology is ready in 2026, exploring the commercial inflection point initially, and L4 will usher in a boom period from 2027 to 2030.

1) Initial exploration of L3 commercial inflection point: in March 2026, Changan Automobile obtained the official special license plate for L3 autonomous driving, which means that "verification of L3 technology maturity is completed and it has entered the early stage of mass production and commercialization". Previously, vehicles of L3 and above could only operate in closed parks or specific test areas. After obtaining the official license plate, vehicles are allowed to carry out on-road use pilots on expressways and urban trunk roads in Chongqing (such as Inner Ring Expressway and Yudu Avenue). At present, these vehicles are operated by Changan's mobility company, and consumers can now experience them by ride hailing instead of buying them directly. This method can reduce the uncertainty of initial implementation through professional management. As of April 2026, two OEMs, Changan and BAIC, have obtained the official special license plate for L3, with the qualification for commercial pilots on public roads.

2) OEMs' strategic deployment of L3/L4 intelligent driving: two major routes of L3 skipping and L3+L4 parallel development

At present, the industry has formed two clear strategic paths for high-level intelligent driving:

One is L3 skipping route: take Robotaxi as the core breakthrough, skip L3 mass production in strategy and directly develop L4 technology. Although L3 is compliant, the investment (computing power/redundancy) in it is close to L4 but the experience is limited, leading XPeng/BMW to choose to skip L3. Taking XPeng as an example, it plans to launch 3 mass-produced OEM L4 Robotaxis in 2026, start the normal road test of L4 autonomous driving in H1 2026, officially launch the demonstration operation of Robotaxi in H2 2026, complete the tripartite verification of technology, customers and business, and realize non-safety officer commercial operation in 2027.

The second is L3+L4 parallel development route. The core choice of current mainstream OEMs follows a steady rhythm of "qualification verification → mass production internal testing → L3 launch → L4 implementation", covering both private passenger cars and Robotaxi tracks simultaneously, with a complete technical route and clear mass production rhythm.

In terms of time layout, 2025-2026 is a critical period for L3 road testing, mass production & access and product delivery. SAIC, Geely, BAIC and GAC have clarified product technical readiness of L3 intelligent driving capabilities for private vehicles in 2026 (pending legal permission), marking 2026 as the "mass production first year" of China's L3 autonomous driving. Car owners will legally obtain the right to take their hands off the steering wheel on specific roads (highways/expressways) for the first time. 2026 is the starting point for large-scale operation of Robotaxi. Mainstream OEMs (SAIC, Geely, BYD, BAIC, GAC) all anchor the substantive commercial implementation of L4 in 2026, and focus on core areas of first-tier cities such as Shanghai, Shenzhen and Beijing.

In the medium and long term, 2027-2030 is a window period for the implementation of L4 in complex scenarios and the wide adoption in private cars. Changan and Dongfeng have clarified large-scale mass production of L3/L4 by 2030, indicating that the industry will fully penetrate from the "business/operation end" to "consumer users" in the next 5 years.

Insight 4: Cockpit-driving fusion is expected to accelerate, and automobiles are evolving rapidly into AI Super Agents.

The ultimate form of intelligent vehicles is a digital living body and mobile intelligent terminal with autonomous capabilities. At present, the industry is gradually developing from the architecture of separate cockpit and driving domains to the direction of cockpit-driving fusion, cockpit-driving integration and chassis full-domain fusion, and eventually moving towards the form of intelligent mobile robots with autonomous decision and autonomous execution capabilities.

In March 2026, IM Motors launched the IM Ultra Agent, an AI super agent which realizes in-depth collaboration of three domains of intelligent driving, intelligent cockpit and chassis based on IM Fusion Nova architecture. At the hardware level, full chassis-by-wire is the foundation for realizing full-domain vehicle control. IM Motors LINGXI Digital Chassis adopts full-stack wire-controlled solution, with four-wheel steering response time as low as 20ms, and response efficiency about 4 times that of traditional steering system. It is also equipped with an aviation-grade triple safety redundancy architecture, with the system failure probability lower than 10FIT, providing a stable and reliable hardware foundation for high-level intelligent driving and vehicle dynamic control.

At the software level, vehicles are equipped with Alibaba Tongyi Qwen large model, providing multi-modal interaction and continuous evolution capabilities for IM Ultra Agent. IM AD ZETA, an intelligent driving system jointly developed with Momenta, adopts a new-generation reinforcement learning large model as the physical AI foundation oriented to L4 autonomous driving, realizing the integrated upgrade of perception and decision capabilities on vehicles. The large model can realize real-time linkage between decision layer, intelligent driving domain and chassis domain, and support one-sentence voice commands directly to vehicle control, making cross-domain collaboration and full-scenario assisted driving move from concept to practical application.

When intelligent driving, intelligent cockpit and chassis are integrated, a single AI command can coordinate the intelligent cockpit and intelligent driving AI large models:

Scenario example: During the evening rush hour after work, the user issues the instruction: "I'm too tired, want to go home, and buy a cup of hot Americano by the way, preferably without getting off the car to pick it up."

The on-device Alibaba Tongyi Qwen AI large model completes natural language understanding and user intention disassembly, dividing the requirements into three categories:

Vehicle control requirements: "I'm too tired" → activate the seat massage function, executed by vehicle control Agent;

Life service requirements: "Buy a cup of hot Americano" → purchase hot Americano coffee, link with IM Motors’ takeaway Agent to complete coffee selection, payment and pick-up point association;

Mobility path requirements: first go to the pick-up point, then return to the destination to go home.

Finally, the IM AD ZETA intelligent driving large model unifies overall planning, completes dynamic path planning, real-time road condition prediction and full-process intelligent driving execution, realizing a one-stop experience of "picking up food without getting off the car + automatic homecoming".

From 2023 to 2025, Chinese passenger car intelligent driving market has completed the initial iteration from "available or unavailable" to "good or bad". The installation structure presents a clear pattern of " stabile basic market and high-level outbreak", and the route differentiation between independent and joint venture brands has become more prominent. In 2026, a key inflection point for China's automotive intelligent driving to move from "quantitative change" to "qualitative change", underlying architecture has fully entered the era of physical AI, with multiple technical routes such as VLA and world model developing in parallel. D2D functions are developing faster from high-end optional configuration to mainstream standard configuration, L3 autonomous driving is ushering in initial commercial exploration, and cockpit-driving fusion is promoting the steady evolution of automobiles into AI Super Agents. The dual paths of chip self-development and strategic cooperation are reshaping the industry competitive pattern. It is foreseeable that in the next two years, the competition in intelligent driving capabilities will no longer be limited to the simple stacking of algorithms and computing power, but will more depend on enterprises' systematic construction of physical world understanding, data closed-loop efficiency, depth of software and hardware collaboration and breadth of scenario coverage. In this process from quantitative change to qualitative change, the real decisive factor will belong to those players who can take the lead in realizing in-depth integration of "cognitive intelligence" and "action intelligence".