During the National Day holiday, those who took road trips must have seen numerous reminders on highways about using intelligent driving with caution; even those who didn't drive saw news about "falling asleep while using intelligent driving." This actually indicates that advanced driver assistance systems (intelligent driving) have begun to become widespread. However, both industry practitioners and consumers are secretly wondering, what stage has China's intelligent driving actually reached now?
As a senior practitioner who wrote "Autonomous Driving Product Manager," I really didn't expect how to answer this simply.
After all, there are product launches every day, and concepts change every year: mapless, end-to-end, VLA, world models, agents. Every company claims to be number one, and hearing too much of it just makes you more confused.
Later, I changed my approach. I found a company with the steadiest pace and the most transparent track record, reviewed its OTAs version by version over the past five years, and used it as a ruler to measure against Xpeng, Li Auto, and Horizon Robotics.
This company is Huawei ADS. From ADS 1.0 in 2021 to ADS 5 this year, there are a total of 26 version nodes.
After sorting it out, my conclusion is very clear: China's intelligent driving has a total of six steps. The first five have been completed, and we are now standing at the door of the sixth.
This article from Vehicle does not evaluate which company is the strongest; it only focuses on one thing: how China's intelligent driving climbed step by step, where it is now, and what the next step is.
Let's take it one by one, hoping to provide industry practitioners and consumers with some information about the development of China's advanced driver assistance systems (intelligent driving).
01. Why Use Huawei ADS as the Ruler? Steady Pace, Transparent Track Record, High Installation Volume, But Not First in Every Aspect
Let's clarify the methodology first, otherwise the subsequent conclusions won't hold up.
First, the pace is steady. Since 2023, Huawei has released a new generation of ADS every April, put it into vehicles 3 to 4 months later, and then spent another year refining the details via OTA. This hasn't been disrupted for five years.
Second, the track record is transparent. Since October 2024, the official OTA long images have printed version numbers, from V3.1 to Max V5.0.2. You can check monthly which feature arrived on which day.
Third, the installation volume is high. Huawei is a supplier; AITO, Luxeed, Stelato, Maextro, Avatr, and Arcfox all use it. The Qiankun Intelligent Driving installation volume exceeds 2 million units (self-reported by Huawei).
But one point must be made clear upfront: Huawei was not the first to arrive at every step.
For end-to-end deployment in vehicles, Xpeng was 4 months earlier than Huawei; for full-coverage spot-to-spot, Li Auto was 3 weeks earlier than Huawei. So when I use Huawei as the scale, I must definitely look at other companies during the same period for comparison.
To put it bluntly, Huawei's significance lies in its stability. Once it arrives, it basically indicates that this step is also established in the industry.
02. Steps One and Two: First Highways, Then Removing Maps. From 2021 to February 2024, Solving "Can It Be Driven Nationwide?"
2021 was the starting year for China's intelligent driving.
In January, Xpeng launched the public beta for highway NGP; in December, Li Auto OTA 3.0 introduced highway NOA. Huawei ADS 1.0 was released in April, but the first vehicle, the Arcfox HI edition, wasn't delivered until May 2022, a wait of 13 months in between.
The commonality at this step is reliance on HD maps. In September 2022, Huawei launched urban NCA in Shenzhen, opening it city by city; each city needed a map first.
Problems quickly emerged: HD map collection is expensive and updates are slow. Relying on it simply cannot expand nationwide.
Thus, 2023 became the "Year of Going Mapless":
• July 2023: Huawei ADS 2.0 mapless urban driving launched in 5 cities
• November 2023: Xpeng 4.4.0 mapless city NGP fully rolled out in 25 cities, expanding to 243 cities a month later
• February 2024: Huawei urban NCA pushed nationwide; it took 8 months to go from 5 cities to nationwide
• July 2024: Li Auto OTA 6.0 mapless NOA fully rolled out
There is another detail: during Huawei's 2.0 period, its 6 OTAs did not mark version numbers even once. At that time, companies were competing on "how many cities are opened," and no one cared about version numbers.
I believe February 2024 was the first watershed for China's intelligent driving: urban NOA changed from "pilot cities" to "available nationwide."
03. Steps Three and Four: End-to-End, Then Spot-to-Spot. From 2024 to 2025, This Is the Climax of the Entire Competition
After removing the maps, the next question is: does it drive like a human?
No matter how many rule codes are written, they can't cover all road conditions. So companies replaced rules with models, which is end-to-end.
After end-to-end was deployed in vehicles, the next was spot-to-spot: starting from a parking space in a parking lot and stopping at a parking space at the destination, without interruptions in between.
Looking at the timeline for these two steps, all three companies arrived within half a year:
| End-to-End Deployment | Spot-to-Spot | |
|---|---|---|
| Xpeng | May 2024 (XOS 5.1.0) | Jan 2025 (5.5.0) |
| Huawei | Sep 2024 (ADS 3.0) | Dec 2024 (V3.2 Beta) → Mar 2025 No Interruptions |
| Li Auto | Oct 2024 (OTA 6.4) | Nov 2024 (OTA 6.5, First Full Rollout) |
Comparing with steps one and two: Huawei's mapless urban was 4 months earlier than Xpeng, and mapless nationwide was 5 months earlier than Li Auto.
By steps three and four, the leader and the followers were only one or two months apart.
This was also when Huawei started publicly releasing version numbers. When the feature gap is small, they have to rely on version pacing to speak.
My judgment is that these two steps are the climax of China's intelligent driving competition. Three major events—mapless nationwide, end-to-end, and spot-to-spot—were implemented within these 14 months from February 2024 to March 2025, and each one will change users' judgment on "whether intelligent driving is usable."
04. First Half of Step Five: Intelligent Driving Enters the 150,000 CNY Tier. Features Can't Differentiate, So the Competition Becomes About Price
After steps three and four, the feature lists are becoming increasingly similar. The only thing companies can use to differentiate is one thing: how cheap can it be to get it.
Huawei's trajectory is the clearest. The high-end package dropped from 36,000 CNY in ADS 2.0 to 30,000 CNY in 3.0, and ADS 4 is priced at 32,000 CNY, but it is split into four tiers: SE, Pro, Max, and Ultra.
The Pro Enhanced Edition in November 2025 uses the Limera in-cabin laser visual solution, bringing urban NCA into the 150,000 CNY vehicle tier.
Others arrived almost at the same time:
• Nov 2025: Horizon Robotics HSD first deployed in Exeed ET5, starting at 149,900 CNY, with lifetime free high-end intelligent driving
• Jan 2026: Li Auto AD Pro 4.0, running urban NOA with a single Journey 6M chip at 128 TOPS
• First half of 2026: Horizon Robotics holds a 22.8% share in the chip market for urban NOA of independent brands, ranking second (self-reported by Horizon Robotics)
This is more important than any new feature.
Urban NOA has changed from a selling point to a threshold; a 150,000 CNY car not having it actually becomes news.
05. Second Half of Step Five: Driving Without Setting Navigation. In 2026, Four Companies Reached This Step Successively Within 5 Months
After prices came down, in 2026, all companies did the same thing: removing the last threshold of "you need to set it up first."
Previously, NOA required entering a destination first. Now, in open spaces, country roads, parks, and underground garages, you can activate it and drive at any time.
• May 2026: Xpeng 6.2.0, navigation-free NGP roaming
• Jun 2026: Horizon Robotics HSD V2.0, world model plus reinforcement learning, arbitrary point-to-point
• Jul 2026: Huawei Max V5.0, anytime anywhere NCA, RCA roaming cruise
• Sep 2026: Li Auto 9.1.1, Mach VLA 2.0 distilled to nearly a million older vehicles
Then in September, a very interesting signal emerged.
On September 25, Huawei officially announced Max V5.0.2, adding group travel, multi-person intercom, AR find-my-car, and charging station busy/idle status. On September 22, Xpeng pushed 6.3.0, focusing on the Master Agent, planning the entire trip with a single sentence. You can refer to our previous article "Xpeng Second-Gen VLA Breakdown: Understanding the Frontier Vehicle-End Large Model Architecture of Intelligent Driving and Cockpit in Four Charts".
You will find that among the newly added items, very few are actually related to "driving." Of course, recently DeepRobotics and Horizon Robotics have been promoting all-scenario reverse parking. Actually, Tesla has already launched this feature at least half a year ahead, just without promoting it separately.
Huawei ADS 5 calls its architecture WEWA 2.0, which includes the Driving Agent; Xpeng calls it Master Agent; Li Auto calls it Livis, directly talking about embodied intelligence. All three are moving towards "mobility agents."
To put it bluntly, there is almost nothing left to add to advanced driver assistance itself.
06. So, What Stage Has China's Intelligent Driving Actually Reached? The Ceiling of L2 Is Basically Touched, and Liability Still Lies with Humans
Combining the previous chapters, my answer is: the first five steps are completed, and the sixth step is still at the door.
Why say the first five steps are completed? Three signals:
1. Feature convergence: The official announcements of Huawei V5.0.2 and Xpeng 6.3.0 were only 3 days apart, and their feature lists can be matched item by item.
2. New additions changing flavor: New OTA features have changed from "where it can drive" to AR find-my-car and charging station electricity prices.
3. Price bottoming out: Urban NOA has entered the 150,000 CNY tier, with some models offering it for free for life.
Why say the sixth step hasn't arrived? Because as of today, if an accident happens while driving, the human is still responsible.
No matter how many roads the system can drive, legally it is still L2 advanced driver assistance. You must keep your hands on the steering wheel and eyes on the road. The Arcfox model obtained L3 access in December 2025, but there is still a distance before ordinary people can buy it and use it with peace of mind.
Note: This is my judgment based on public information. The "all-scenario" and "zero takeovers" in various companies' statements are mostly self-reported, and the actual experience varies by vehicle model and road conditions.
07. What Is the Next Step? From "Can It Drive" to "Who Is Responsible," the Competition Has Completely Changed
The biggest difference between the sixth step and the first five is: it mainly relies on regulations, insurance, and liability determination; having a good model is just a prerequisite.
01 Liability: Who Pays for L3?
Huawei's statement is that L3 will be commercially used at scale in 2027, with Huawei and automakers sharing liability, equipped with 6 million CNY insurance; in 2028, L4 will be commercially used at scale for low-speed urban areas. Xpeng plans to trial operate Robotaxis within 2026.
My judgment is that only the first one who dares to clearly write in the contract "I will pay if an accident happens while the system is driving" has truly stepped onto the sixth step; the rest can only be considered marketing.
02 Cost: Who Sells It Cheaper?
High-end intelligent driving has already entered the 150,000 CNY tier; the next step is the 100,000 CNY tier. Horizon Robotics HSD's target for 2026 is 400,000 units, and the 3- to 5-year target is 10 million units. This battle is about chips and supply chains.
03 Agent: Who Manages Mobility?
After the act of driving is completed, what remains are "outside the car" tasks like finding parking spaces, charging, grouping up, intercoms, and ordering food. Whoever strings these into an agent first will occupy the users before the sixth step.
In the next half year, regarding the development of China's advanced driver assistance, I suggest everyone pay attention to these things:
China's Intelligent Driving · Observation Checklist for the Next Half Year
1. Huawei V5.X (official window Sep-Oct 2026): Active and passive fusion, meeting, reversing, and extrication; whether it will be fully pushed.
2. Li Auto OTA 9.2 (planned for Oct 2026): 3D ViT, perception distance 250m+.
3. Horizon Robotics HSD: Deployed in a leading NEV model within the year; can the 400,000 units target for 2026 be achieved?
4. Xpeng Robotaxi: Will the trial operation land within 2026?
5. L3: Besides the Huawei ecosystem, which other companies obtain access, and how are the insurance and liability clauses written?
In the future, when you see intelligent driving news, you can run it through these questions to quickly know which step it is at:
1. Does this feature require HD maps? (Yes = Step 1)
2. Is it written by rules or learned by models? (Rules = Before Step 3)
3. Can it drive from parking spot to parking spot without interruption? (Yes = Step 4)
4. Can it be used without setting navigation, even in a 150,000 CNY car? (Yes = Step 5)
5. If an accident happens, who is responsible? (System is responsible = Step 6)
The 5th question is the most critical. Currently, the answer for all mass-produced cars is still "yourself."
08. In Conclusion: Features Will Become Standard, Liability Is the Real Threshold
After sorting out these 26 nodes of Huawei, my biggest feeling is: China's intelligent driving has run really fast in these five years.
From driving city by city relying on maps to activating it anytime and anywhere today, it took less than four years. The gap between leaders and followers has also shrunk from a year to one or two months.
But the further we go, the fewer problems can be solved by engineers working overtime. The sixth step requires regulations, insurance, liability determination, and whether users are truly willing to let go of the steering wheel.
Finally, remember three sentences:
1 Six steps, the first five are completed
2 Features converge, the competition is about price and scale
3 The threshold for the next step is who is responsible
Next time you see "all-scenario" or "zero takeovers," ask first: Who is responsible if an accident happens? The answer is the true step it is at.
This article is compiled from: Huawei Qiankun official WeChat account and historical OTA long images, Huawei official website, Harmony Intelligent Mobility announcements.
Data cross-referenced with: Xpeng XOS / Xmart OS update notes, Li Auto OTA update notes, Horizon Robotics launch events and HSD push announcements.
Some nodes refer to IT Home, Phoenix News, and other media's transcripts of official long images.
Terminology: "Deployed in vehicles" refers to the first push to users or first vehicle delivery; V5.X, OTA 9.2, Robotaxi, L3 / L4 timelines are corporate plans.
Installation volume, market share, usage rate, and other data are self-reported by the companies.
This article is compiled based on public information. The "six steps" is an analysis framework defined by Jack, not an official grading standard; it contains forward-looking judgments for reference only.
That's all for today's sharing. If you find it useful, please help like, share, and click "watching". You are also welcome to discuss your views in the comment section. Finally, I made a long image of Huawei's intelligent driving OTA to share with everyone.
*Strictly prohibited to reprint or excerpt without permission - To obtain the reference materials for this article: join our Knowledge Planet to download a massive amount of public account reference materials, including the above reference materials.