From 0 to 1, it relied on a team of fierce stallions; from 1 to N, it relied on a group of young people with an average age of 30 and a bunch of terminal veterans.
In the entire history of Huawei's intelligent driving, the Shanghai Auto Show in April 2021 remains the most memorable scene. A BAIC Arcfox engineering prototype without airbags navigated through the jungle of food delivery riders outside a rural compound in Shanghai with zero takeovers throughout the journey. That video went viral on Bilibili and WeChat Moments, making the outside world realize for the first time that a telecommunications company, in a field it had "just started," suddenly played a trump card.
But that was the victory of a demo, not a product.
The real battle began only after the applause faded. And it started in a way that was almost despairingly awful.
In July 2021, Su Qing, the technical leader who had set his sights on Waymo, was transferred away by a document. Almost simultaneously, an acquisition negotiation, premised on "selling the team" and nearly leading to the entire ADS team changing hands, was taking place on the conference table between Huawei and Volkswagen, which ultimately came to nothing.
This article will not expand on these two events themselves. Their true consequence was leaving a rare chaos for ADS:
The technical leader departed, and the new management took over. Morale fluctuated, people took sides, and waited to see how things would unfold due to a failed acquisition. A team once praised for having the right time, place, and people realized for the first time that they could be "sold."
Tackling challenges in a stable state is already hard; tackling them in chaos is even harder.
The time span of this article is from April 2021 to April 2023, which is the birth process of ADS 2.0. Over these two years, the idealists who went from 0 to 1 exited, and the brothers who went from 1 to N took over. The two generations waxed and waned, turning a radical intelligent driving technology from a belief into mass production.
(Note: The technical R&D time for both ADS 1.0 and ADS 2.0 was two years, and subsequent versions from 3.0 to 5.0 have maintained a release rhythm of one version per year.)
01、The Bill of the Old World: HD Maps
To understand why ADS 2.0 had to happen, one must first understand how expensive the old world it overthrew was.
When the Car BU was established in 2019, the default route for the entire industry was HD maps. The logic of this paradigm was very self-consistent: since the car's ability to understand the world is limited, map the world in advance, store it in the car with centimeter-level precision, let perception only be responsible for identifying dynamic obstacles, and let planning and control follow the pre-marked lane lines in the map.
But the accounts didn't look good. Building an HD map for a single city cost between 1 billion and 2 billion RMB; annual maintenance fees were another 200 to 300 million RMB—road construction, route changes, and repainting of markings meant the freshness of the map had to be maintained by mapping vehicles running over and over again. Huawei set up an HD map platform department in Beijing, burning through billions of RMB over a few years, plus two thousand outsourced personnel, only to cover two and a half cities in the end. More fatally, there was the qualification issue: only six cities nationwide allowed HD map surveying and mapping business, while Huawei's cars were to be sold nationwide.
The ADS 1.0 equipped on the BAIC Arcfox at the Shanghai Auto Show in April 2021 was the pinnacle of the HD map solution, and also the endpoint of this solution.
In mid-2021, the skeleton of the Car BU was rearranged. Yu Chengdong was appointed CEO of the Car BU; Bian Honglin was appointed CTO of the Car BU and Minister of the R&D Management Department. The ADS Department under them was headed by Li Wenguang and Zhao Qian. One served as the SPDT President, and the other as the PDU Minister.
This is Huawei's unique matrix management: SPDT manages product definition, marketing, and sales; PDU manages technology and resources along one line. Most of them were veterans of Huawei's terminal business, slightly unfamiliar with intelligent driving technology, so they simply entrusted professional matters to various experts. This hands-off approach, instead, facilitated the golden age for many.
Yu Chengdong's way of making demands was very typical, directly deducing from commercial reality: "You've kept 2,000 people and spent two years, but the HD map can only cover parts of Beijing, Shanghai, Guangzhou, and Shenzhen, running in at most two cities. How can we sell this car?"
But the three words "do mapless" translated into technology was a reconstruction that affected the whole body. In the entire intelligent driving system, what ultimately controls the vehicle is planning and control, so mapless was first pushed forward from planning and control. This task first fell on Wang Xinyu, the Chief Expert of Planning and Control.
02、Wang Xinyu: Uprooting the Map
A Planning and Control Engineer Who Emerged from Military Unmanned Vehicles
Wang Xinyu, a young man born in 1989, completed his bachelor's, master's, and Ph.D. at the Beijing Institute of Technology (BIT). His supervisor was Academician Fu Mengyin, conducting research on military unmanned ground combat platforms. In March 2017, he graduated with a Ph.D. and joined Huawei's Noah's Ark Lab as a fresh graduate, becoming one of the earliest people at Huawei to work on autonomous driving planning and control.
There were two competitions in domestic autonomous driving in those years: one was the "Intelligent Vehicle Future Challenge," and the other was the military's "Crossing Obstacles." BIT participated in both. In 2014, Baidu invited several high-level domestic teams to demonstrate at the entrance of Baidu's Xiamen building. Wang Xinyu's lab car drove a few circles there—after which BIT and Baidu had close cooperation.
In 2016, during his final year of his Ph.D., he interned at Baidu IDL, working on behavioral decision-making with Tao Ji, Yang Wenli, and others. That year, Baidu's autonomous driving big shots collectively left to start businesses. In March 2017, holding offers from Baidu, Huawei, and several others, he chose Huawei and entered Noah's Ark Lab.
Noah's Ark had a technical project called "Cognitive Internet of Vehicles" at the time, with investment from both the group and the Shanghai IOV Internet of Vehicles Innovation Center—the project director on the IOV side was Chen Qi, and there was not a single person on the team with an intelligent driving background. Wang Xinyu was the first formally trained intelligent driving engineer recruited for this project.
His title was the head of the planning and control group and project SE (System Engineer)—responsible for both planning and control algorithms and stringing the modules of perception, fusion, road structure, and planning and control into a runnable whole; he even wrote the simulation environment himself. The planning and control team started recruiting from just him, growing to over twenty people when he left Noah's Ark; the entire cognitive Internet of vehicles project had seventy or eighty people.
At the end of 2017, the car was running. Chen Qi took a look: "How did you get it done so fast? That's great." It was also in this year that a man named Su Qing noticed this team from HiSilicon.
The Iron Triangle and Two MMDs
Su Qing had been leading the definition of an automotive-grade intelligent driving chip—the 1951—at HiSilicon Turing since 2015 and 2016. After the chip was defined, he needed a team that could write algorithms. In early 2018, he directly "bought out" the entire Noah's Ark intelligent driving team where Wang Xinyu was located.
The company's original positioning for this team was "one body, two wings": the body was the strategic task assigned by Xu Zhijun—exploring L4 and L5 independently, using intelligent driving to drive the improvement of Huawei's full-stack AI (Artificial Intelligence) capabilities; the two wings were serving IOV and HiSilicon, two internal customers. Su Qing overturned this design with a single sentence: "What one body, two wings? No, just one wing, no body and no that wing."
At the end of 2018, Volkswagen Group organized MMD—Management Driving—in Anting, Shanghai, an annual routine event where German executives flew to various parts of the world to test drive their own models and local competitors. This year, they decided to take the opportunity to check the intelligent driving capabilities of Chinese suppliers, finding six or seven: on the Audi line were Huawei and another company; on the Volkswagen line were Baidu, DJI, Tencent, UISEE, etc.
The task given by Volkswagen was to drive a lap inside the courtyard of the Crowne Plaza Hotel in Anting, where scenarios were artificially set up—a dummy person popping out to yield, a traffic light set up to stop, all typical tricks of that era. All teams obediently followed suit.
Huawei's side said: We won't drive inside; we'll drive outside. Volkswagen asked, can you handle it? Audi asked, are you unable to handle it?
The team only had a few months to prepare, using a Magotan modified with the help of Hedong. In the end, Huawei's car drove out of the hotel gate automatically, took a large lap around the real streets, took thirty to forty minutes in total, reached a top speed of forty to fifty kilometers per hour, and then drove back. Wang Xinyu himself sat in the driver's seat as the safety officer—at that time, the steering wheel of the modified car was locked tight once it entered the intelligent driving state, impossible to turn; one had to press a button to disengage before taking over, making takeover much more difficult than it is today.
The worst-performing team in that competition was the Baidu camp—the old colleagues who did Robotaxi domestically didn't participate; it was the North American Apollo team that came, scraping the curbs.
In the second round of MMD in 2019, Volkswagen proposed a much more conservative topic: RHP (Rush Hour Pilot)—only testing highways and congested roads, believing this was the consumer pain point. After completing the designated highway route, the Huawei team drove a large section of urban areas, specifically picking difficult sections like construction zones, and took first place again.
Based on these two events, the Car BU was officially established at the end of 2019. Su Qing took charge of ADS, Chen Yilun oversaw perception, and Wang Xinyu oversaw planning and control. The three formed the Iron Triangle.
The team rapidly expanded from over a hundred to four or five hundred people, with over two hundred in planning and control and over two hundred in perception. Interestingly, Wang Xinyu was still at Noah's Ark at this time, managing the planning and control teams on three sides across two organizations—Noah's Ark, HiSilicon, and IOV—forming an awkward dual-reporting state: technically reporting to Su Qing, while administration and promotions were managed by Zhao Qian. It wasn't until the end of 2020 that Zhao Qian came up with a way to have Wang Xinyu cut into the Car BU as an "internal visiting scholar," saving face for both sides.
From Sampling to Optimization: The Triple Jump of Planning and Control
Before talking about mapless, it is necessary to clarify what happened to the craft of planning and control between 2019 and 2021. Because the success of mapless is half due to perception and half due to the evolution of planning and control itself.
The architecture in 2019 was a typical Robotaxi architecture, highly similar to what Baidu Apollo open-sourced:
On the perception side, it was PV (Perspective View) visual perception plus laser perception, followed by post-fusion, and then prediction;
On the planning and control side, it was sampling base—doing policy sampling in the topological space, calculating costs, and then using MPC for control.
This architecture was similar across the board; the real gap was in planning and control: "Specifically, how to sample in what policy space, how to define the coordinate system, and how to make behavioral decisions and interactive decisions respectively vary among companies."
In 2020, planning and control completed its first leap: switching from sampling base to optimize base—from "picking one from limited candidates" to "solving an optimization problem in a continuous space." Interactive game theory was also introduced: no longer just planning the ego vehicle's trajectory, but incorporating surrounding vehicles for joint planning, sampling the future policy space for other vehicles, predicting their behavior, and solving it jointly with the ego vehicle's planning. This was the true differentiation of Huawei's planning and control, being the first to propose interactive game theory.
At the end of 2020, the team used an Arcfox in Shanghai Pudong to shoot a "one-take" demo. The car weaved through food delivery riders on crowded streets, and the video trended on hot searches for several days. This was the first time the outside world knew to what extent Huawei's intelligent driving had reached—before that, only Audi and Volkswagen knew.
What was shown to the world at the 2021 Shanghai Auto Show was exactly this: strong rules, heavy planning and control, and reliance on HD maps. Standing at that point in time, planning and control had won.
It was also in this year that Su Qing left.
After Su Qing left, a group of perception backbone members left one after another. The new management quickly took over. They fully delegated specific technical routes, and many later described the actual boundaries of power like this—leaders set the tone, engineers set the route.
The Catfish Activates, All Hands on Deck for the Mapless Battle
In the spring of 2022, the backbone of ADS was combined to report to the new leadership.
The question at the meeting was not a technical one. Yu Chengdong was pushing the Smart Selection car business at the time, discussing which intelligent driving solution to use for the first version of the AITO M5.
The Bosch solution had already been preliminarily finalized, but that didn't even count as high-level intelligent driving; it could basically only do highway navigation and was almost immovable in urban areas. Yu Chengdong wanted to upgrade to high-level intelligent driving, believing it was the core competitiveness. He initially approached DeepRoute.ai—DeepRoute had always claimed "mapless" externally, with a first-mover reputation in multi-sensor fusion perception and low-cost inference engines, launching a mass-production front-loading solution with a cost of less than 20,000 RMB in 2021. Yu Chengdong tested their car in Shenzhen and thought it ran well, planning to use it to replace Bosch.
But Yu Chengdong still wanted to try if ADS could actually deliver. He had experts evaluate it at the meeting, and the conclusion was: we can do mapless too, and we can do it better than DeepRoute.
Once this sentence was spoken, it became a military pledge.
The window was only one year—by the end of 2022, ADS had to face DeepRoute in a head-to-head PK in front of Yu Chengdong. In the first half of 2023, mass production had to be finalized, and in the second half, it had to be put on cars. If they couldn't do it, ADS's position in the Smart Selection car system would be gone.
Some people remembered that Old Yu often drove DeepRoute's car to work that year; when the ADS people saw it, they all held in a breath of anger.
In the middle of that year, Wang Xinyu's situation was very special. His title changed from "Chief Expert of Planning and Control" to "Chief of ADS Business," managing both the planning and control and perception large groups. This was a state that shouldn't normally appear in Huawei's organizational habits—department chiefs usually only press one line. But this temporary, fire-fighting authorization allowed him to directly step onto the field, taking two large groups that were originally building their own walls, dismantling and reassembling them according to the needs of a single campaign.
If it were just internal KPIs, the pressure wouldn't be that great.
On September 30, 2022, Tesla held AI Day in California, revealing two things: the Occupancy Network, and a neural network that could infer lane topology relying only on standard navigation maps. Elon Musk's meaning was very clear—he was done with the HD map route; pure vehicle-end perception could solve urban areas.
Almost at the same time, XPeng started talking externally about removing HD maps, promoting "light maps" as the main selling point for the next-generation solution.
Tesla, XPeng, DeepRoute, and Huawei—three opponents with different directions pointed to the same conclusion: the mapless route is correct, and no one is willing to wait for Huawei.
Actually, there was originally a more stable route inside ADS—the crowdsourced light map done by Chen Tongqing and Qin Tong, internally called RoadCode. It didn't use professional mapping vehicles, letting the mass-production fleet run the roads themselves. The advantage of this route was a smooth transition; the disadvantage was that it still relied on the freshness of the map—if the map was bad, the car wouldn't dare to drive. After AI Day, the judgment within the team began to change: "Should we be more aggressive, not relying on HD or RoadCode at all, or only relying on them for a small number of long and difficult scenarios?"
The final strategy was only six words: explore while mass-producing. RoadCode wasn't cut, but downgraded—the vehicle end continued to build lightweight maps, but the entire software no longer trusted it; NCA could run with or without it, and all basic information came only from AutoNavi's SD navigation map. Without RoadCode, road structure cognition and planning and control had to bear the load, meaning—Chu Jingkai's road structure cognition and Qian Chaojie's behavioral decision-making were mainly responsible for getting it done.
This was a very counter-intuitive decision: actively putting something they had worked hard to build for years into a "dispensable" position. But it later became exactly where Huawei pulled away from XPeng—XPeng's map had flaws and had to send cars to collect, repair, and calibrate, only able to open city by city; Huawei's map was allowed to have flaws because it wasn't relied upon in the first place.
White-Box Mapless, Unique Globally
The reason mapless was so difficult was that all difficulties ultimately converged on the same scenario: intersections.
One had to know how many exits there were, where each lane led, and match it with navigation intentions. Relying purely on vision, when the front of the car just poked into the intersection, it couldn't see everything and didn't know where to exit; the prediction of surrounding vehicles failed simultaneously because even the shape of the road wasn't certain, making it impossible to judge which direction others wanted to turn, and interactive game theory was out of the question.
A single intersection ground perception, maps, prediction, and planning and control into the dirt simultaneously. The final approach had no tricks whatsoever: all hands on deck. Except for the parking group, the entire department went all out to solve intersections.
Wang Xinyu's first step was to tear down walls within planning and control.
Planning and control had over two hundred to three hundred people at that time, cut into four blocks: route planning to Qian Chaojie, decision module to Dai Zhengchen, planning module to Tong Liuzhu, and control module to Zhang Renxie.
This division of labor was very comfortable in the map era—the map gave lane lines, intersection topology, and drivable areas; routing followed navigation, decision-making chose within given lanes, planning drew lines within lane lines, and control followed the lines. Everyone's input was certain. Mapless took away this certainty. Everyone's input became a real-time estimate that would jump, shake, and suddenly disappear. No one could just guard their own block.
The second thing Wang Xinyu did was to pull the perception department in to work together.
Technically speaking, the most labor-saving approach for the mapless solution was to have perception output real-time road network information with the same precision as HD maps, so planning and control wouldn't need to change a single line of code, the interface would remain unchanged, and the schedule would be the shortest. But perception couldn't do this, so how perception and planning and control interfaced became the key issue.
To use an analogy, the road structure cognition line—how many exits an intersection has, how lanes connect, which leads where—belongs to the perception department organizationally, not planning and control. Wang Xinyu pulled this line to work together. This part of the work was static perception, handled by Chu Jingkai.
Chu Jingkai, mentioned above, was an old subordinate Su Qing led at HiSilicon. He did planning and control early on, and later merged under Wang Xinyu's leadership, partnering for several years. Having someone who did planning and control define perception output meant he naturally knew what the downstream needed, knowing which errors planning and control could bear and which they couldn't. This sounds trivial, but it was actually the technical premise for the whole thing to be pushed forward.
One more thing to mention: the earliest developer of road structure cognition was Wen Feng from the 2012 Lab. That was still at Noah's Ark, when Wang Xinyu was responsible for planning and control and joint debugging, and Wen Feng was responsible for road structure cognition. The group Wen Feng led was called the "Spatial Cognition Group" (special cognition), mainly doing cognition of static road structures, with about ten people in the group. Later, Wen Feng didn't cut over to the ADS department. At the ADS 2.0 stage, Wen Feng went to be the director of the Montreal Institute in Canada, staying for three years, and didn't return to China until the end of 2025.
After the wall-tearing and interface issues were determined, it was the research breakthrough within the planning and control department.
This step went deeper than the outside world imagined. It wasn't just adding a layer of filtering to smooth out the jitter, but replacing the entire mathematical foundation: the planning originally done in the topological space was moved to the Euclidean space, completely shifting from policy sampling (sampling base) to multi-modal joint optimization (optimize base)—not picking one from several candidate trajectories, but directly solving for an optimal trajectory in a continuous space, while taking into account the interactive game theory between the ego vehicle and other vehicles. The decoder was also self-developed, and the entire solver was heavily optimized for the computing power characteristics of the 1951 chip.
In other words, a voice inside ADS at the time was: mapless must be tied to end-to-end to be done. Because rule codes couldn't exhaustively enumerate those uncertainties, they had to be handed over to neural networks. Huawei was the only one that didn't believe this consensus, insisting on solving it with a white-box approach.
Note: The 1951 chip was defined by Su Qing around 2015 at HiSilicon Turing. Externally, the chip was never talked about individually, only appearing in the form of a domain controller (MDC)—MDC 610, 610 Pro, and the 810 with dual 1951 chips. From ADS 1.0 to 4.0, a full four generations of products ran on this chip, with only the 1951 Light and VR generations iterated in the middle, but the core remained unchanged. It carried the load for nearly ten years until 5.0 switched to the new generation 1952. Looking back, after the US sanctions, there were no new chips for intelligent driving; it was this chip defined ten years ago that stubbornly supported the entire rising period of Huawei's intelligent driving.
Huawei ADS VS DeepRoute, Two Rounds of PK
In 2022, ADS and DeepRoute had two rounds of head-to-head PK.
The M7 in the middle of the year and the M5 at the end of the year—the internal ADS 2.0 solution advanced rapidly, and both ultimately chose the internal solution.
The most interesting was the latter. That was the real final kick.
At the end of 2022, in Shenzhen, personally accepted by Yu Chengdong. The ADS team rehearsed the routes to be tested many times in advance. What was put on was the prototype of the entire ADS 2.0 system—the BEV dynamic perception led by Huang Qingqiu (how the perception team solved BEV + early fusion is also a very interesting story during the ADS 2.0 period, expanded below) + the mapless solution led by Wang Xinyu. That was the first time ADS officially removed HD maps in a formal setting, and also the final closed loop of the perception side's AI transformation from target perception to fusion perception.
Two cars. Wang Xinyu drove one, picking up Yu Chengdong; Huang Qingqiu and Qian Chaojie drove the other, picking up Wang Yanmin, the head of the Smart Selection car business.
After the demonstration, the conclusion was that the effect was better than DeepRoute's. "After the final PK, at first they said one for each company, but then they PK'd for a while longer and felt we were indeed much better than DeepRoute."
Years later, many people maintained decency in their evaluation of DeepRoute as an opponent: "DeepRoute's technical capability is definitely good. I tried their car in Shenzhen, and the overall feel was at least on par with ADS 1.0 at the time, with some data dimensions even better. They proposed the 'mapless' concept earlier than us, which has to be admitted. But overall strength is definitely still Huawei's; DeepRoute is a good company, but its size can't compare to ours."
Of course, in hindsight, this might have been Yu Chengdong's most effective management move—introducing DeepRoute as a catfish to activate the inside.
"Human nature is like this; if you have PK, everyone's combat power will definitely be stronger." But at the time, no one thought this was just motivation—"They had already signed a designated point with DeepRoute, and DeepRoute was the A point, prioritized; Huawei ADS was the B point, only used if the A point couldn't handle it."
In early 2023, mapless was confirmed internally to have run through. In April 2023, it was unveiled at the Shanghai Auto Show. In September 2023, mapless capabilities were officially delivered and put on cars.
Thus came the slogan at the launch event—Drive anywhere in the country, drive wherever there is a road.
At that time, the industry was comparing "the number of opened cities." Li Auto and XPeng announced every other month that they would open 10 or 100 cities next month. Wang Xinyu had someone check how many township-level administrative regions there were in China, getting the number of over 60,000, and put it into the launch event materials—a table was listed in front, XPeng 50 cities, a certain company 100 cities, Huawei's 60,000 plus. Later, he simply distilled it into ten words: Drive anywhere in the country, drive wherever there is a road—once this sentence came out, it directly pierced the dimension of the number of opened cities.
Looking back years later, the entire ADS led domestic peers by the longest gap of time exactly during that period, leading others by a full year.
The old map system collapsed accordingly. The HD map department in Beijing was dismantled, all two thousand outsourced personnel were laid off, and formal employees were transferred to the map group in the perception department, doing positioning, navigation map adaptation, and crowdsourced mapping. Later, the HD format was converted to the RoadCode format for maintenance, and finally taken offline comprehensively. As for lighter-weight maps, they no longer relied on manual drawing, but were automatically generated by algorithms.
Technology Pushes the Organization, Technology Pushes the Process
This mapless transformation left a foreshadowing at the organizational level: the Prediction Department.
The Prediction Department was originally an independent module under the Perception Department, on the same level as dynamic perception and static perception. The first head was Yang Xuguang. After Yang Xuguang left, Ding Wenchao took over for a short period, and then it was changed to Tai Lei.
Mapless directly lifted the difficulty of prediction to another order of magnitude. Predicting where others want to go requires knowing what the road looks like—with maps, the intersection topology was given, and prediction only needed to allocate probabilities among a few given paths; without maps, even how many exits the ego vehicle had at an intersection had to be estimated on the spot. Prediction became guessing others' intentions in a constantly deforming world. The interactive game theory of the planning and control team—sampling the policy space for surrounding vehicles, predicting their behavior, and solving it jointly with the ego vehicle's trajectory—was exactly stuck in this position.
Later, ADS's response was to move the Prediction Department from the Perception Department to the Planning and Control Department, bringing it closer to the user. This small group of twenty to thirty people eventually disappeared completely in the ADS 4.0 end-to-end era a few years later—all AI capabilities were integrated into the AI Model Development Department, and prediction was no longer an independent module.
In addition, there was the process.
Li Jieru, QA (Quality Assurance) in Huawei's Quality Operations Department. Huawei says "process determines quality," so Huawei's QA doesn't truly manage quality, but manages the process. What he had to do was formulate the R&D process and delivery process, and then watch everyone go through it.
Li Jieru took this to the extreme.
He transferred from the wireless business half a year after the Car BU was established, at the end of 2019. At that time, the Car BU was in an awkward position: doing autonomous driving, taking the road from 0 to 1, the business process simply didn't match the company's standard IPD process. At this stage, the role of QA was very subtle—saying it's important, it can be crucial; saying it's unimportant, it can also be insignificant. QAs with strong capabilities and those without coexisted in this position in the company.
Li Jieru was the former.
Huawei's IPD process, to outsiders, seems like an ironclad block, with hard rules one after another. But Li Jieru was different. He wouldn't just copy and paste the standard process; instead, he would talk to experts one by one, asking carefully: What is this step for? Why is it needed? Can it be cut? What are the consequences of cutting it?
The feeling for many was that those seemingly rigid process regulations in the company, in his hands, didn't have a single one that couldn't be changed. He cut the unreasonable ones, kept the necessary ones, and then took the modified things to align with the company, stamped them, and turned them into a new compliant version.
ADS 1.0 did this, and ADS 2.0 did this too.
But when the new industry went from chaos to order, the process was built from scratch, and everything was settled, they instead felt it wasn't fun anymore. After 2.0 and before 3.0, the Car BU's process was basically settled, and the executors could operate step by step. Li Jieru then turned and left, going to the next new department.
03、Huang Qingqiu: How AI Ate Rules Step by Step
The AI Germination Period of the Perception Department: Using AI to Solve Target Perception
Huang Qingqiu was a key but low-profile role in the ADS 2.0 stage: his main focus was the "modeling" of perception—that is, using deep learning instead of traditional hand-written rules. However, before expanding on his story, we must first return to the starting point and look at the initial signs of using AI in the perception department.
That was the long dormancy that began in the ADS 1.0 period. Understanding that turning point from white-box rules to black-box neural networks is the only way to truly understand the weight of the ADS 2.0 perception transformation.
The story starts with Xu Bin, the former head of Huawei's target perception, and Chen Yilun's minus-one after the Car BU was established in 2019.
In November 2018, he was recruited into the Internet of Vehicles Lab by Chen Qi, doing perception. At that time, ADS perception was completely rule-based—following Mobileye's path, relying on humans to write rules to solve one corner case after another. The result was that the code became more and more complex, and the problems were less and less thoroughly solved.
The "perception" defaulted by the industry at the time was dynamic target perception (identifying dynamic obstacles like cars and people), because everyone relied on HD maps. The static environment didn't need the car to "see" itself; the map had already drawn the lane lines, curbs, and traffic lights.
In early 2020, Xu Bin's line began to try using neural networks for target perception. By the end of the year, the rule-based approach was completely overthrown and turned into NN (Neural Network). Target perception became the first module in the entire ADS to embrace neural networks.
The significance of this wouldn't be seen clearly until many years later. Once target perception ran through, the next path naturally emerged: NN-ify the remaining modules one by one.
At that time, three lines were being pushed in parallel—Xu Bin was responsible for target perception, Meng Zhaoshi was responsible for the NN-ification of static perception, and Yang Xuguang was responsible for the NN-ification of prediction.
Using their words at the time, the most core parts of autonomous driving were OD, LD, and PD—target perception, environment perception, and vehicle control. NN-ifying all three and stuffing them into one network is what everyone is talking about today as end-to-end. And Huawei's later external promotion of "mapless, drive anywhere in the country" was essentially NN-ifying static perception; the subsequent promotion of end-to-end was essentially NN-ifying planning and control. Step by step, it was all that path seen clearly in 2020.
This was the prelude to Huang Qingqiu's later thorough AI transformation. The seeds planted by Xu Bin and his team in 2020 echoed Huang Qingqiu's harvest a few years later, together forming the complete context of perception AI-ification from germination to landing.
However, this line was interrupted in 2021. After Su Qing left, this group of perception people gradually resigned.
Huang Qingqiu, as the spark of the perception department, was retained. He was also the soul figure who later strung together the entire AI-ification of Huawei's intelligent driving. From laser perception, early fusion BEV, GOD, to end-to-end, he led these key AI battles all the way through.
Before the 2021 Shanghai Auto Show, Using AI to Solve Semi-Solid LiDAR
Huang Qingqiu, born in 1994. Bachelor's in Automation from Tsinghua University, studying under Zhao Mingguo, doing humanoid robots at Tsinghua. He was the captain of the Tsinghua University Robot Football Team—Huoshen (Vulcan) Team, and also participated in making an autonomous driving bicycle. Then he went to the Chinese University of Hong Kong for his Ph.D., entering the MMLab, with supervisors Lin Dahua and Tang Xiaoou, doing long video understanding. In his own words, his bachelor's was doing the robot's "cerebellum," and his Ph.D. was doing the "brain."
In August 2020, he entered the ADS dynamic perception group as a Genius Youth of the second batch, doing research on visual perception, such as turn signal perception based on temporal networks, but it was more exploratory. And his mentor was exactly Chen Yilun.
The real turning point was in March 2021.
The Shanghai Auto Show was less than a month away. Previously, the POC stage used rotating mechanical LiDARs—360-degree scanning, extremely high point cloud quality, 100,000 RMB each; what was to be put on the show was RoboSense's semi-solid solution—the field of view was only 25 degrees vertically and 120 degrees horizontally, and the point cloud quality was much worse, but the price dropped by more than an order of magnitude.
This was essentially a cost-reduction move, but a disaster for algorithms. The fusion perception group couldn't handle it. Chen Yilun had Huang Qingqiu lead the team to tackle it. "I didn't know a single person over there, and I hadn't done LiDAR before."
What he did was change the entire LiDAR perception from rules to neural networks—taking the neural network experience accumulated from doing visual perception and moving it entirely to point clouds. At that time, there were actually two sets of solutions in the group pre-researching AI-ification, but neither was very good; after he went over, it was equivalent to integrating the two solutions into a complete AI-ified solution, even making LiDAR perception the most credible module in the entire perception system for a time.
Solved in a month. He stayed in the fusion perception group, and the team under him expanded rapidly—from six or seven people when he joined the group, to 50 people at the end of ADS 1.0, spanning Shanghai, Beijing, and Shenzhen.
The perception architecture at that time was modular: visual perception, LiDAR perception, and millimeter-wave perception each output results, connected to post-fusion, then to the prediction module, and finally to planning and control. Besides dynamic target perception, there was another set of static road perception with a similar structure—first through a network, then fusion, then planning and control. The entire design basically referenced Baidu Apollo's architecture: perception, fusion, prediction, planning and control.
The biggest problem with this architecture was that each module had to first "abstract" the physical world into structured information understandable by humans—how many meters the distance to the front car is, length and width, speed—and then pass it down. And abstraction inevitably loses details: you can exhaustively enumerate distance and speed, but you can't exhaustively enumerate turn signals, brake lights, wheel angles, whether there are people in the car, and other long-tail information.
The Thorough AI-ification of the Perception Department: Early Fusion + BEV
From after the April 2021 Shanghai Auto Show to early 2023, Huang Qingqiu was always doing the same thing: tearing down this wall layer by layer.
First, LiDAR plus millimeter-wave, making a fusion module called "RL perception," and then pulling vision in as well. At the same time, Huang Qingqiu attracted a key figure, Genius Youth Zhu Xinge—Huang Qingqiu's junior schoolmate at CUHK MMLab. This fusion architecture was exactly from the hands of the two of them.
Thus, the point cloud of LiDAR, the images of cameras, and the signals of millimeter-wave radar were stirred together in advance in the neural network, uniformly projected onto a "bird's-eye view from God's perspective," and the model learned by itself "what is in front, where it is, and whether it can be hit," instead of three sensors each speaking for themselves and then relying on manual rules to vote.
This is the "early fusion BEV perception network," and Huawei was the first in the industry to bring it to mass production.
It needs to be pointed out that early fusion is actually a unified interface for downstream—for planning and control, whether the upstream is early fusion or post-fusion, what it gets is a unified perception result, with no difference. What will truly revolutionize planning and control is end-to-end, which directly strings perception and planning and control into one network later, which will happen a few years later.
Under this architecture, the value of LiDAR was truly released for the first time.
The most difficult problem with LiDAR is noise—water splashes in the rain will generate a large number of LiDAR points; how to distinguish whether this point is a water splash or a real obstacle? The core technology of this "LiDAR denoising" was developed between 1.0 and 2.0, optimized in every generation since, and has never been talked about externally.
But later, many competitors chose "LiDAR on and then off"—they found that LiDAR not only didn't play the expected role but dragged down the overall performance, and had to be forced to remove it. Huawei's difference lies in being able to fully leverage the advantages of LiDAR while effectively avoiding and solving its shortcomings.
Here also hides a thing almost unknown to the outside world: Huang Qingqiu simultaneously took on another role—defining specifications for the sensor department AIS.
The "sparse top, dense middle, sparse bottom" arrangement pattern of Huawei's LiDAR, and the product definition of Lidar + Camera pixel-level alignment, were all proposed by Huang Qingqiu from the perspective of perception problems and completed jointly with AIS experts. This is exactly the result of Huawei's software and hardware vertical integration. Those who buy hardware externally cannot achieve this level.
In 2022, ADS 2.0 began to switch to Huawei's self-developed LiDAR. The principle changed completely—RoboSense's was a MEMS micro-mirror, while Huawei's self-developed one was a line-scanning 2D scan. The change in hardware principle caused the noise pattern to change as well. The team did a lot of data collection and re-optimization for this; fortunately, the architecture was designed to be general, and the large neural network architecture didn't need to change.
This architecture also conveniently solved a more realistic business problem.
Huawei's cooperative models were increasing—Arcfox, Avatr, AITO, and later Luxeed. If every time a car was changed or a sensor layout was changed, the algorithm had to be re-tuned, the team simply couldn't keep up. The commercial requirements of the Harmony Intelligent Mobility (HIMA) system for ADS were actually only two: platformization, regardless of whether the car is tall, short, fat, or thin, it must be solved with one neural network; and one track, all cars share one set of algorithms. Huang Qingqiu's architecture was exactly the foundation of "one track" at the algorithm level (the "one track" at the software level was mainly promoted by Wang Guanjun, and later Rong Pengqun also made good results as a software expert).
In 2022, this architecture won him one of the most valuable individual awards within Huawei—the "Outstanding Individual Award." According to internal standards, gold individuals are about 1%, while outstanding individuals are one in several thousand to ten thousand; the entire ADS usually only has one person per year.
Here, another person must be mentioned: Li Ya.
Came from Momenta, now a Level 20 expert, and one of the four people who have won the highest-level gold individual award in ADS.
The AI-ification of the entire visual perception was led by him. BEV has a natural limitation—it is good at expressing macro elements like roads and vehicles, but not good at expressing fine-grained information like traffic lights. Therefore, a large amount of perception work in the visual domain still needs to be completed directly in the image domain, which has always been deeply cultivated by Li Ya. By the 2.0 and 3.0 stages, he pushed visual perception to the extreme: identifying small obstacles on the road 200 meters away, judging whether there are hard objects in a plastic bag and whether it can be run over—these extremely fragmented corner case attributes are not suitable for BEV or GOD, and all rely on his visual perception module to bottom out.
"Did a lot of dirty and tiring work, and is also a core contributor to many key safety issues." Huang Qingqiu mentioned Li Ya twice.
In addition, there is another person, also heavily mentioned by Huang Qingqiu: Jiang Xuesong.
The background is this: the BEV early fusion for dynamic perception had just run through, and Zhu Xinge pulled vision into the same network. The two lines briefly intersected in 2022—on the day of the PK with DeepRoute, Wang Xinyu's mapless planning and control and Huang Qingqiu's BEV perception, the two core algorithms formed a complete ADS 2.0 system, installed on two cars.
Fusion was originally the core hub of perception. In the old architecture, vision, LiDAR, and millimeter-wave each output results, handed over to the post-fusion module to be pieced together into a unified target list using rules, and then passed down. Early fusion judged the winner of this dark line. The intermediate features of the three sensors were directly sent into the same neural network, and the network fused them itself. What was post-fusion supposed to do—bottom out.
After BEV landed, the post-fusion module wasn't deleted but was kept to specifically handle extreme situations—Jiang Xuesong had been responsible for this since ADS 1.0 and had been guarding it ever since. The significance of this module's existence was only one: active safety's tolerance for false triggers is close to zero. A million cars running for a day cannot false-brake even once. Neural networks cannot give this kind of guarantee, so there must be a layer of human-written, auditable logic pressed at the very bottom.
A former technical backbone ultimately settled as a safety redundancy. In the context of intelligent driving, this is actually a very high evaluation. This is the role transformation of an individual in technological iteration, and also the rational choice of "guaranteeing the lower limit" over "rushing the upper limit" during engineering landing.
The Deeper Ambush of the Perception Department: Shenzhen, a Data Pre-research Team Built from Scratch
In the second half of 2021, Huang Qingqiu made a decision that didn't seem important at the time but later determined the direction of ADS 3.0 and even 4.0: building a new team from 0 to 1 in Shenzhen.
At that time, ADS only had two R&D bases in Shanghai and Beijing, while Shenzhen was talent-dense, and other departments had already set up points. Huang Qingqiu himself was based in Shenzhen and traveled to Shanghai for business for a long time. Coincidentally, ADS decided to layout in Shenzhen, so he took the opportunity to propose establishing an algorithm research team.
In July, the first person joined; in October, the fifth person joined. One could feel that Huang Qingqiu was very cautious in picking this team. The office was in Longgang Yungu, "inviting guests for a meal and finding a large table couldn't even be filled, one side was completely empty."
This team was positioned as a pre-research line from the beginning—"It's not appropriate for a new team to directly take on core business, so start with various pre-research projects." But it was exactly from this pre-research that the two most important people for ADS 3.0 later grew:
Su Peng, a Genius Youth recruited by Huang Qingqiu, one of the first ten employees of the Shenzhen team. In the 2.0 stage, he led the team in pre-researching a pure vision solution—verifying the feasibility of the pure vision route under a high-computing-power platform. In the 2.0 era, the company decided not to push pure vision for the time being, and this line was shelved; but it didn't die. It was later reactivated and became the technical source of ADS SE (that is, the ADS Light pure vision downgraded version). After being shelved, Su Peng turned to the R&D of the GOD network—General Obstacle Detection. The problem this network had to solve was that the original moving target detection only recognized cars and people, while dropped sacks, broken tires, fallen water barriers, and rolling stones that often appeared on highways were completely unrecognized by the system. GOD found a universal obstacle expression method, allowing everything that might be hit to be uniformly expressed in the same space.
Guo Wen, one of the first three employees of the Shenzhen team, originally did GT Factory—an automated data production reconstruction chain. Initially a supporting facility for GOD, it later expanded into a data foundation serving the entire perception system. He later grew into the Minister of the Data Development Department.
(If you experienced the ADS 3.0 period, welcome to add the author Wu Tong's WeChat: lI__O0o)
In the second half of 2022, GT Factory transitioned from a pre-research state to a mass-production-oriented rhythm. The changes it brought were fundamental: before this, data production capacity equaled manpower—"however many people, however much data"; after automation, this shackle was broken. This system iterated through dozens of small versions and over a dozen major versions, reaching a state where it could automatically produce in batches by early 2023.
At this time, Tesla had not yet revealed any details about the data closed loop. "Although the overall sense of each company's solution would be similar, the details varied greatly, and no one would make it public."
Wang Xiruo participated in the entire process of the data system construction and had a very simple summary of what kind of people this work needed: first, to be meticulous—not nitpicking formats, but having a persistent desire to get to the bottom of every kind of data or modality. Passing over a slight anomaly would bury many problems; second, to open up one's mind—there is no rigid, rule-following way to solve problems.
A Methodology Grown in the Crevices of Huawei's Process
If looking only at technology, the success of Huang Qingqiu's line can be attributed to a victory in architectural selection. But almost everyone who has worked with him talks about things other than technology first when talking about him.
Huawei's organizational tradition is to split management and technology into two people: the Minister (TM) manages people and delivery; the Technical Leader (TL) manages the technical solution. Huang Qingqiu was a rare exception: he was both the Minister and the Technical Leader.
Those familiar with him evaluated him more directly: "He took the manager position as a technical expert. This is actually very rare in Huawei, because Huawei's manager is a pure management position, doesn't need to understand technology too much, but needs to understand delivery better." Someone also mentioned a detail: the day Huang Qingqiu became Minister, the team's private discussion concluded—he was the kind of person "everyone felt was suitable for this position, and didn't even resent how fast he could be promoted."
There were a few un-Huawei things in his management style.
The first was information transparency. He would take the initiative to report to the team—not talking about papers, but product planning, business planning, what cross-departments were doing, and what future plans were. "Many pieces of information are completely inaccessible without contact or reaching this level, so explaining everything to everyone makes it feel relatively transparent and reliable."
The second was the alternation of pre-research and mass-production lines. "Doing mass production is actually quite hard; doing it all the time might make everyone a bit tired, so this is actually an alternation. Do a period of landing, keeping feet on the ground; then do a period of reaching for the top. Although early fusion was only produced in 2.0, he was already pre-researching it in 1.0; although end-to-end was only produced in 4.0, it was already being pre-researched in advance in 3.0."
The third was bypassing the process. "If you completely follow the IPD process—project initiation, persuading leaders, dealing with doubts—half a year is gone. Many times, you have to quietly build the prototype yourself first, and then push it with the results." When he did LiDAR fusion, he took the opportunity to verify the feasibility of visual fusion under this name, running it through before initiating the project. This style is not very Huawei, but it is worth the organization to learn from and preserve.
The fourth was planting people a year in advance. "I want to do this thing and need this kind of person, but it doesn't mean I have to let him do this full-time on his first day of employment. I will recruit people first, put them in related directions, and when the business timing is mature, he has already run a section of the road internally." Zhu Xinge was recruited for early fusion; Su Peng was recruited to do pure vision pre-research.
Huang Qingqiu's own summary of the most training aspects of this job had two points:
One is the ability to grasp the focus, "being able to tell at a glance which things are valuable and which are noise";
The second is the awareness of team cooperation—"I know very well what everyone is good at, and know how to cover for them."
Leiphone's "Zuolin Youli" public account will release "The Huang Qingqiu I Know." Friends are welcome to add the author Wu Tong's WeChat in advance, WeChat ID: lI__O0o.
Many Genius Youths actually can't do this, because most of them think they are technical geniuses and have a strong sense of individual heroism. But Huang Qingqiu is not; many times he is doing miscellaneous work and setting up the stage. Later, the AI department managed by Huang Qingqiu included more than 10 Level 19 and above technical experts, and 7 Genius Youths—at least half of whom were dug up by himself, such as Zhu Xinge, Su Peng, and Wu Minyan, all Genius Youths recruited by him through various channels.
Interestingly, it is Huawei's way of using these people. The outside world thinks that half the reason Genius Youths are dug in with high salaries is for academic assets—who the supervisor is, where the lab is, how broad the network of schoolmates is. But what truly makes them generate value is precisely forcing them to shift from "very high quality but hard to land" things to things that can be delivered. Huang Qingqiu himself is the most typical sample.
04、Hu Wenxiao: Letting Users Understand the Black Box
Don't Make Users Learn Your Technical Architecture
The technology is done, but there is the last mile: how does the user know what this car is thinking.
This mile was managed by Hu Wenxiao. He previously responsible for user experience at internet companies and NIO, and was a rare "ToC lineage" in ADS. Hu Wenxiao's addition filled the key link of "translating technical capabilities to ToC"—ADS had no ToC, all were senior experts, tech guys, and professionals. "His first feeling after going was the disenchantment of Ph.D.s: 'I thought a Ph.D. in my department at NIO was amazing, but ADS is full of Ph.D.s everywhere.'"
In 2021, Hu Wenxiao had only 11 people under him; by the end of 2022, it grew to over twenty, nearly thirty, a large part of which was outsourced manpower—because there were more projects and they had to pave the way for the future.
In the Su Qing era, all aesthetics of HMI were decided by Su Qing alone. "He presented it however he wanted." Su Qing had an obsession with Tesla: when cooperating with BAIC, the car originally didn't have a double-click stalk design like Tesla. Su Qing believed Tesla's solution was safer and stubbornly required the other party to add a stalk. But Hu Wenxiao leaned more towards physical buttons at the time—he had done human factors evaluation at NIO, and Huawei ultimately used the buttons on the left side of the steering wheel—but personal opinions didn't matter at that stage.
After Li Wenguang took office, this logic changed. The new leader didn't do this direction before and wasn't familiar with the business after taking over, so he adopted a completely different strategy: let professionals do professional things, fully authorizing experts in various fields.
"For me, it was a golden period. I didn't win any awards before 1.0, and innovations in interaction were hard to reflect in the UI visual level; during 2.0, he actively cooperated with surrounding CBG, UCD, Human-Computer Interaction Lab, and other teams, winning four product design awards including Red Dot and iF in 2 years. It also laid the foundation for the brand-recognizable ADS."
The core experience concept Hu Wenxiao pushed in the 2.0 stage was called "integrated driving and parking."
Before 2021, due to insufficient capabilities, the industry was used to selling functions broken down by technical modules: ACC adaptive cruise, ICA intelligent cruise assist, NCA navigation cruise assist, AVP automated valet parking, EPA auto parking... a bunch of abbreviations. Users getting the car had to first learn this glossary to know what functions they could use on what road sections.
"The experience was very poor. What I have always pushed from the user experience side is 'integrated driving and parking'—no distinction of scenarios at all, whether it's highway, urban, park, or underground garage."
This judgment was advanced at the time and later became the industry standard practice. In the 2.0 stage, "integrated driving and parking" was not just a concept; it had solid technical support: on the parking side, AVP (automated valet parking) was added, connecting from parks to urban areas, from parking spaces to parking spaces; cooperating with the slogan "what you see is what you can park"—you can bookmark your frequently visited parking space, and the car can drive intelligently all the way to that parking space.
The most troublesome blind spot Hu Wenxiao remembered was a twelve-story three-dimensional parking building in Shanghai. "What if the parking space you bookmarked is occupied by someone else? You have to find a new parking space—the intelligent driving can't exit, right?" The system must be able to recognize that the parking space is occupied, and then roam again to find another one. The three segments of urban NCA, park roaming parking, and APA auto parking were completely strung together, which was the most important layout of ADS 2.0 at the experience level.
An Emerging Species: The Joint Association
For the organization, Hu Wenxiao had another job: the Experience Joint Association.
Huawei's process is complex and rigorous. Any demand has to go through a very long process, round after round of reviews, and making a small modification is very troublesome—this IPD process was designed for hardware delivery and is not suitable for rapidly iterating user products.
Hu Wenxiao made a suggestion to Li Wenguang: set up an "Experience Joint Association," driven by user experience, with people from planning and control and perception coming to evaluate user-side things—should we do it? Should we change it? Is it a problem?
"User-side things are different from technical-side things—dynamic, rapidly changing, requiring rapid decision-making, rapid trying, rapid feedback, and rapid modification."
This was the thing he considered the most successful in the 2.0 stage. The cost was also clear: this meeting relied on his personal driving force.
But these weren't the most tiring moments. "The most tiring period in the 2.0 stage happened in Chongqing during the 2022 pandemic."
That was the E12 project—Changan's car. The previous E11 had good quality but average sales, and Changan had a lot of opinions about Huawei. To push the project forward and change E12's original cockpit solution to Huawei's, Hu Wenxiao and a few others flew all the way from pandemic-locked Shanghai to Chongqing.
"It's equivalent to me having to represent Huawei with the strength to convince others." He had to report to all of Changan's executives, explaining the benefits and advantages of the solution, as well as the reduction in their equipment costs—"The project required using the least modification cost to make others accept it, but also making others feel it was very beneficial."
He got sick immediately after finishing the talk. On the third day, when they were preparing to go to Seres, he tested positive and flew back urgently for isolation.
This was 2022, a year later simplified into "the ADS 2.0 R&D period." In the memories of those who experienced it, it was a year of flying business trips, locked-down airports, and a body that collapsed after giving a presentation. Talking about it now, it seems very far away.
05、Yu Chengdong: Translating Technology into Sales
The product is packaged, and another person took the stage.
A Car BU CEO Who Drives Himself to Work Every Day
The outside world's impression of Yu Chengdong is usually "Huawei's strongest salesperson" and "far ahead." But inside ADS, engineers' evaluation of him was unexpectedly positive, and for very specific reasons.
"When a leader reaches this level, being reliable and not making mistakes is very important; it's very possible to make very low-level mistakes. For example, with intelligent driving, Old Yu drives it every day, so he knows exactly what the situation is with intelligent driving, and the people below can't fool him."
Yu Chengdong's car has no driver; he drives himself to and from work every day; his phone always uses the next-generation prototype, kept in a case and used daily as the main phone for several months. He has a problem feedback group where he posts problems every day—where did it brake slightly today, hurry up and reply what's going on.
"This ensures that command won't be blind command, and the cows blown at launch events won't be blindly blown. He has confidence in his heart because he has tried it and knows."
His way of making demands was also very typical: not starting from technical feasibility, but deducing backwards from commercial and product selling points. Mapless came this way, the catfish came this way, and safety came this way too—the technical team actually hesitated at the time. No matter how strong intelligent driving is, it can't have zero accidents. Would loudly shouting about safety backfire? But the market proved that high-end car owners value their lives, and the conversion effect of the safety label was extremely good.
"His judgment path is 'what consumers care about -> how to sell the product -> deduce backwards what technology must break through.' This point is an extremely effective external pressure for the technical team."
April 16, 2023, The Ten Words Remembered
The launch event on April 16, 2023, was the moment when the four lines met on the same stage.
Wang Xinyu's mapless solution became the "not relying on HD maps" and "drive anywhere in the country, drive wherever there is a road" city-opening roadmap on stage. From that moment on, the industry game rule of "number of opened cities" that had lasted for two years was pierced.
Huang Qingqiu's BEV early fusion became the multi-sensor fusion perception in Yu Chengdong's mouth—"LiDAR, millimeter-wave radar, cameras, irregular obstacles can also be recognized."
The 3D rendering videos made by Hu Wenxiao's team played on a loop on the big screen.
That day also saw the first external release of the GOD network—General Obstacle Detection: rollover vehicles, dropped large cardboard boxes, falling rocks, fallen large trees, all can be recognized and decelerated to a stop. Strictly speaking, this was an advanced release: GOD truly matured and was large-scale commercially used a year later in ADS 3.0; at the 2.0 launch event, it was more like a preview.
New M7: Technology Cashed in as Sales
What truly turned this technology into money was the new AITO M7.
When the M7 first launched in July 2022, its performance was flat; by late August and September of that year, as Huawei's intelligent driving solution truly got on the car, sales began to pick up. In October, it sold 5,226 units in a single month, its highest record at the time. Li Xiang later publicly admitted: "The launch and operation of the AITO M7 directly crippled the Li Auto ONE. We have never encountered such a strong opponent, and for a long time, we were completely powerless to fight back."
But the real explosion came a year later. In September 2023, Huawei invested heavily 500 million RMB to upgrade the new M7. Yu Chengdong used three "ceilings" in a row—ceiling of intelligent cockpit, ceiling of intelligent driving, ceiling of intelligent safety. Orders for the new M7 broke 15,000 units in less than 24 hours after launch; from September to December 2023, monthly sales were 5,248, 12,193, 17,039, and 25,545 units respectively; by January 2024, AITO delivered 32,973 units in total, becoming the monthly sales champion of new force brands in the Chinese market for the first time.
This wave of explosion coincided exactly with the time when ADS 2.0's mapless capability went from a demo to a nationwide rollout.
One detail was repeatedly mentioned: the new M7's first true breakout relied not on the launch event, but on short videos spontaneously uploaded by a large number of users—when encountering pedestrians or vehicles suddenly cutting in, the car automatically braked to a stop, saving a life. These videos went viral online, becoming an important driver of the sales explosion. As for whether this was the market department going with the flow or Yu Chengdong's actively planned communication strategy, even insiders couldn't tell clearly, "but the direction was set by him."
Later, someone evaluated this business reversal very directly: "The same ADS, couldn't sell in volume on Arcfox and Avatr, but sold in volume on AITO."
Arcfox was too expensive and ugly, Avatr was too niche and didn't even have storage space; the new AITO M7 was a large five-seater and seven-seater, practical, suitable for families, plus Yu Chengdong's personal IP and traffic, and it immediately took off.
Before September 2023, the Smart Selection car (within Huawei's terminal system) was almost dead, and ADS had been losing money. After this, ADS saved the Smart Selection car, and the Smart Selection car also saved ADS.
Regarding the attribution of this victory, many people had the same sigh:
"The success of ADS and Smart Selection cars, Old Yu's personal IP accounts for half, and our efforts account for half."
"The three most correct decisions Yu made in the ADS 2.0 era: cutting HD maps, introducing DeepRoute for PK, and the Smart Selection car new M7."
"Technological breakthroughs are the foundation, but Yu Chengdong translated 'mapless drive anywhere in the country' and 'active safety saves a life' into a language consumers understand, turning technical assets into sales assets."
After the cars were sold, the flywheel started to turn. Millions of real road condition data began to be sent back. During the ADS 2.0 period, it was still a small-scale data closed loop. After the new M7 picked up volume, it was invested in as a core system engineering project—this is the story of ADS 3.0.
06、Epilogue: The Year of Unprecedented First Place
2021 to 2023, these two years, Huawei, previously considered slightly inferior in mass-production capability, completed a positional reversal.
In the ADS 1.0 era, the industry generally recognized XPeng as doing the best; by ADS 2.0, perception was fully AI-ified, multi-sensor fusion was completed, HD maps were completely left behind, and Huawei stood at the front of the first tier of domestic intelligent driving for the first time, and was the first to shout "drive anywhere in the country," leading peers by a full year.
This reversal didn't have a single hero. It was the result of Wang Xinyu's white-box breakthrough in mapless, Huang Qingqiu's early fusion attack on the perception side and the quietly built foundation of the data base, Hu Wenxiao's brand recognition established at the experience level, superimposed on Yu Chengdong's last-minute change of generals against DeepRoute at the commercial level, all cashed in together.
This battle also belonged to those who rarely had their names publicized online: Chu Jingkai, Qian Chaojie, Zhu Xinge, Su Peng, Guo Wen, Jiang Xuesong, Zhang Renxie, Xu Chunjing, Rong Pengqun, Long Xuan, Li Wenguang, Zhao Qian... History writing is always simplified to a few names, but those who truly turned technology from PPT to mass production were this entire layer of people who almost never accepted interviews.
Using the words of one of them, "I am very fortunate to participate in this revolution at this age, rather than when I'm a bit older."
That was an era when the technological paradigm hadn't converged, and everyone could leave their fingerprints in it; a group of people with an average age of around thirty, using overtime, arguments, and all-nighters, pushed a backward machine to the forefront.
That was their golden age, and also Huawei's intelligent driving golden age. It was brief, noisy, not without regrets, but as real as a car that has run on real roads—all deductions in the laboratory are illusions, and ultimately must be driven out solidly.
At the end of almost every interview, Leiphone would ask the same question: Is this a personal success or a platform success?
The answers were all consistent: Both.
The platform provided hundreds of billions in computing power budgets annually, matrix management, and IPD processes, stuffing a group of smart brains into the same version plan, able to operate on its own without relying on a single genius. But precisely at those critical nodes—the decisive turn of perception AI-ification, the first run-through of the end-to-end prototype, the final bottom-line guarantee of planning and control, and the decisive move of the new M7 forging ADS into a consumer-grade IP—it was specific people who pushed that crucial handle.
Huawei's greatest place lies precisely here. It is not a hall that advocates individual heroism, but can aggregate the wisdom of a group through a set of processes into a thrilling victory.
Preview of the Next Article
In the ADS 3.0 era, perception is gathered into a large GOD network, and decision planning is pressed into a PDP network. But taking one more step forward, the extreme approach is to directly string the two together. Perception and planning and control merge, and the entire system collapses into a neural network. From the raw signals of the sensors in, to the trajectory the car will take in the next second or two, there are no human-written rules in the middle, only a black box that directly controls the car.
That is the story of ADS 4.0.
That revolution fulfilled many people, buried many people, and finally leveled everyone. The ultimate of genius is to personally execute oneself. It sent Huawei's intelligent driving to the undisputed first place domestically, and also let everyone see the technological ceiling clearly for the first time.
All this was almost inevitable. When the slope of technological evolution slows down, the best batch of people doing intelligent driving in China, with the experience of mass production, the run-through data closed loop, and a temperament disciplined by Huawei yet unwilling to be disciplined, flowed to the next battlefield being ignited.
The evolution of technology is never controlled by people. One generation has its own mission.
Welcome to add the author Wu Tong's WeChat: lI__O0o
* Please note self-introduction (company, position, topics for in-depth communication)