Introduction
On March 30th, 2022, the official Weibo account of the automaker, Li Auto, released its self-developed intelligent driving system, named Li AD Max. The perception, decision-making, planning, and control software are all developed in-house, with the AEB emergency braking function optimized for Chinese road conditions. The upcoming Li Auto L9 will come equipped with Li AD Max as standard across all models.
China has become the main battlefield for the intelligent automobile industry. By 2021, the penetration rate of L2 level assisted driving technology for passenger cars has reached 22.2%, breaking through the high-level assisted driving stage. Top-tier companies are gradually proving that mass production cars + self-developed technology is the main theme of top competition. Meanwhile, in recent years, China’s emerging forces have already made it to the forefront of the world. This year, Li Auto hopes to take another step forward with the L9.
It is little known that Li Auto’s self-developed system started its research and development just after the Spring Festival of 2021 and was launched in less than a year.
Why was Li Auto able to self-develop so quickly when only starting its endeavor?
This is the sequel to the Li Auto entrepreneurial story following “they created Li ONE!”. It is a story about self-developed assisted driving technology.
This article is 10,500 words long and will take 25 minutes to read.
We Need to Self-develop
In 2021, there was a phone call that Liang Xiangpeng would never forget.
It was on the first day of the lunar year on February 11th when the caller, Li Xiang, who was the VP of Intelligent Driving at Li Auto, asked him over the phone: “Liang, have you thought about self-developing intelligent driving technology yet?“
Liang Xiangpeng didn’t hesitate to answer: “I’ve already thought about it. We need to self-develop completely.“
“Do you have confidence in this?” Li Xiang asked.
“Of course, we have confidence. No problem,” Liang Xiangpeng replied almost without thinking.
“Okay, if you have confidence, let’s do it,” Li Xiang said and hung up.
A few minutes later, Li Xiang announced the decision in an internal WeChat group of partners: “This year, we will newly launch Li ONE with the self-developed intelligent driving system. Everyone should cooperate with Liang Xiangpeng to make this happen.”
In fact, Liang Xiangpeng had already discussed self-developing technology within the team since the end of January. After a thorough inventory, everyone believed that, considering the current company resources, this would be a challenging endeavor due to a lack of manpower, as well as a lack of experience.# 2021 Ideal ONE will change the hardware architecture for ADAS in October 2020. At that time, the team’s idea was to retain the original supplier architecture for basic ADAS functions while adding a new system for navigation-assisted ADAS. The ADAS architecture has two controllers and the communication and switching between the two systems is quite complex. At that time, the company’s ADAS R&D team had less than 100 people, less than half of whom had the ability to code.
Some team members proposed to continue to rely on suppliers, but require the algorithm code to be open, i.e. the so-called “white box mode”. When there is a demand, the company sends people to modify the code at the supplier’s site. This status quo persisted until early 2021.
Lang Xianpeng, an experienced person in the autonomous driving industry, understands that if supplier solutions are adopted, communication and coordination between various processes will certainly consume a lot of time and energy. Moreover, few suppliers can implement higher-level ADAS, and this path will not go far.
Therefore, from the company’s strategic perspective, it is necessary to self-develop ADAS, which has strong significance for both product capabilities and overall efficiency improvement, and is a core capability that Ideal must possess as a truly intelligent automotive company.
On February 1, 2021, Li Xiang, unable to bear it, typed the following in the company’s intelligent driving R&D group:
“I reiterate a clear decision: including basic ADAS, navigation-assisted ADAS, and autonomous driving functions such as visual parking, NOA, and L4, we must self-develop ADAS and not allow any external suppliers. If we really don’t have the ability to develop, I would rather shut down this company. I don’t want to create an electric vehicle company that is both backward and without prospects.”
Lang Xianpeng then said in the group:
“The self-developed project must be capable of developing our own capabilities. Otherwise, I will be the first to resign.”
On the first day of the Lunar New Year, ten days later, with Li Xiang’s announcement in the high-level group that the ADAS of 2021 Ideal ONE will be completely self-developed, Lang Xianpeng and the team officially embarked on the starting line. At this time, with less than four months before the launch of 2021 Ideal ONE, after the decision to completely self-develop, everything under the Mobileye system was reset and everything had to start from the first line of code.
On February 26, 2021, Ideal Motors held a conference to strengthen morale, and again emphasized the importance of self-developed ADAS. The self-developed ADAS includes three projects: basic ADAS, navigation-assisted ADAS, and AEB automatic emergency braking system, and for Ideal Motors, everything starts from scratch.The most difficult thing is not only the development difficulty of the project itself but also the urgent time. The release date of the 2021 IDEAL ONE is May 25th, and the basic auxiliary driving part, including ACC, LCC, automatic parking, etc., needs to be independently researched and on the line before this date.
Backtracking, the latest time for functional road testing is April 15th. Before this, software performance needs half a month of debugging time, so all basic auxiliary driving functions must be on the line before March 30th.
March 30th! When everyone calculated this time, they gasped. From February 26th to March 30th, there is only one month left. Once the project is delayed, the 21 IDEAL ONE won’t be delivered on schedule and the “timed bomb” that Lang Xianpeng has bound to himself will also explode.
Under such conditions, it is imaginable that the development did not go smoothly when it started hastily. All kinds of unknowns and accidents emerged in March, such as one of the four test cars at the time deviating to the left while in lane keeping mode, while the other three cars with the same software version were completely normal. No problem was found in the initial check of all links. The team spent two weeks to solve this strange bug.
By March 29th, the first phase of the functionality was barely completed, and the next day, Liu Liguo, the head of IDEAL’s whole vehicle development, participated in the verification test and then poured cold water on Lang Xianpeng: “Lang Bo, I think with the current status, it will take you at least another four months to finish.”
The team felt pressured upon hearing these words, but Lang Xianpeng showed his unique calmness. He understood Liu Liguo’s “harsh” style very well. Since Liu Liguo did not say that the current status was impossible to achieve, the next step is to find ways to compress four months of work into 55 days.
Time is certainly not enough, but delivery cannot be delayed. This is the ultimate sacrifice. In the remaining 55 days, in order to deliver the basic auxiliary driving function on schedule, the team almost maximized efficiency. Finally, at 10 am on the day of the press conference on May 25th, Lang Xianpeng sent a message in the company’s senior management group: “After being confirmed by the whole vehicle and quality department, the basic auxiliary driving has reached the delivery status, and the press conference can proceed as scheduled tonight.”
So far, the first stage of the self-research project of IDEAL in 2021 has ended, and the first “timed bomb” on Lang Xianpeng has been dismantled, which has relieved some of the pressure within the team. In the atmosphere of the ultimate sacrifice, Lang Xianpeng withstood the pressure and turned his wait for more than three years into a solid first step. However, at this time, Lang Xianpeng did not think that it was time to celebrate. He clearly understood where his tension came from.
Indispensable necessary conditions# A Long-awaited Moment
On January 28, 2021, was a memorable day for Lang Xianpeng, as XPeng announced the official release of its NGP function for navigation-assisted driving. Lang Xianpeng borrowed a car that evening to experience it himself. On the way back home, there were several scenes of ramps, and the interference of the environment’s light at night was not small. However, the NGP function performed stably all the way, and there was no manual takeover during the journey.
XPeng’s excellent performance prompted Lang Xianpeng to reflect. In contrast, the 2020 Ideal ONE model could not achieve NOA function because of system restrictions.
That night, Lang Xianpeng did not sleep at all. Just over 20 days ago, everyone in the company was happy about selling 32,624 cars in 2020. At that time, everyone thought that it would be great if 50,000 were sold in 2021, and it was considered very challenging to set a target of 60,000 cars. However, after analyzing the market situation, everyone realized that if the sales could not reach 100,000 units, it might be a problem for the company to survive.
NOA is the embodiment of an enterprise’s core R&D capability for assisted driving technology. To become a truly intelligent automotive company and win market recognition, NOA must be done and done well. However, at that time, there was no supplier in the industry that could provide this level of assisted driving solution. Therefore, the NOA function was also a “certification label” of the first-tier intelligent automotive companies at that time.
In the case of suppliers unable to achieve advanced assisted driving functions, the OEMs can only rely on themselves to build such systems, which actually verified Lang Xianpeng’s prediction three years ago.
At the end of 2017, Lang Xianpeng met Li Xiang. At that time, Lang Xianpeng was the Director of High-precision Maps and Autonomous Driving Technology at Baidu Intelligent Automotive Business Unit. By chance, Lang Xianpeng and Li Xiang established contact, and they talked at length about topics related to the autonomous driving industry. At that time, they agreed on one point: To do well in autonomous driving, it is necessary to build a complete closed-loop system.
Except for intelligent vehicles, this closed-loop system requires a powerful algorithm training system and a large amount of road data. Compared with autonomous driving suppliers, OEMs with user and mass-production capacities have innate advantages in the most critical data link and are more hopeful in achieving autonomous driving in the future.
In January of the following year, Lang Xianpeng joined Ideal Automotive. However, Ideal Automotive was in a life-or-death stage of entrepreneurship in 2018, and financing repeatedly hit the wall, with extremely limited funds that could not support the independent development of assisted driving projects. Lang Xianpeng understood that the independent development of assisted driving would be carried out after the company’s financial improvement, and everything was just a matter of time.
In the face of difficulties, Lang Xianpeng and the self-developed project of the company entered a dormant period, until the first day of the Chinese New Year in 2021, when Lang Xianpeng received a call from Li Xiang, the long-awaited wait for three years finally came to an end.During these three years, Ideal Auto has also begun to build the necessary training and data platforms for the closed-loop system and has applied shadow testing, making the launch of the self-developed intelligent driving project seem decisive, but it was actually a prompt decision made under favorable conditions.
In the 2021 self-developed project, basic assisted driving is just a warm-up exercise. With the delivery of the 21st Ideal ONE, the self-development of NOA and AEB will be the real challenge.
Self-development Goals
Deformable Performance Indicators
In 2013, the assisted driving industry was far from as hot as it is today. Due to an accidental opportunity, Wang Jiajia joined the well-known Tier1 supplier Bosch China.
In the following years, Wang Jiajia became an elite in this field, and during this time, he began to pay attention to the development of several top new forces in China. In the opportunity for business exchanges, he established contact with Li Xiang and others. Before joining Ideal, Wang Jiajia was already the youngest director of Bosch China.
In May 2021, Wang Jiajia joined Ideal Auto as the Senior Director of Intelligent Driving R&D. Then, Lang Xianpeng handed over the task of self-developing assisted driving, including continued optimization of basic assisted driving and the most important development work of NOA and AEB.
At that time, NOA was scheduled to be delivered in September, and there were still many bugs that needed to be resolved in the already installed basic assisted driving software. Wang Jiajia, who had just joined the company and saw that there were only single-digit software control teams, felt unprecedented pressure.
But he had no time to regret. He had to quickly recruit troops, quickly improve their skills, and align everyone’s goals as soon as possible.
Wang Jiajia planned and managed the research and development in accordance with his mature experience. He broke down the project into time nodes, with upstream departments doing development and downstream departments verifying after the previous step was completed. Everything was done step by step.
However, the first delivery node of the project couldn’t be achieved, followed by the second and third. At this point, Wang Jiajia gradually felt unable to breathe, and he saw himself, who was once proud and confident, almost reaching the edge of being unable to continue.
From the results, the problem seemed to be in efficiency, but the real problem was hidden deeper.
After the May delivery of the vehicles, the work of optimizing the performance of basic assisted driving continued. For the automatic parking part, the performance indicator that was set at the time was that the success rate of parking after OTA could not be lower than the version on May 25th, and the OTA was scheduled to be pushed in June.
Before the push, Lang Xianpeng, as usual, tested the new version of the software and found that the parking spaces in the garage could not be parked in after he went back. Lang Xianpeng then gave feedback on this issue, but the parking team replied that it was unlikely and that it must be an extremely low probability problem.
Lang Xianpeng told the team: “I was able to park every day before. After the update, I can’t park in my parking space. This is definitely not a small probability event. You need to decide for yourselves what to do with this.”Let Lang Xianpeng did not expect that the parking team actually went online with the software without addressing the issue. Shortly after the release, a large number of complaints flooded in from the whole user community, all indicating that they could no longer park into the parking spaces after the system update.
It’s the most frustrating thing when the warned issue still occurs. However, Lang Xianpeng discovered that the initial performance target set for parking optimization was not to have a lower parking success rate than before the optimization. Based on the parking data from this version of OTA, the parking team actually achieved this performance target. Because the success rate of new parking spaces increased while the success rate of old parking spaces decreased, but the weighted average success rate was indeed maintained at the same level as before.
It seems that there is no problem, but for users, the actual experience they got is that their old parking spaces used to work, but now they cannot park into them after the update.
In Lang Xianpeng’s opinion, this is a result of mechanical execution of KPIs. As a result, the ideal internal workplace culture has been advocating and implementing the OKR work philosophy, and the biggest difference between the two is the understanding of goals. The project’s goal is to deliver high-quality and high-performance products to users, but many people only concentrate on achieving specific indicators such as parking success rate.
This is why everyone seems to be working so hard but still making some mistakes. The goals and milestones were discussed together, but often the problem was found as the work progressed.
The difficulty of R&D work seems to be on the engineering and technical level, but in fact, the problem falls on management during the actual implementation. In the retrospective meeting, Lang Xianpeng attempted to communicate with the team using the OKR concept, but the effect was not significant.
In the first few months of the NOA project, many people had not yet truly applied the OKR system promoted by the company because the team had entered a large number of new people. These talents with world-renowned background were initially more willing to use their own familiar experience and methods to promote and execute the project. When Lang Xianpeng discovered this phenomenon, he did not rush to stop it. His many years of management experience told him that it is difficult for people to realize the problem until they hit the wall.
After several delays in milestones, it was already September, and the team had been rapidly expanding with the addition of dozens of people every week. NOA, which was initially planned to be delivered on September 30th, was delayed until October 30th. However, the sustained high pressure and continuous negative feedback made the overall atmosphere become very depressed and frustrated.
It was finally time for Lang Xianpeng to take action. At the critical stage when the problem was fully exposed, on September 21, 2021, the day of the Mid-Autumn Festival, the self-developed team finally sat down together. Before outputting the methodology, Lang Xianpeng first harshly criticized the team, “Aren’t you all very capable? How did the OTA have problems? Why can’t the team coordinate together? Why can’t we deliver it by September 30th?”After creating a good atmosphere, everyone continued to discuss earnestly about what the goal actually is. Lang Xianpeng used the ideal LSA framework to organize things, and spent three days discussing with everyone and came up with the goal: high quality, high performance, and on-time delivery. The next step was to create Key Results (KR) around this goal. The top three KR were safety, product power, and delivery milestones, and the responsible person for each KR was confirmed.
At this meeting, Wang Jiajia fully absorbed the company’s management system for strategic goal, and summarized the team’s OKR as a philosophy question plus a math question. The goal O is the philosophy question, to unify everyone’s direction and values, and the KR is the math question, where everyone works together to break down the values and weights in a system of evaluation.
Only when everyone in each link has autonomous and unified goals and actions, can the team get rid of the linear system of the industrial age which relies on time axes. No longer will they wait for the entire process because of missed milestones, instead they can streamline their cooperation with each other. When a problem arises in one link, another link can adjust and allocate resources promptly. Wang Jiajia understood that the key to R&D management is to control the pace, not the time.
After the meeting, the quality team, R&D team, road test team, and project management team each set their own OKR under the project’s OKR, unified their goals, and broke down the specific values. Things went much more smoothly from then on. True strategic collaboration can only be established through consensus, and efficient execution can only be achieved when there is consensus.
Although Wang Jiajia had been walking on the edge of collapse, his pursuit of software product performance did not slack, and high performance and high quality besides on-time delivery remained his development criteria. He was not satisfied with the extensibility of the algorithms and frameworks in the software version delivered in the first phase of the self-developed project. The team patched up the software inconsistently between May and July after the NOA project started.
As the project continued, Wang Jiajia increasingly felt that the existing software architecture was limiting subsequent development. It was difficult to achieve high-quality NOA on this framework, and it was impossible to continue like this. In July, Wang Jiajia planned to refactor the software code, but there were only two months left until the established delivery time, and refactoring the code alone would take a month, making time extremely tight. If something went wrong, the project would be greatly affected.Under tremendous pressure, Wang Jiajia still persisted in making the decision to refactor the software, while also being prepared to leave in case of failure. At the end of August, the new version of the software was basically completed, but at this time, Wang Jiajia felt even greater pressure. Many things from the old version of the software were still missing in the refactored version, and if the old version could still be fixed and used, now it has become rough. As a result, the overall performance of the refactored version was even worse than the old version, and the team worked harder for a month but ended up with a worse situation.
Wang Jiajia’s decision was thus questioned by many, but he believed that the refactored software would gradually demonstrate its value in subsequent feature development. However, before that, he could only lead the team to accelerate development in the face of questioning, and make the refactored software surpass the old version as soon as possible. In Wang Jiajia’s memory, this was the most difficult period of time in the entire self-research project.
At the beginning of September, there were still many minor problems with NOA’s performance, and the car was still swaying during testing. However, as the development of the new version of the software continued, these problems were subsequently solved. At the same time, everyone also found that there were fewer BUGs under the new version of the software architecture, the functionality was richer, and the scalability was also better. As the advantages of the refactored version of the software gradually emerged in subsequent development, the intention of Wang Jiajia’s insistence on refactoring the software began to be understood by everyone.
By the end of October, NOA was already in a pushable state, but the team raised the safety target to a higher level, and conducted a full coverage road test across the country for a whole month, running more than 3 million kilometers in total, covering every highway in the country seven or eight times, and making corresponding adjustments for all the problem sections found during the testing.
The NOA project was finally delivered in December, and along with it was the self-researched active safety system from this iteration.
Inconspicuous Full-time ADAS
Small Function, Big Task
AEB, Autonomous Emergency Braking, is a system that automatically brakes a vehicle in dangerous situations to prevent or reduce collisions. AEB has a history of application in the automotive industry. In 2014, the European NCAP began to include AEB as an active safety function in the overall safety assessment of vehicles. In the following years, collision tests in the United States and China also followed suit, and it has now been widely used.
Compared to cutting-edge features like NOA, AEB is not particularly new and has low visibility in daily use by consumers. In fact, in the entire automotive industry, it has been rare for manufacturers to develop AEB in-house, as they have mostly relied on supplier solutions. However, AEB is unassuming, but it is the only ADAS function that works in all weather and road conditions. Creating a high-coverage AEB functionality is a daunting task, especially since most AEB suppliers come from Europe and America and have not optimized their research and development according to the specific challenges of Chinese road traffic (such as pedestrians crossing the street, electric delivery scooters, and lateral vehicles). All of these factors have made self-developed AEB an urgent matter for Chinese automakers.
Before embarking on a self-developed AEB project, the 2020 Ideal ONE model used a supplier’s AEB scheme, which performed poorly. In order to improve the safety of the system, many improvement requests were made to the supplier. Unfortunately, the supplier’s goal was to deliver the project on time, so implementing changes would take time and inevitably delay the project. This led to many reasonable demands being refused, which left Ideal in a passive position.
In order to improve the AEB’s performance and streamline system iteration, Ideal decided to develop its own AEB for the 2021 Ideal ONE model.
The official launch of the AEB project was in April, and the project resembled the self-developed basic assistance-driving project in its early stages, where everything had to be done from scratch, including algorithms, code, tools, and development environments.
In the initial stages of AEB research and development, implementing functionality was relatively easy compared to achieving the accuracy of AEB triggering. In other words, the team needed to ensure that AEB only triggered when it was actually needed. This meant solving problems such as “not triggering when necessary” and “triggering unnecessarily.” The key to optimizing this process was through a closed-loop data system, as previously noted by Lang Xianpeng.
Among AEB suppliers, Mobileye has the lowest rate of false triggering, making it the world’s best in class. When setting up OKRs, the team’s requirement for false triggering was to be on par with Mobileye’s standard, which included reducing the average false triggering rate to once every 100,000 kilometers.
By the end of July, the positive adjustment work for field testing was completed, and in August, the AEB team began planning the various work and milestones for reducing false triggering rates. Meanwhile, the road testing team began preparing for data collection and testing data scenarios, and the backend started implementing data labels.The quality department often provides feedback regarding issues users encounter when using AEB. For the AEB team, these tickets from users are crucial data scenarios that have been carefully screened, aiding in the optimization of the system’s accuracy.
At the system level, an ideal “shadow test” was very beneficial. The test software runs in the background of the vehicle, and the vehicle’s real driving data is provided to the software, but the software’s computing results are not executed by the vehicle; rather, they are sent back to the R&D headquarters.
Therefore, without affecting the user experience or safety, the AEB testing software was able to complete real-world road tests seamlessly by relying on data loops and obtaining results without any noticeable changes.
Compared to previous methods, these system-level efficient closed-loop configurations have significantly improved efficiency. In Xu Zhitao’s memory, in the past, when traditional AEB in cars had a false trigger, the customer would complain; then, the developers would go to the 4S store and read the data from the user’s car, which could take up to three days to obtain a single instance, resulting in significantly lower efficiency.
From August to October, the AEB false trigger rate reported by the system reduced to one occurrence every 10,000 kilometers. However, the following road to reducing the false trigger rate to one occurrence per 100,000 kilometers was filled with various challenging corner cases; the difficulty of these issues compared to before was not within the same dimension.
Xu Zhitao expressed that the process from one in 10,000 to one in 100,000 was painful because the previous methods were no longer feasible at this stage. Many cases involved extremely challenging issues.
“For example, the determination of the attributes of the vulnerable road user group. An individual trial scenario is easy, but after launching it in the market, a lot of challenging scenarios arise, such as during peak hours, where many citizens riding electric scooters cross the vehicle’s surroundings. In addition, when pedestrians suddenly walk out between parked cars in city alleys or interior roads of residential areas where many cars park on both sides, these scenarios plagued us for a long time.”
Xu Zhitao further explained the difficulties of these cases: “The calculations for these scenarios must be accurate for target attribute calculations. For example, during rush hour, the car’s surroundings are complex, and there are many targets. If perception algorithms and logic are not optimized for this scenario, two different target attributes may migrate, resulting in inaccurate or even incorrect physical attributes of the target. For example, a pedestrian in an alley who is planning to cross the street sees a car coming, so the pedestrian moves their leg slightly, changing from crossing horizontally to stopping or even moving vertically. At this point, the pedestrian is no longer crossing laterally, but the perception algorithm is not optimized, so they continue to be recognized as such, eventually leading to misoperation. For the end driver, this means a false brake, which is very scary.”
The rapid movement changes of pedestrians imply that the algorithm must have strong convergence.During the process, everyone realized that in ideal AEB perception before May, the team had been using supplier technology. However, after establishing a perception team, a deeper team integration was needed. So the company rented a large conference room in Beijing to move the regulation and control team, who had originally been stationed in Shanghai, to Beijing in late September to work face-to-face with the perception team.
Exploratory work often has no shortcut, and problem-solving is accomplished through the combination of smart people and stupid methods. At the early stage, when the two teams sat together to discuss, everyone had a headache because they did not know how to proceed. They tried to solve problems by trial and error. During the bottleneck period of the AEB project, with continuous exploration and overcoming difficulties by the perception team and the regulation and control team, everyone gradually found a way to solve the problem, and the feedback optimization between the perception and regulation and control teams played a crucial role.
For example, when the perception team had a problem with algorithm convergence, they had a method, but did not know what extent to do it or whether to do it. At this time, the downstream team would give the optimization direction and indicators from their professional perspective, and tell the perception team to react within a certain number of milliseconds and make the attribute correctly.
Similarly, when it takes a lot of time and effort to modify from the perception end, the downstream will also discuss with the perception whether to optimize by adding logic from the downstream.
In this way, the perception and regulation and control teams formed a closed-loop feedback collaboration mode. Later, everyone tackled one problem point by point, and combed through the data to overcome one Corner Case after another.
The purpose and value of work:
In fact, during the self-research project, everyone also thought about what is the goal of AEB engineers?
When working with suppliers, this goal is very simple, which is to deliver a product agreed upon by everyone to the OEMs, and the product is mainly tested for cost-effectiveness.
In ideal work, one of the goals of the entire intelligent driving team, including AEB, is to deliver high-quality, high-performance products to users, and the product strength is tested in terms of product value, R&D efficiency, and supply relationships. More attention is paid to some very detailed operational indicators.
In an early OKR meeting, Xu Zhitao set an easily achievable performance indicator for pedestrian crossing scenarios. Lang Xianpeng saw it and asked Xu, “Do we need to protect the pedestrians outside the car? We must protect the passengers and the driver in the car.” Xu Zhitao instinctively felt that in the limited time and resources, the main optimization items should be reflected in the parts related to passengers in the car, and protecting passengers in the car was the core goal.
Lang Bo said that he was not satisfied with this answer and asked Xu to think about it. Then he said to the meeting host beside him, “Give Zhitao an Open Problem List. Tomorrow I will ask him whether there is any problem with his goal.”
This incident made Xu Zhitao feel embarrassed, especially when someone followed up with him after the meeting and asked whether Lang Bo had figured out the problem he left for him.Actually, he realized shortly afterward that pedestrians outside of the car were also lives, and needed to be protected just as much. However, he was worried that once the horizontal crossing performance indicator he had set was raised, it would involve the task planning of the upstream and downstream departments, adding extra workload for everyone. At that time, every team already had a lot of things to do, so if this disrupted someone’s original plan, they would definitely be unhappy.
The next day, when Xu Zhitao was asked again, he proposed to raise the AEB covering speed of the pedestrian crossing scenario from 40 km/h to 50 km/h, and this time Lang Bo did not continue to “trouble” Xu Zhitao.
But in the beginning, there were complaints from the small team about such things. Although the speed of the scenario only increased by 10 km/h, the increase in difficulty was far more than 1/4. Later, to achieve this performance indicator, the team spent an additional month on development and testing.
There was another thing in the self-developed project that deeply touched everyone. The company’s background system had a set of case scenario material libraries, which were originally intended to collect failed scenarios to help everyone better optimize and iterate software based on the scenarios.
This set of material libraries also gave AEB engineers a perspective they had never had before to re-examine their products. There were segments of some major accidents in the case scenario library. When these AEB engineers saw collisions that happened in the real world for the first time, especially in accidents involving pedestrians, they actually felt uncomfortable deep down.
These accidents all had one thing in common: either the driver did not notice, or did not have time to react, but if the AEB system were used, perhaps there could be some redemption. After seeing these accidents, everyone’s mood was difficult to calm and settle for a long time because some scenarios were already under development at that time, and they made this hypothesis: if everyone’s efficiency could be a little higher and the software could be put online faster, perhaps this accident could be avoided.
From that moment on, they began to feel like doctors, saving potential accidents that had not yet happened. After AEB was delivered online later, everyone would also feel a different sense of satisfaction when looking at the AEB trigger data statistics in the background: every number on the screen is a potential accident that was successfully avoided. The invisible hand that was written by code, pulling users back from the brink of death, was created by the efforts and hard work of the AEB team.
This group of people had a deeper realization of the meaning and value of their work, and transformed this understanding into the driving force of their work. Indeed, they regarded AEB development as a lofty cause to fight for.
Wang Jiajia had an even more simple-minded view, thinking that AEB was the project that best reflected basic skills in this industry. Every 3 million kilometers of mileage could avoid 15 collisions, but such life-saving things relied heavily on data samples piled up by people in the laboratory. Other than basic software and algorithm skills, it was in the experimental field where people’s tenacity was truly tested, experiment after experiment.However, the feeling of satisfaction after the project’s achievements is still the most exciting. After the AEB project topped the DCD car ranking, Ideal became a research target for its competitors. Wang Jiajia proudly stated that AEB team of their competitors are now inquiring about them.
In fact, when XPeng NGP was first launched, Wang Jiajia, who was a supplier to XPeng, was deeply shocked. At that time, XPeng’s full-stack self-research project had already established a team of hundreds of people, and the technology of each self-research link had made constructive progress.
At the initial OKR co-creation meeting, there was not much confidence in competing with XPeng. After all, when XPeng’s self-research started to make great strides forward after years of accumulation, Ideal had just begun.
After the project was finally delivered, the atmosphere changed. With the unremitting efforts of everyone, they personally achieved what they thought was impossible. And the team of less than 100 people at the beginning of the year had already reached more than 300 people. At first, the teams from different work backgrounds had major difficulties in cooperation, but in the process of achieving goals, many people truly felt the power of the organization and understood the importance of the closed-loop data system in the development of autonomous driving.
Those engineers from big companies who always acted conservatively also began to have confidence in competing for first place.
In conclusion
With the official launch of NOA and AEB, Lang Xianpeng finally dismantled the “resignation time bomb” he personally installed. Ideal’s large autonomous driving team has also finally completed the process from 0 to 1.
Lang Xianpeng said, “This is strategy. Strategy does not ask if something can be done, but only if it is necessary. When the company needs you to stand up, you have to make it happen no matter what. After the strategy is determined, what is left is execution.”
In order to achieve necessary strategic goals, the self-researched large project has also experienced many challenges from 0 to 1 in this year, and the pressure has continued from the beginning of the year until the end. However, under the battlefield, the scars are everywhere. Some people have grown in the team, while others cannot bear the loss of control of their lives and emotions, and choose to leave. Among them are several core employees who Lang Xianpeng valued very much.
At this time, Lang Xianpeng, who was executing strongly to the point of cruelty in the project, revealed his compassionate side: “Bearing responsibilities that do not match your age is a disaster for most people, but they still bear those things they should not bear. The excellent cases we see now are just the survivorship bias. When we see a person who is so amazing and capable at a young age, it is only because they are the survivors.”At the end stage of Ideal 2021 self-developed project, Lang Xianpeng said to the team, many people who left the company in the early stages of the project had no confidence in the company and thought that the company was just bragging and shouting slogans, and that Ideal couldn’t achieve it. However, fortunately, most people chose to stay, and the project was eventually completed. For those who participated in the project delivery last year, they experienced a very important battle, and there must be people in this team who will become the top talents in the world’s autonomous driving field in the next 10 to 20 years.
With the increase of newly delivered cars, Ideal ONE’s assisted driving has now exceeded 240 million kilometers, of which 2021 contributed more than 60 million kilometers. The accumulated mileage of NOA exceeded 15 million kilometers, averaging around 4 million kilometers per month. With the increase in sales, it is expected that the total mileage of NOA will exceed 100 million kilometers this year, while other competitors have publicly announced data of only 20 million kilometers per year.
This year, the key factors that helped Ideal’s self-research success included the decisive decision-making of the high-level, the dual push brought by internal and external pressures, the early layout of the closed-loop data iterative system, the network-style organization management driven by goals, and the team’s efforts to achieve the goal. The integration of many factors allowed Ideal to challenge their limits in one year and achieve timely compensation of NOA functions. They also become the only company in the industry to achieve self-developed AEB.
During the 2022 China Electric Vehicle 100-People Conference, Li Xiang proposed in a closed-door meeting to make AEB a standard configuration for cars, and this proposal was supported by Miao Wei, the Vice Chairman of the Economic Committee of the National Committee of the Chinese People’s Political Consultative Conference.
With the listing of L9 in April, Ideal cars insist on the standard configuration of auxiliary driving, which will allow more consumers to use safer intelligent driving products and further expand the scale of the full-stack self-developed closed-loop system.
In the first echelon of new forces, Ideal has obvious advantages in product definition and vehicle sales data, but due to the lack of inherent resources, Ideal’s efforts in intelligence are actually in a follower position. However, it is the excellent domestic car brands that push the level of independent research and development of domestic cars in intelligent driving one step higher with their full effort to catch up and surpass each other. The brands that compete and cooperate with each other have a relationship that is both competitive and friendly. At the company level, everyone is a competitor to each other, but in the race of domestic cars, they are constantly learning from each other and making progress while competing with the world’s best car brands.
For the history of China’s autonomous driving, the rise of new forces in recent years is just the beginning. The process from 1 to 10 to 100 is still full of variables, but at the same time, it also contains a huge opportunity to rewrite the global position of Chinese cars.Finally, feel free to download the Garage App to keep up with the latest news on new energy. If you want to get more real-time communication, you can click here to join our community.
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.