Author: Su Qingtao

“Did Hesai choose the wrong technological route? Can Hesai’s products only be sold to Robotaxi companies for demos and not for mass production?” Since the first half of 2021, many insiders in the autonomous driving industry and investors have been questioning the competitive landscape of the lidar market.

However, in late September of this year, Hesai officially announced its first mass-produced product for the front market: the AT128, with “a monthly delivery volume of over 10,000 units,” becoming the first domestic lidar supplier to announce this number. In fact, due to the Ideal L8 being offline in August, Hesai’s AT128 has started delivery since July.

At this point in time, according to Hesai, the AT128 was released less than a year ago in August 2021 (if the Shanghai epidemic did not delay the progress, the schedule would have been faster). Li Yifan said that compared with other peers’ product rhythms, Hesai’s AT128 is the “latest release and the earliest submission.”

So, how did Hesai do this? What new ideas did Hesai have in the process of achieving from 1 to 10? With this series of questions, Jiuzhang Autonomous Driving conducted an interview with Hesai CEO Li Yifan at the end of October. The interview notes are now organized as follows.

The biggest challenge experienced

Jiuzhang Autonomous Driving: Judging from the mass production and delivery performance, Hesai is now a top player in the lidar industry, but it must have overcome many difficulties to get to this point. What is the biggest challenge Hesai has faced from its establishment to now?

Li Yifan: The biggest challenge is still the execution of strategy. When we planned our first product in 2015, we didn’t win with “cost-effectiveness” like typical Chinese hardware companies, but decided to make the most high-end category first, hoping to win market recognition with the best product performance. Looking back, we didn’t bet wrong on this strategy, but it’s difficult to execute, from R&D, sales to customer support, the difficulty is enormous.

We have no regrets about choosing the Hard mode

Looking back, we do not regret choosing the Hard mode. If we had started with a 16-line LiDAR of 10k-20k RMB, the early development and sales would have been relatively easy, and the market would have been considerable. However, this choice would have made us focus more on the supply chain and cost reduction, and we would not have had enough energy and motivation to develop high-end products. We would also have had no chance to collaborate and push the leading-edge technologies with the world’s top players, and in the long run, it would have been a regrettable thing.

Why is the product release not proactive?

Ninebot ZhiJia: Typically, many foreign chip and LiDAR manufacturers enter the mass production and delivery stage 1.5 to 2 years after product release. However, in China, I found that Horizon’s J5 began mass production and delivery less than a year after its release, and Hesai’s AT 128 went into mass production and delivery from the end of August last year to the end of July this year, also less than a year after release. Since the mass production progress is rapid, why is the product release later than others?

Li Yifan: Releasing the product when it is still far from maturity is usually for financing, or to target the financing of competitors, but we have neither of these demands.

The focus should be on precise positioning instead of quantity

Ninebot ZhiJia: Is your ideal to become Hesai’s first production customer related to your and Li Xiang’s relationship from Lakeside?

Li Yifan: No, it’s not. Li Xiang’s decision-making is still very rational. How could he buy our product just because we are classmates? It is because he recognizes our product enough.

On the other hand, since we are more familiar with Li Xiang and have learned a lot in the communication process with him, we have invested more resources in our cooperation with him.

Ninebot ZhiJia: What are your gains from communicating with Li Xiang and collaborating with Ideal?

Li Yifan: Li Xiang is very thoughtful and good at expressing his ideas. He has given us many advanced concepts, such as his most important concerns about organization and strategy.# Translation

Li Xiang seldom leaves his office, it seems like he is always sitting in there, saying, “Every day, I just think about how to solve problems with organized methods,” which is quite admirable. We also use Feishu under the influence of Li Xiang.

Nine-Chapter Intelligent Driving: I found that Hesaix emphasized less on the number of fixed-point projects, but more on the scale of a single fixed-point. The Ideal One (the first front-loading mass production project based on Ideal L9) was the target. So, the cost-effectiveness of taking small projects compared to such large projects is not high.

What I can think of is that too many small projects may affect the stability of the team. For example, engineers may feel that the problems solved on projects B, C, and D are the same as those on project A, doing a “low-level repeated construction,” and their own skills do not improve much. Hence, they may resign. Besides, what are the other problems?

Li Yifan: There is something worse.

When working on small projects, the engineer’s enthusiasm may be insufficient or even resentful, so they will not invest many resources in serving customers well. In the eyes of customers, their small project is undoubtedly very important, and they cannot understand the supplier engineer’s “carelessness.” Instead, they will ask, “What makes you not support me well?”?

It can be said that if you do too many small projects, it is difficult to handle the relationship with customers. We have also learned a lesson in this regard before.

Therefore, we believe that it is a better choice to do large projects for impressive customers, and the number of projects should be relatively small. These points are just what we have achieved in the Ideal L8 project.

Use the Power of Organization to Make up for Manufacturing Shortcomings

Nine-Chapter Intelligent Driving: Everyone was in great pain once it entered mass production, but if that hurdle could be overcome, it would be a transformation. So, for Hesaix in the past six months, what are the challenges that have brought the management the most significant pressure during mass production but have been successfully dealt with?Li Yifan: We, the tech entrepreneurs of our generation, are naturally interested in starting from scratch – at this stage, we innovate with “something from nothing”. However, mass production, which involves going from 1 to 10, is not as interesting.

At the 1 to 10 stage, the first problem entrepreneurs face is their capacity being insufficient because they lack experience in this area. However, this is normal. In our industry, the vast majority of entrepreneurs lack experience from 1 to 10.

We became aware of this problem two years ago. As an innovative enterprise, after designing our product, the question was whether we could achieve mass production that meets vehicle regulations. This requires operational and manufacturing abilities, so should we address this capability gap or not?

We started designing the production line in 2021 and began capacity ramp-up in 2022. During the epidemic in the first half of the year, several hundred engineers lived in the factory to debug equipment. Currently, the annual production capacity of this production line is 300,000 units, but because the degree of automation is over 90%, less than 30 people are required to handle some auxiliary work.

To make up for the deficiency in manufacturing capabilities, we invested a lot of time in learning from our shareholders Bosch and customers like Ideal, as well as doing a lot of management consulting.

Li’s inspiration was significant. In the end, we had to rely on the organization to solve the problem. After starting to build the production line, we hired a former production manager from a certain computer enterprise, who had managed tens of thousands of people, to become the factory director. Under his leadership, the original “research and development-focused” organization became one that was “more production-oriented.”

However, at this stage, there were also some entrepreneurs whose awareness was insufficient. For example, they did not realize that “it is different now than before” or that their own capacity was insufficient. They also did not think about attracting sufficiently mature professional managers to solve current difficulties.Autonomous Driving at the Ninth Chapter: A friend who had been responsible for production technology said that the biggest challenge faced by mass production of LiDAR was not in technology, but in engineering and production technology. “Don’t be fooled by many companies claiming to value engineering. How high is the status of those responsible for engineering in your company’s organizational structure? If the status of those responsible for engineering is not high in your company’s organizational structure, I don’t believe that your engineering can be successful.”

Li Yifan: I agree.

Here, from 0 to 1 research is carried out by Chief Scientist Sun Kai. From 1 to 9, engineering is managed by CTO Xiang Shaoqing. From 9 to 10, production is managed by the factory manager.

The overall person in charge of engineering is CTO, who is also one of the co-founders, and has a high status. Moreover, CTO manages nearly 1,000 people (excluding workers), while the other co-founders, Sun Kai and I, each manage over 100 people.

Autonomous Driving at the Ninth Chapter: Why did you choose to build your own factory instead of entrusting production to an OEM factory?

Li Yifan: Many companies put all their interests in the 0 to 1 phase (research and development), and this pure interest-driven approach is prone to problems. While many other companies also pay attention to the entire process from 1 to 10, they are disconnected between the engineering phase (from 1 to 9) and the production phase (from 9 to 10). They believe that “once this thing is designed, it’s done” and “just hand the drawings over to the factory (OEM factory) to build, and I don’t have to worry about it.”

Autonomous Driving at the Ninth Chapter: I have noticed that most foreign LiDAR companies prefer the light-asset model, entrusting production to OEM factories, while domestic companies such as Hesai and Livox emphasize self-built production lines.

Li Yifan: I have a view that “research and development is part of production” but it has not been recognized by the industry.Jiuzhang Autonomous Driving: I agree with you, this is called “starting with the end in mind”. You need to know first how to design for easier production and better quality. Then, design it so it’s “easy to manufacture” during the R&D phase – I’ve worked in manufacturing before and often see problems arising from poor design during the R&D phase, so I have a deep understanding of this.

Li Yifan: In industries with highly mature and standardized processes, production doesn’t necessarily have to be a part of R&D. However, in emerging industries without prior experience in large-scale manufacturing, production must be a part of R&D. Tesla’s integrated body shell is a typical example.

In our industry today, we are experiencing the first large-scale manufacturing of lidar in human history. No one can say they have enough experience, so at this time, if it’s not the designing company itself doing it, who can do it for you? If the contract manufacturer has the ability to “do it all at once,” then why would they want to contract with you instead of becoming a “lidar company”?

Jiuzhang Autonomous Driving: I have doubts about this too. Previously, when I saw some companies entrusting their production to a contract manufacturer who has never made this product before, I wondered, “Is the contract manufacturer omnipotent?”

Li Yifan: You can only make standards and let others do it for you after you fully understand it yourself. If you do not even have a clear standard, nor can produce small batches of products, then can you expect the contract manufacturer to assist you in everything? After all, the level of skill of the contract manufacturer’s personnel cannot be higher than yours in producing small batches.

I think contract manufacturers are essentially copiers, and the necessary condition for copiers to copy “good things” is that the quality of the “original” is good and it does not need to be changed anymore, right? However, the lidar is clearly a product that is still being iterated on. Currently, it has not iterated to the stage of good enough quality, so what will come out of the copier if you stuff it into one? If the copying is not done well, even if there are many orders, delivery will be challenging.# Translation to English Markdown

Of course, this issue is always subjective, but I actually think it’s pretty clear – Why would someone as cost-sensitive as Musk build his first factory in such an expensive place as Silicon Valley? And why did he sleep in the factory for a while? It’s because manufacturing is part of R&D. What Musk likes to do most is “factory patrol,” and he fires workers who “aren’t doing well” on the spot.

Nine-Chapter Intelligent Driving: If you do it yourself, the first batch of workers on the production line needs to be highly skilled.

Li Yifan: Absolutely. In fact, our first batch of workers were engineers.

Nine-Chapter Intelligent Driving: Under what circumstances do you consider working with contract factories?

Li Yifan: When the product’s performance has been repeatedly verified by the market and the demand is sufficient.

From 1 to 10, the definition of “talent” needs to be redefined

Nine-Chapter Intelligent Driving: In the autonomous driving industry, many companies are facing a common problem: they have recruited many talented individuals, but because most of these individuals have their own opinions, having too many talented people may lead to difficulty in forming a cohesive team, thus causing poor overall execution for the company. In my experience, Haylion’s execution is relatively good within the industry. Do you have any experience you can share regarding this issue?

Li Yifan: Based on what you’ve said, it seems like your definition of “talent” is still limited to just individuals who have high academic qualifications or who previously worked in prestigious companies. Software companies are indeed more interested in these types of talented individuals, and individuals who meet this definition may indeed have a problem forming a cohesive team.However narrow the definition of “genius” is — resume alone is not the only criterion for measuring one’s genius. Our company does not lack brilliant minds with impressive background, but we realized two or three years ago that if you want to achieve the transition from 0 to 1 to 1 to 10, you will have to re-define the term “genius,” because the geniuses under your current definition are mostly scientists who may not be suitable for the needs of a company in the 1 to 10 stage.

Before entering the mass production stage before the attack, we fought side by side. Not only do we need geniuses to shoot arrows on the front line, but we also need geniuses in logistics and support. The marketing and sales departments also need geniuses. Under this premise, the traditionally defined geniuses may not be considered geniuses anymore.

However, usually only when the company enters the 1 to 10 stage, entrepreneurs will realize these problems.

Product, business model, ceiling

Jiuzhang Autonomous Driving: Are you providing completely standardized products or customized ones to customers? If customized, what is the proportion of customization?

Li Yifan: Taking AT128 as an example, this product has quite a few customers, and their versions may differ in certain aspects. First of all, the appearance is different — it needs to be adapted to different car models and sometimes the shape of the light projection needs to be modified; secondly, the software and communication are also different.

Jiuzhang Autonomous Driving: Do you only provide hardware, or do you also provide software algorithms?

Li Yifan: There are many layers of software algorithms, such as signal processing algorithms that convert raw data into point clouds, and point cloud processing algorithms that process point clouds into “images.” For most customers, we only provide raw point clouds, and they do the point cloud processing algorithms themselves; but for some customers, we also provide point cloud processing algorithms.

Jiuzhang Autonomous Driving: Two years ago when we talked about this issue, you said that the algorithm of Lidar is only an value-added service, and you would not make it a formal business. Has your opinion on this issue changed now?Li Yifan: I still hold this view.

Firstly, although it seems that many of the point cloud processing algorithms of self-driving companies/host factories are not as professional as those of laser radar manufacturers, they hope to improve in the long run.

Secondly, many customers want to do multisensor fusion, which means that the original point cloud data from the laser radar is essential. If the laser radar manufacturer does the point cloud processing algorithm, how can downstream fusion be done?

Thirdly, in China, there is an objective reality that host factories have a low willingness to pay for software, which means that software algorithms are difficult to become significant sources of income.

Some chip companies also provide algorithms, but their main revenue source is still chip fees. The proportion of software in total revenue is very low.

Market structure

Jiuzhang IMa: In theory, traditional Tier 1 giants should also have technical reserves in the field of laser radar, but why are start-up companies like Hesai more active in the front-end now?

Li Yifan: Giants are not almighty. In my opinion, the advantages of giants are mainly money and the systems they have accumulated.

Let’s talk about money first. More money makes things easier, but there are still objective barriers to R&D, such as the design iteration cycle, chip fabrication cycle, and customer validation cycle. Developing hardware products is a bit like giving birth to a child. It’s not like finding five people to give birth together, and the baby is born in two months.

Similarly, Hesai now has more than 1,000 people. The product we iterated for two years, even if the giants spend money to recruit 10,000 people, cannot be shortened to two months. Especially for large companies, it’s like a big ship. When it sees the direction and runs at full speed, it can be fast, but it’s not easy to turn around and quickly gather so many people and money from the existing system and find the right direction.Usually, what the giants are good at is something they anticipated 10 years ago, prepared for 8 years ago, and almost figured out 4 years ago, while new LiDAR is a completely new market. Startups respond much faster to the demands of a new market than the giants.

In addition, the giants have many business directions, but each one will not be the “top priority project”. Therefore, it is difficult to say how much manpower they can invest and whether they might change their direction halfway through. In our case, LiDAR is not only a “top priority project”, but also the “only project”. We have no way out, so we must go All In.

Of course, compared with the system of large companies, we small companies are indeed at a disadvantage. For example, the regulatory-level design and certification of LiDAR for automotive use is mostly about system and process rather than strictly defined innovation. If we rely solely on our own trial-and-error, it would be very painful and inefficient, which is unsolvable in my opinion. The only solution is cooperation.

After we figured this out, we asked ourselves, who is the most knowledgeable player in the world? The answer is Bosch. Bosch Group is the largest Tier 1 in the world and very good at automotive electronics. We have been in contact and cooperating with them since 2017, and we have even received investment from Bosch, just to learn more from the player with the most experience in the regulatory-level system construction.

Nine-chapter autonomous driving: How high is the technical barrier for LiDAR? If the barrier is not high, and more new players join the market after startup companies mature the market, won’t your market share become lower and lower?

Li Yifan: This is indeed a problem in the hardware industry. I think the essence of whether the barrier is high or not depends on whether the core technology of the industry is advancing quickly enough. If the industry is constantly iterating quickly, then the barrier owned by industry-leading players will surely be relatively high.

The long-term competitive barrier is in the chip.Nine Chapter Intelligent Driving: Currently, it seems that Hesai is among the top-tier companies in the field of LiDAR. So, in the medium to long term, who do you see as your main competitors?

Li Yifan: In the short term, the competitive barrier for LiDAR lies in manufacturing. Therefore, players with strong manufacturing capabilities will win out in stages. In the medium to long term, however, the competitive barrier for LiDAR lies in the chip. Thus, the companies with the strongest chip capability will have the most competitiveness in this area. For the past few years, Hesai has been investing heavily in the chipification direction of LiDAR, based on this consideration.

Nine Chapter Intelligent Driving: Do you think LiDAR will become a universal technology with a high level of standardization behind it, which can lead to a high degree of market concentration? Or will it be highly customized with a low degree of market concentration?

Li Yifan: Great question! This depends on the barriers to entry in this industry. If manufacturing is the barrier, the manufacturing force will be dispersed and the market concentration will be low. However, if chip is the barrier, market concentration will be high. For example, the market share of the first-ranked company may be as high as 60%, and the market share of the second-ranked company may be around 20%.

As mentioned earlier, we believe that the short-term competitive barrier for LiDAR lies in manufacturing, while the long-term barrier lies in the chip.

What is your view on the “threat” from computer vision technology?

Nine Chapter Intelligent Driving: Musk insists that with powerful visual algorithms, L4 autonomous driving can be achieved without the use of LiDAR. So, what is your opinion on the progress of computer vision algorithms, and will it threaten the market volume of LiDAR?Li Yifan: I am a Tesla owner and I use Autopilot every day while commuting. It really improves my driving experience, especially during traffic jams. However, as an autonomous driving practitioner, I always remain nervous when using Autopilot, constantly focusing on the road ahead, stepping on the brake, and keeping my hands on the wheel, ready to take over at any time. It’s like being a safety officer risking my life – unexpected braking, failure to detect cutting in… I have experienced most of the issues encountered by netizens.

If you carefully analyze these common dangerous scenarios of Autopilot, many of them are due to the lack of robustness of the visual and millimeter-wave radar solutions alone. These issues can be solved by adding a laser radar. I often turn the process of testing Autopilot on my way to work into a process of analyzing the installation position and field of view design of ADAS laser radar, and discuss the results with the product manager as soon as I arrive at the company.

Musk did say that it wasn’t necessary to install a laser radar in the car, but I have a theory – he doesn’t really think that a laser radar is useless, but rather that there is no suitable laser radar for Tesla yet. However, Tesla’s Autopilot functionality must be released on schedule, so he must adhere to his “laser radar is useless” theory. Otherwise, it would mean acknowledging that “Tesla does not have sensors that can make Autopilot safer, and instead gives customers an insufficiently safe product.” Therefore, once the laser radar becomes mature and cheap, I believe there is no reason for Tesla not to use it.I am a photography enthusiast. There is a story in the photography industry that has a similar sentiment. In 2005, when Canon released the world’s first full-frame digital camera 5D, Nikon fans all over the world asked when Nikon would upgrade to full-frame. However, Nikon’s CEO publicly stated: “Nikon does not believe that full-frame is the future trend and hopes everyone will continue to purchase Nikon’s APS-C format cameras and lenses”. A few months later, Nikon’s full-frame D700 was launched, and the CEO’s statement was forgotten…

Likewise, I firmly believe that Musk does not dislike LiDAR, but rather has not yet found a LiDAR that meets both performance and cost requirements. In fact, whether or not to install LiDAR is not a technical decision but a business decision—depending on the comparison of revenue and costs.

Regarding cost, the cost of LiDAR is clearly rapidly decreasing and may eventually become only 2-3 times the cost of cameras. Furthermore, in many corner cases, accidents can occur without LiDAR. Given the same conditions, would anyone think that a car without LiDAR is safer than a car with LiDAR?

Let’s make an extreme assumption: suppose that in 10 years, visual algorithms have become powerful enough and there are 100 million autonomous vehicles worldwide. If none of these vehicles install LiDAR, they would cause only 10 deaths, but if they had installed LiDAR, the number of deaths could be reduced to 5. Do you think this is valuable?

Jiuzhang Auto: Yes. When it comes to matters of life, we cannot just calculate economic benefits. As vehicle safety continues to improve, consumer thresholds increase. In the end, things related to life safety are only qualified with a score of 100.

Li Yifan: This is the concept of Westerners. The risk of accidental death is either 0 or 1. Therefore, when the price of LiDAR drops to within 1,000 yuan, would consumers still choose to lower their vehicle safety level just to save a little money?### How to Define the “Primary Sensor”

Nine Chapters Intelligent Driving: In the past few years, the common belief in the autonomous driving community was that “laser radar is the primary sensor, and cameras are the secondary sensor” for multiple-sensor fusion. However, over the past two to three years, an increasing number of people have started to say “use the camera as the primary sensor and use the laser radar as a redundant sensor”. Does this mean that the market demand for laser radar will decrease?

Li Yifan: How do we define who is the “primary sensor”?

Nine Chapters Intelligent Driving: An explanation of the head of autonomous driving at a new energy vehicle company is: “When the identification results of two sensors are inconsistent, the decision-making system decides whose opinion to follow, and that sensor becomes the primary sensor.”

Li Yifan: If the camera is the primary sensor to listen to when the identification results of laser radar and cameras are inconsistent, why not just remove the laser radar? And if we listen to the laser radar when the two are inconsistent, then the meaning of calling someone the “primary sensor” is not significant—similar to saying “I am the primary one at home, but when there is a disagreement, I listen to my wife”.

In my opinion, in the era of multiple sensor fusion, sensors with different principles complement each other in different scenarios, and there is no need to tout a sensor as the “primary sensor”. Adding a laser radar is safer than not adding one, and I don’t think anyone would disagree. From the perspective of information redundancy and information complementarity, more is better, and this is an indisputable scientific truth.

The controversy of this matter actually lies in whether the extra system security is “worth the price”. Today, laser radar is nearly a hundred times more expensive than cameras, and the vast majority of people still think it is “worth it”. As deep learning technology evolves and the cost of laser radar decreases, this “100 times” will eventually converge to “10 times” or “1.2 times”, which is a topic that we can boldly speculate about. After all, ultimately selecting or not selecting a sensor is just a balance between system performance and overall cost.Ninebot Smart Driving: On the surface, it seems to be a technical issue, but ultimately it is a financial one.

Li Yifan: It’s not just about money, it’s also about safety, and balancing the two is actually an economic problem.

First of all, people say that life is priceless and safety is priceless. Is safety really priceless? If it is, then every car should be equipped with the most expensive and best safety equipment, the thickest steel plates, the most airbags, and the best-performing laser radar… This is obviously not the case.

Why? Why does a car decide not to install 8 airbags after having a plan with 6 airbags? The answer is that 6 is already safe enough. The slightly increased safety of adding two more, compared to the increased cost, is not worth it.

Similarly, when will laser radar become “unnecessary”? It’s when the camera-based solution is already “safe enough”. We decide to cut off the laser radar from the cost perspective because its additional cost is not enough to support the slightly increased safety it brings, after all, the camera is already “safe enough”. We don’t have enough evidence to prove that the day will never come because the improvement of deep learning technology and computing power is constantly evolving.

However, in a few years or decades, one day, everyone will finally feel that unmanned vehicles are “too safe” and laser radar is no longer needed. It is highly probable that this technology has been validated and popularized, and millions of Level 4 autonomous vehicles are produced every year. The dream of unmanned driving has been realized, and laser radar has also fulfilled a great mission and formed a sufficiently large industrial cluster.

So the question is, at that time, is it possible that laser radar will also be cheap enough for everyone to easily afford?

Control Over One’s Fate

Ninebot Smart Driving: Does laser radar companies have a sense of control over their own future in the entire autonomous driving industry?Li Yifan: This possibility definitely exists, and it is the issue that we are truly concerned about. The rise of any supplier is based on the growth of your customers as scheduled. Therefore, we must diversify our risks, not just focus on the autonomous driving industry, but expand to multiple application scenarios and serve the broader robotics industry, including many mid-range and low-end applications. It is unlikely that these industries will not develop, right? I think robot technology will definitely have a future in the long term.

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.