Produced | Tiger Sniff Car Team
Author｜Wang Xiaoyu
Edit｜Zhang Bowen
Lidar, by emitting laser beams to measure the distance and orientation of objects in the surrounding environment, so as to determine the relative position of the vehicle and obstacles. The autonomous driving algorithm will finally issue various operating instructions to the vehicle based on the data sensed by these sensors.
TeslaCEO Elon Musk has publicly attacked the lidar route for many times: “Lidar is only used by fools” and insists on adopting a pure visual route. But on the contrary, almost all domestic auto companies have chosen the lidar route. By pre-embedded lidar, new cars are labeled as “intelligent”. Recently, everyone has “rolled up” on the number of hardware and the number of models.
According to Huxiong statistics, there are at least 12 LiDAR models that have been mass-produced/planned to be mass-produced in 2022, and most of the new models were unveiled at the Guangzhou Auto Show. Among them, carryHuaweiThere are more than three models with autonomous driving technology, while Xiaopeng Motors has two LiDAR models, and the Salon Automobile, a high-end brand of Great Wall, is equipped with four LiDARs exaggeratedly.
This gives consumers a sense of sight of the “first year of lidar”.
In the final analysis, car companies are madly launching lidar models in order to raise the ceiling of the hardware. No matter which autopilot software functions can eventually be realized, at least the hardware parameters cannot be lagging behind. The son said: “Workers must first sharpen their tools if they want to be good at their work.” Although the idea is good, with only one page of PPT and one show car, why should consumers spend four to five million yuan to pay for the lidar? ?
The domestic lidar manufacturer Saiteng Juchuang told Huxi that the time of industrial explosion has come, and the lidar industry has begun to accelerate its development around 2016. High-performance automotive-grade lidar products have matured and been mass-produced-“2021 It is already the first year of mass production of lidar.”
1. The battle of routes: there is no ultimate plan yet
At present, the domestic lidar market is undergoing a process of transformation in a technical route that is almost the same as that of power batteries.
In the early stage, the ternary lithium battery was the main stage, and the company pursued high performance while ignoring the high cost. However, with the emergence of more “domestic substitutes”, lithium iron phosphate batteries with excellent cost are sought after, but there are certain shortcomings in performance. Each technical route has its advantages and disadvantages, and companies will try to solve the problem from a technical level. But for a long time, the ultimate solution may not be seen.
Lidar, as a “what you see is what you get” sensor, can enhance the redundancy of the sensing system, supplement the scenes where millimeter wave radar and cameras are missing, and is a “necessary artifact” for high-end autonomous driving.
According to the structure of the scanning module, lidar can be roughly divided into three types: mechanical, semi-solid, and solid.
Classification of main technical routes of lidar (Source: Hesai Technology Prospectus)
Mechanical lidar is the earliest and most mature technical route to enter the market. It refers to the scheme of arranging multiple lasers in the vertical direction and driving the photoelectric structure to rotate 360° through a motor, thereby turning points into lines to form a three-dimensional point cloud. The number of lines is proportional to the resolution, and it has high resolution and high resolution. Features of ranging.
“Mechanical lidar is to use a mirror to revolve around the laser source to reach more angles of coverage. You see that some vehicles have instruments on it. This is the principle.” A practitioner from an autonomous driving company told Tigersniff, because of the high assembly requirements, it is difficult to achieve mass production mechanically. For example, the Velodyne 32-line lidar HDL-32E requires 32 sets of transmitting light sources and 32 sets of receiving light sources for one-to-one debugging, which is prone to failure.
In addition to the manual operation of laser stacking, the average failure time of the mechanical type is only 1000-3000 hours, which is significantly different from the minimum 13000 hours required by car regulations, and it is difficult to achieve mass production of passenger cars. Therefore, mechanical lidar is generally only used as the main radar for L4/L5 autonomous driving test or operating vehicles.
HDL-32E turns like this
2019 is the year that the “mechanical lidar” will break China.
This year, the American lidar company Velodyne sued the Chinese lidar startup Hesai Technology and Sagitar. According to court documents, the company sued Hesai Technology and Sagitar Juchuang that the products being sold infringed on many aspects of its No. 7969558 US patent (High Definition Lidar System), which was granted to Velodyne in 2011. The founder of David Hall.
But fortunately, the “558” patent mainly restricts the follow-up research and development of mechanical lidar by other manufacturers. However, domestic manufacturers headed by Sagitar and Hesai have already deployed semi-solid lidar first. Although helpless, this happens to be the trend.
Semi-solid lidar, because there are fewer rotatable parts, the more stable it is, and the lower the manufacturing cost. For example, in the rotating mirror solution, its transceiver module remains stationary, and the motor reflects the light beam to a certain range of space while driving the rotating mirror to move, so as to realize scanning detection. The micro galvanometer solution uses a high-speed vibrating two-dimensional MEMS micro galvanometer to achieve scanning and measurement of a certain range of space.
However, because of the different scanning methods, there are certain differences in the data presented by lidars of different technical routes. An autonomous driving solution provider told Huxi, “The low-cost and stable lidar has limited coverage, and some of the products will cause heavy scans due to the complexity of the scanning mode (multiple scans in the same direction in a short period of time). Shadow problems, causing inaccurate ranging.”
At present, the Xiaopeng P5 with a starting price of less than 200,000, is equipped with two HAP lidars from DJI Livox, which belong to a type of rotating mirror semi-solid lidar. The working principle is that the laser is refracted by a rotating prism. Modeling is done by changing the light path to scan to more places. Because the number of lines of laser emission and reception is reduced, the material cost is greatly reduced.
However, HAP Lidar also has a shortcoming in that it lacks real-time performance, and the point cloud density will be affected by the scanning time. It often shows the highest density in the middle and gradually lower point cloud characteristics (in the shape of a chrysanthemum). In order to solve this problem, Lanwo adopted algorithm adaptation so that this lidar can achieve the equivalent of 144 lines with an integration time of 0.1 seconds, which is no different from the effect of mechanical lidar.
The difference in hardware does not mean that the functions that will eventually land will lag behind.
You should know that the HAP lidar used by Xiaopeng P5 is not the main source of decision-making data, but as a data supplement to the front environment to improve the safety of the vehicle during advanced assisted driving.
This is like power batteries. On the one hand, car companies and manufacturers are improving the performance of low-cost products, and on the other hand, they will also seek to improve the performance of low-cost products.idealThe ultimate solution. Hesai Technology told Huxi: “There are no moving parts in pure solid-state lidar designs such as Flash and OPA. Theoretically, the volume can be reduced to the smallest of all solutions. It has always been considered the ultimate form of vehicle lidar.”
However, due to the current technical limitations of basic components, the product detection range of the Flash solution is far inferior to that of MEMS or mechanical, and it is mainly used as a blind-compensating radar; the OPA solution is still a long way from mass production, and the related prototypes are still very much in progress. Early development stage.
2. Cost dispute: Cost is reduced by a hundredfold, selling cabbage price is not a dream
As early as 2017, Audi released the world’s first mass-produced L3 self-driving car, the Audi A8L, which was also the first mass-produced car equipped with lidar. It was jointly developed by parts manufacturer Valeo and German sensor technology company Ibeo. The SCALA lidar on the A8L is the “originator” of semi-solid lidar.
The first mass-produced car A8L equipped with Lidar
Of course, not every car company has such a strong appeal as Audi, allowing suppliers to develop the most advanced lidar for it. In addition, users who buy millions of luxury brand models are not sensitive to prices. However, when the price of the model is reduced to 300,000 yuan, the cost difference caused by each additional lidar will be particularly eye-catching.
The early lidar market was mainly used by technology companies regardless of cost for autonomous driving tests. For example, in 2016, a 16-line mechanical rotary lidar manufactured by the American manufacturer Velodyne sold for about US$8,000 (approximately RMB 51,000), and a 64-line lidar manufactured on the same principle sold for US$80,000. (Approximately RMB 510,000).
At the beginning of this year, IFC issued a research report “Heavy volume is imminent, Lidar will start the first year of pre-installation”, saying that it has been confirmed that pre-installed mass-produced models equipped with Lidar will exceed 300,000 units in 2023, and the price will be concentrated at 400,000-800,000. Yuan. But in the long run, the lidar for high-end autonomous driving in the future will gradually control the cost of the vehicle to less than US$1,000.
According to Huawei’s plan, the price of its equivalent 96-line lidar will drop to US$200/unit, and the future goal is to drop the price to US$100/unit to solve the cost bottleneck of Lidar boarding, which is L3 and above. The mass production of level-of-self-driving cars is possible.
“Lidar is a precision sensor integrating light, machine, and electricity. Its core components are not only computing chips, but many core optomechanical components and technologies do not follow the’Moore’s Law'”, Sagitar. Tiger sniff said. There are hundreds of subdivision devices in the lidar. In the production process, the material cost and equipment debugging cost are high; in the product structure, the mechanical part increases the volume and weight of the radar.
If you want to reduce costs, you need to change from the bottom chip. “The chip upgrade can turn the cost of lidar and the thorny problem of mass production into a ‘semiconductor’ problem.” Hesai Technology told Huxi.
In fact, in the BOM cost (material cost) of lidar, the laser transceiver module, which occupies the bulk, includes electronic components such as lasers, detectors, lasers, laser drivers, and analog front ends. By integrating these laser components on the chip, the cost of materials and the cost of installation and debugging can be reduced. At the same time, the cost of the chip can be further reduced by continuously improving the semiconductor manufacturing process.
Different from the “price war” in the conventional sense, the practice of reducing prices through “chips” not only did not sacrifice the performance of the product, but also “by the way” brought a series of performance and functional improvements.
For example, Sagitar’s “Intelligent Solid State Lidar M1” is based on two-dimensional MEMS smart chip scanning technology. Not only the internal structure is more streamlined, the components are less, and the integration is higher, but it also has a “gaze” function that can dynamically adjust the resolution and refresh rate.
Sagitar also told Huxi: “In the MEMS solution, the two-dimensional MEMS intelligent scanning chip has a relatively high technical barrier. The MEMS chip we use is completely self-developed. The chip has passed the verification of vehicle regulations and is currently in the process of large-scale vehicles. The planned mass production will be delivered to designated customers with M1.”
But as we all know, chipization is a long-term and extremely prone to failure. Therefore, in theory, only when manufacturers ship in large quantities and automakers apply in large numbers, can the performance and reliability of new products be verified by the market, and in turn promote the iteration of new technologies.
To put it bluntly, this is the same as the power battery back then. The early lithium iron phosphate batteries have not been optimistic because their performance is not as good as ternary lithium, but because of their lower cost and higher safety, after Tesla and BYD both adopted lithium iron phosphate, they promoted the scale effect. . More and more manufacturers are going to solve its shortcomings through technical means.
3. The battle for scale: the game between car companies and manufacturers
“Scaling is one of the core conditions for cost reduction.”
Sagitar Juchuang also stated to Huxi: “The breakthrough and maturity of lidar technology is based on product design, streamlining the structure, improving manufacturability, and reducing the material costs and manufacturing costs required for mass production. Large-scale forward The installation order is the trigger condition for the lidar to enter the mass production stage of the car, and the large-scale effect of mass production will also bring the cost of lidar components and production and manufacturing costs down.”
The automobile is an industry where the scale effect is particularly obvious. The power battery has been proven, and now it is in the lidar industry.
“Cost and scale have always been in a dynamic balance. Cost reduction has driven large-scale applications, and large-scale applications have reduced costs. Both are in positive feedback. This can be seen from the decrease in the cost of lidar in recent years.” Zhang Yu, director of perception of Qingzhou Zhihang, told Huxi.
However, lidar is still in the early stage of the passenger car market, and there will be a certain game between car companies and manufacturers that have just started cooperation: on the one hand, car companies want to purchase high-performance lidar at a lower price; on the other hand, manufacturers only have The price of lidar can be lowered only after the continuous increase in procurement volume has formed a scale effect. Therefore, in the initial stage of cooperation, both parties must reach a more consistent understanding of price and scale.
Pre-installed mass-produced semi-solid lidar AT128 (picture source: Hesai Technology)
Large manufacturers generally lock in the price and production capacity of lidar through investment.
The last major financing time in the industry was on November 16, Hesai Technology announced that it had obtainedMilletWith the additional financing of US$70 million in production and investment, plus the previously announced financing of more than US$300 million, Hesai’s D round of financing has exceeded US$370 million. Relevant statistics show that this is the largest single financing in the domestic lidar field since 2021.
This is also the second time Xiaomi has blessed Hesai since June. Other leading investors in this round of financing include Hillhouse Ventures, Meituan and CPE. Since June, Hesai Technology has reached cooperation with at least 12 OEM and autonomous driving R&D companies, including Ideal Auto, Lotus, Jidu, AIWAYS, and Chinese Express. In addition, lidar manufacturers such as Yijing Technology, Innovusion, Leishen Intelligent, and Tanwei Technology have also received a new round of financing.
On the OEM side, Ford,DaimlerMultinational OEMs such as Volvo and Volvo will basically select at least one manufacturer for investment. for example,Wei LaiThrough NIO Capital, it has invested in Innovusion’s Series A, Series B and Series B+ financing for three consecutive times. Innovusion’s B+ financing announcement also wrote: “This round of financing will be mainly used to support the mass production and delivery of the Weilai ET7 lidar for mass production.”
Obviously, the lidar “goods grab war” is intensifying.
But this kind of involution is beneficial in the eyes of industry professionals. “This is a bilateral cooperative relationship. Both parties promise something, and then everyone builds this trust. Then you can do it. In fact, whoever throws in the investment at the beginning, as long as one party throws it away, Then turn it around.” An insider of an autonomous driving supplier told Huxi.
“We have also been staring at MEMS lidar, and they (lidar manufacturers) said that if you reach 100,000, I can give you two or three hundred dollars. If you have hundreds of thousands, I can totally do 100 dollars. Do it.” The insider of the above-mentioned autonomous driving supplier told Huxi.
According to Sullivan’s forecast, the global lidar market will reach USD 13.54 billion by 2024, with a compound growth rate of 64.65% from 2020 to 2025. Analysts believe that with lidar performance advantages + price drop + downstream demand urgent, three-factor resonance promotes lidar to become the mainstream solution for mass production of supporting vehicles.
In terms of current market share, there are five Chinese manufacturers in China that are more prominent, but French manufacturers still dominate.
According to the data of consulting company Yole Développement, the market share of lidar manufacturers in the automotive and industrial markets is ranked, Valeo ranked first, accounting for 28%; Sagitar RoboSense accounting for 10%, ranking second; Luminar, DJI Livox,DensoFive manufacturers, China Mainland, and Cepton, tied for third place with 7% share; Innoviz, Ibeo, Innovusion, Huawei, Hesai Technology, Innovusion Tudatong, and Velodyne each had a share of 3%.
Fourth, the battle for functions: the dilemma of car companies
The mass production of lidar on the car is just the beginning of the realization of autonomous driving.
This is like preparing a table of good ingredients, but whether the good ingredients can be cooked into a table of delicious dishes depends entirely on the level of the chef.
“Now the software capabilities of car companies must not be able to keep up. This is very certain.” An insider from an autonomous driving supplier told Huxi. At present, the industry has a relatively unified understanding of automatic driving functions: first leave enough hardware surplus, and after mass production is on the car, the functions are continuously upgraded through software OTA.
Take the Xiaopeng P5 equipped with 2 lidars as an example. When it was delivered to users in the early days, it did not have the “urban NGP” function mentioned in the promotion-based on navigation, it can be used in some urban scenes. Realize automatic driving, such as the function of “turn left without protection in mixed traffic between people and vehicles”, and also realize “voice lane change”.
On the 1024 Technology Day, Xiaopeng showed the actual measurement video of the city’s NGP on P5: the route is 15 kilometers in length, starting from the parking lot, passing through the downtown area of Guangzhou, and finally reaching another parking lot. Basically, this can be understood as “urban autonomous driving”, but considering issues such as laws and regulations and user safety, conscious companies will still be adequate when propagating.
Polar Fox Alpha S Huawei HI Edition
Many car companies are talking about the “DOOR TO DOOR” automatic driving function, including the Polar Fox Alpha S Hi version equipped with Huawei’s automatic driving technology. But to this day, there are only a handful of actual vehicle test videos. Although everyone is criticizing Tesla’s unreliable autopilot, at least people in the United States have achieved urban autopilot capabilities through OTA upgrades.
“Tesla has always been like this. He installed very powerful hardware first, and then iterated the software. This is the case in this industry. The hardware iteration cycle is very slow. If you don’t install it now, you may be behind in one or two years. But. , The software can be OTA at any time.” The insider of the above-mentioned autonomous driving supplier said. In fact, many traditional car companies have only learned the first half of Tesla, but not the second half.
Going back to the realization path of autonomous driving, the core thing is to solve three problems: “Where am I?”, “Where am I going”, and “How am I going”.
According to the jargon: perception, decision-making and execution. More vividly, the perception layer is equivalent to the five senses of a person, perceiving the surrounding environment, collecting data and transmitting it to the decision-making layer; the decision-making layer is equivalent to the human brain, processing the data transmitted by the perception layer, and outputting corresponding operating instructions to the execution layer ; The executive layer is equivalent to the limbs of a person, and executes the instructions given by the brain.
Lidar, because it has the advantages of accurately acquiring three-dimensional information of the target, high resolution, strong anti-interference ability, wide detection range, and near all-weather work, it occupies an important position in the intelligent driving environment perception system. However, lidar is only a sensor for data collection, and does not have the ability to make decisions and execute. In other words, Lidar just assumes the role of “eyes”.
Clairvoyance is not enough. The key is to make good use of the data generated by lidar, that is, to rely on algorithms. In essence, all kinds of autonomous driving functions rely on the development of specific application algorithms. The environment of autonomous driving is complex and changeable. It is the ultimate goal of Lidar application algorithm development to accurately and quickly extract effective data from the complex lidar point cloud data, and to correctly understand and analyze useful information.
But there are still many problems. For example, at present, there is no unified framework and judgment standard for the application algorithm of autonomous driving lidar, which has strong pertinence and certain particularity. Often the higher the accuracy, the worse the adaptability, and the use range is quite limited. In the face of various complex and changeable autonomous driving scenarios, it is necessary to make the algorithm expandable and portable, and it is particularly necessary to improve the self-adaptability of the algorithm.
In addition, there are data-driven functional iterations, which are also a hurdle to test car companies.
“Perception and planning are all driven by data. The advantage of data-driven is that in the intelligent driving system, the order of magnitude of corner cases encountered is more than one million, which is impossible to optimize by engineers alone. Therefore, truly powerful intelligence Driving must be driven by data, and iterate and optimize a large number of scenarios that are small in probability but will be encountered through a data-based method.” Zhiji Automobile Co-CEO Liu Tao told Huxi.
But now it’s fine. It doesn’t matter if traditional auto companies don’t have software capabilities, as long as they have money.
“Many car companies don’t even have the software capabilities themselves, so they directly tell the lidar supplier that you just give me the algorithm and finish it. Then the lidar supplier may not have the algorithm, and they may come to us in the end.” The above-mentioned automatic Insiders of the driving supplier said.
This is not surprising. In the field of ADAS assisted driving,IntelIts autonomous driving company Mobileye once occupied 75% of the market. The reason for success is that it is not only a popularizer of visual sensors, but also a complete solution provider of “visual sensors + computing chips + intelligent algorithms”. This kind of cooperation model that is close to “plug and play” has been greatly affected. Loved by car companies.
At the GTC conference in 2021, NVIDIA released a completely self-designed intelligent driving solution-NVIDIA DRIVE Hyperion 8, which is a computer architecture and sensor set for fully automated driving systems, it is equipped with NVIDIA self-developed chips, NVIDIA Recommended cameras, NVIDIA recommended radars, and even NVIDIA standard development kits.
In addition, NVIDIA also directly demonstrated the perception algorithm developed by the team, and also mentioned its “high-precision map surveying and mapping capabilities.” Although NVIDIA did not directly tell the car companies that we can provide you with a full-stack autonomous driving solution, it basically shows that NVIDIA has the ability to be a full-stack supplier of autonomous driving.
In the early days, NVIDIA only provided chips to manufacturers, and now it has opened up various fronts. In fact, it is also forcing domestic chip manufacturers to provide more software services. For example, in the previous paragraph, the AI chip company Black Sesame Intelligence, which has just received investment from Xiaomi, told Huxi: “Because software, algorithms and supporting development support are important indicators that reflect the ease of use of chips. The current domestic chips are compared with international chips. Brand, its advantages are mainly reflected in flexible and efficient support, personal support to customers for development and services.”
“Whether the development tool chain is complete is an important indicator of the ease of use of the chip. In conjunction with the Huashan series of autonomous driving computing chips, Black Sesame Intelligence also released the Shanhai artificial intelligence development platform. It has more than 50 AI reference model library conversion use cases, reducing customers Threshold for algorithm development.”
In general, it is not important to have full-stack software self-research capabilities. Lidar manufacturers, autonomous driving solution providers, and autonomous driving chip manufacturers will always help car companies to release the hardware value of lidar. In essence, it depends on whether car companies are willing to spend money.
Waymo’s test car lidar cleaning method
In the end, there is a crucial issue that needs to be solved by the car companies themselves: “How to maintain the delicate lidar?”
A person in the autonomous driving industry expressed this concern to Huxiong: “The anti-fouling problem of lidar and subsequent maintenance problems, car companies currently do not provide similar solutions. For example, if it is dirty, how should the car owner wipe it clean; Yes, does the car owner pay a lot of money for it?”
Getting on lidar is just the beginning of the trouble. But if everyone gets on you but not on, it is also the beginning of trouble. The involution of the lidar industry stems from the deformed involution of the car industry.
Write at the end
“Right now, all trades and industries are very serious. Every consumer will have his own choice. We can’t make all consumers like us, but we are the first to get out of the inner roll and do our own unique things well. “Xia Heng, co-founder and president of Xiaopeng Motors, said at the recent Guangzhou Auto Show.
The “involution” of lidar has already formed, which is bound to be a good thing for consumers. But at the same time, more selectivity means that there may be mixed situations. Therefore, in addition to the hardware configuration, it is even more necessary to recognize the software capabilities of a car company. At the very least, you should also be responsible for the user’s real experience, not the cottagecell phoneSimilarly, rely on the number of cameras and pixels to attract consumers.
As Liu Tao, the co-CEO of Zhiji Automobile, said: “High-tech should not be a show of technology, but a real service for the user’s driving experience.”
Domestic car companies, don’t let Lidar become the leek harvester in 2022.