Image source @ Visual China
The | is deep, and the author | Dawn, edited by | Ai Xiaojia
Lidar is a bit mythical.
Since last year, the attention of lidar in the automotive circle has increased significantly. This year, the heat has increased unabated, and the high-end new energy models newly listed in major cities are almost standard with lidar. It seems that without this part, it is not worthy of being called a smart car.
On April 18, Jidu secretly announced some details of the concept car, and the location of the two lidars was exposed. It is such a detailed problem that it actually attracted two CEOs in the new forces - He Xiaopeng and Li Xiang, plus Jidu CEO Xia Yiping, and the three of them had a battle online.
The car has not yet been mass-produced, the algorithm is still under development, and more key information has not been disclosed, but the lidar has preemptively "got on the car". The atmosphere was quite well set off.
"The marketing method of the automobile factory, the labeling of the brand." An insider at an autonomous driving vendor commented on Deepway.
Lidar, simply and crudely understood, is a device that emits a laser, reflects a bounce back, identifies an object, and detects a distance. The principle is not complicated.
At this stage, the so-called smart car, can it not be installed with lidar?
OK.
Tesla is not installed, relying on pure vision, but everyone will not doubt Tesla's level of intelligence. According to CEO Musk, "Humans don't drive lasers by their eyes. ”
So, with the installation of lidar, does it play a big role at this stage?
Not too big.
Xiaopeng P5 is equipped with two lidars, and the main role at present is to fill the blindness, which is not a core role in the entire perception system. The radar of WEI's ET7 claims to be able to achieve ultra-long-distance detection, but the current function has not been unlocked, and it is necessary to wait for the OTA upgrade. As for other car brands that follow the trend, the lidar model is released, but the mass production delivery is still a matter of no spectrum.
Lidar is very good and powerful, but the current product maturity is not as high as the outside world expects. Expecting too much is sometimes not a good thing.
Zhang Lichen of solid-state lidar supplier Tanwei Technology told Shentu that from the perspective of industry history, the speed of lidar progress in recent years is still quite fast, "Maturity of this piece, I think there is a 2-3 years of scale mass production polishing, should be able to give a satisfactory answer." ”
Until then, the road that lidar has to go, the pit to step on, and much more.
Who is touting lidar?
The touting of lidar is driven by the new forces of car manufacturing.
At NIO Day last January, NIO released its fourth production car, the ET7. At that time, the car was very hot, in addition to Wang Feng pushing a hot search on the site platform, another point of concern was that ET7 was equipped with a lidar.
This is the first officially released lidar model among domestic automakers, and it will also win the title of "the world's first lidar mass production car".
The ET7 was bold, not only attaching the lidar to the roof, but also making it bulge out to look like a watchtower. Earlier, the Audi A8 and Mercedes-Benz S-Class both hid it.
Lidar on the roof of the NIO ET7
This bold design makes people with obsessive-compulsive disorder unacceptable and feels abrupt and ugly. But there are also people who like it, feeling avant-garde and high-tech. One positive and one negative, controversial and topical, coupled with weilai traffic, lidar is on fire.
However, it is interesting that the car company that finally won the title of "the world's first lidar mass production car" is not Weilai.
Three months later, at the Shanghai Auto Show, some restless car companies began to keep up and also launched lidar models. This includes Xiaopeng.
Xiaopeng P5 lidar on the left side of the front face
The Xiaopeng P5 is directly on two lidars, one on each side of the front face. The point is that the P5 began delivering in October last year, a step ahead of the NIO ET7. The title of "world's first" was snatched away by Xiaopeng.
From the beginning of the title grab, the wind of lidar was blown up.
Lidar does have some advantages over cameras and millimeter-wave radar.
"Lidar on ADAS can definitely help vehicles identify some of the more visually illegible corner cases (extreme scenarios) on high speeds, such as special-shaped vehicles and large white trucks, which is definitely a safety plus." Zhang Lichen's analysis of the deep way.
In just over a year, more than a dozen car companies have intensively launched nearly 20 new models equipped with lidar. From the new car-making forces such as Weilai, Xiaopeng, Ideal, Weima and Nezha, to the new brands launched by traditional car companies such as Zhiji, Jihu, Eian and Mocha, lidar is listed as the standard of new models, and at the same time it has become a selling point when promoting externally.
In particular, Wei Xiaoli, which best represents the level of intelligent car manufacturing in China, and the next-generation models with innovative significance - Weilai ES7, Xiaopeng G9, Ideal L9, and Jidu, which is known as "automotive robot", are equipped with lidar.
The set of dual lidar on the front cover of the car
It is still difficult to quantitatively assess how much lidar has improved the performance of vehicles, but it is clear that the outside world has gradually put the words lidar and intelligence into a situation.
He Yuhua, the founding partner of Hegao Capital, found in contact with the main engine factory that domestic car companies are very aggressive in using lidar. "Everybody wants to differentiate." He said to Deep Way.
Lidar has become the brightest boy in the car circle.
Lidar was mythologized
There is no doubt that lidar is a smart product. It is a key component of the car in the realization of high-level automatic driving, together with the camera, millimeter wave radar, constitutes the perception system of a smart car, becoming the "eyes" of the car.
But has a car equipped with lidar become a smart car?
not necessarily. At least for now, lidar is not as powerful.
Lei Peng, senior commercial director of Weigan Technology, told Shentu: "In the L2 stage, the role of lidar is not experienced by drivers, mainly for visual and millimeter wave redundancy. ”
Taking the "world's first lidar mass production car" Xiaopeng P5 as an example, the camera and millimeter-wave radar are still the most important sensing devices. Lidar is not a charcoal in the snow, it is the icing on the cake.
Zhang Lichen revealed to Shentu that the two lidars of P5 are from DJI's company Lanwo, which has low cost and fast mass production, but its irregular scanning characteristics cannot meet the perceived needs of high-speed driving. P5 mainly uses it for data supplementation and autonomous parking, not the main source of decision-making data.
Therefore, the outside world may have exaggerated the role of lidar on P5.
Weilai's strategy is different from Xiao Peng's, wanting to eat a fat man in one bite. ET7 is equipped with Tudtong 1550nm lidar, the performance is very good, just released last year is the "farthest detection distance of 500 meters" label, at 10% reflectivity detection distance can reach 250 meters. The industry has high expectations for this product.
However, an industry insider who had just finished test driving the ET7 told Shentu that Weilai sales told him that the NAD upgrade date was expected. "The lidar installed now is a hardware decoration for the time being."
He Yuhua told Shentu that the current function of lidar has not been well developed, and many cars equipped with lidar do not have strong automatic driving capabilities when they leave the factory, which belongs to the "castrated version". The story told by these OEMs is that the vehicle is pre-installed with hardware, and the software is upgraded through the OTA stage, so that the owner does not have to buy another car for the latest self-driving features. But in fact, there are many restrictions on OTA upgrades, and the specific extent of the function is unknown, but the owner needs to pay in advance.
Not only are passenger cars immature, but the L4-class Robotaxi and Robotruck are actually immature.
In the past, lidar has been used in L4 level autonomous driving tests. These test vehicles do not have to consider the size and appearance of the problem, the use of ultra-high-performance lidar. But even so, there are still many corner cases (extreme scenarios) that cannot be overcome, which in itself shows that lidar is not a panacea.
"Take Robotaxi, for example, the sensors are armed to the teeth, the actual experience is still not satisfactory, and the commercial gimmick is much greater than the actual." Zhang Lichen said.
Qin Dongyong, product manager of an autonomous driving company, told Shentu that some autonomous driving companies that do closed or semi-closed scenes have "suffered losses" in lidar.
He gave three typical examples, one is that the ranging level within the entire field of view is not the same, such as the 200-meter ranging ability on paper, only 100 meters at the edge of the field of view, but the manufacturer will not say clearly, it needs special testing to know; second, in the case of sunlight interference, the noise level of the lidar point cloud will be very different from the theoretical value; third, when encountering high reflectivity objects and low reflectivity objects, there may be high anti-ghosting and high anti-expansion.
Lidar appears "highly anti-inflation" / courtesy of respondents
"These problems may not be tested in the early stage of testing, which may be a friendly time period or environment, but in the real mass production scenario, they are all exposed." He said.
The losses suffered by autonomous driving companies on lidar, and new car-making forces and new brands will also face in the future.
Now, for passenger car NOP/NGP, the industry mainly uses the L2+ function. L2+ does not use lidar, and the software and hardware have a relatively mature mass production program. Lidar on the car, but also only for the sensor to provide a redundant solution, a bit of downgrading the use of meaning.
Qin Dongyong's analysis of the deep way: "At this stage of lidar in passenger cars, it can be understood as: 1, hardware embedding; 2, laser effective data collection provides data for subsequent multi-sensor fusion and model training; 3, gimmicks." ”
It's also too early to get on the bus in bulk
Car companies buried lidar in advance, which is to prepare for future upgrades, so that the hardware capabilities cannot keep up when they do L4 automatic driving in the later stage.
But another problem has emerged, the current lidar, whether in the application side or the industrial side, is not mature enough.
To mass-produce and install lidar, it is necessary to meet multiple elements such as performance, volume, cost, and safety at the same time, and no one aspect can have too many shortcomings. But the reality is that today's lidar is difficult to take care of at the same time.
Taking the Xiaopeng P5 as an example, Lanwo's plan solves the cost and sacrifices performance. Therefore, in the next car G9, Xiaopeng plans to cooperate with Sagitar Juchuang and replace it with a higher performance lidar.
The Tudtong 1550nm lidar used by Weilai ET5 and ET7 has no choice in detecting the distance, which is less harmful to the human eye, but the cost is higher, and it protrudes from the roof, which will affect the wind resistance coefficient of the vehicle. Zhang Lichen said that 1550nm has always had the problem of low yield rate in mass production.
This has left a problem for car companies, to load the car in batches, you have to make trade-offs, but also have to bear a certain risk.
Looking deeper, this is not a problem for a single lidar manufacturer, but a problem for the entire industry. Lidar has a technical route battle, each with its own advantages and disadvantages, but none of them can be called the perfect solution.
The industry started with mechanical lidar, which can rotate and has a horizontal field of view scanning range of up to 360 degrees. However, mechanical lidar has a fatal flaw: the internal rotating parts are easily damaged, and the stability is relatively low, which will affect driving safety.
Mechanical LiDAR on domestic driverless taxis source / tech otaku
So the industry figured out another route - solid-state lidar. There are no rotating parts in the laser, and the stability and service life of the lidar are improved, which is considered the optimal solution. But the problem is that solid-state lidar is not yet technically possible.
Finally, the industry took a compromise solution, launched a hybrid solid-state (semi-solid-state) lidar, neutralized the advantages and disadvantages of mechanical rotary lidar and solid-state lidar, more suitable for large-scale mass production, and has now become the mainstream scheme of the industry.
On the mixed solid state, it differentiates into different directions, such as micro-galvanometers, uniaxial rotors, prisms, etc. The representative models of the micro-galvanometer direction are Xiaopeng G9, GAC AION LX, WM M7, and the representative models of the single-axis rotator are Polar Fox Alpha S and Great Wall Mech Dragon. However, none of these models have yet to be mass-produced.
LiDAR on the roof of the WM7
Until now, the entire industry has not found the optimal solution for lidar to get on the car. The hybrid solid-state route that car companies are now advocating is actually just a transition.
Vehicle regulation certification is another big mountain before the mass production of lidar. Lee wanted to question where set-degree lidar was installed, on the grounds that it didn't pass pedestrian collision regulations.
Zhang Lichen told Shentu that the vehicle regulation system is more complicated, and there is a unified requirement for temperature, vibration, and production line standards, but there is no specific standard for the single product of lidar.
"European and American countries have strict requirements for car regulations, resulting in a very low number of lines or a short distance, which can only be used to make up for blindness, which does not play a decisive role in the upgrade of the automatic driving system." Large and small batch samples of domestic steps can be used first after the stability test, directly on the high line number of long distances, stability will take time to repair. He said.
Domestic lidar is so radical that no manufacturer has produced more than 100,000 units a year, not to mention the stability of large-scale mass production. "The lidar barrier is still relatively high, and it is not overnight to make a vehicle-grade product." Qin Dongyong said.
Cost issues also need to be considered. Compared with the visual solution, the current price of lidar does not have any advantage.
"The cost is too high, the popularization is difficult, it can only be used in high-end cars and high-end equipment, and the models that are generally shipped are still without lidar." DVDs were expensive when they first came out, and the price dropped sharply after the popularity was large, so let the bullets fly a little longer. An insider of a domestic computer vision listed company told Shentu.
At the end of 2020, Huawei released a 96-line lidar, claiming to be able to produce 100,000 sets / line per year, which is installed on the Polar Fox Alpha S Huawei HI version. But until now, the car has not been mass-produced, and Huawei's lidar has not been delivered in batches.
"From the maturity of the industrial chain and the landing progress of the application end, the real large-scale mass production has landed two or three years later." Zhang Lichen analysis.
Will there be lidar in the ultimate solution?
Whether or not lidar is installed, for car companies, it is not only a multiple choice problem, but also a mathematical problem.
If it is simply to pile up hardware, lidar will not have such a high heat and will not be mythologized.
"I think the problem with algorithms is much bigger than perception." Zhang Lichen said.
Perception, planning and control are the three parts of the automatic driving system, corresponding to human driving, that is, the eyes see the road, the brain commands, and the hands and feet operate. Lidar belongs to the perception device, equivalent to one of the "eyes" of the car, the other "eye" is the camera, but the imaging principle and data form of the two "eyes" are different, so that the "brain" can read, it is necessary to do data fusion.
Multi-sensor fusion, the computing power, algorithm requirements are very high, many car companies installed lidar, in fact, do not have the ability to do fusion.
Musk is a staunch opponent of lidar. Tesla has always insisted on using a pure visual perception scheme of cameras, not only without lidar, but even with millimeter-wave radar. In this route, Tesla does not have to do complex multi-sensor fusion, but it is very demanding on the visual perception system.
Tesla source / unsplash
The mainstream domestic route is a multi-sensor solution equipped with lidar. The industry generally believes that higher-order automatic driving above L3 is inseparable from lidar.
Who is the ultimate solution of these two lines is still inconclusive.
Proponents of Tesla believe that lidar will eventually be removed, and multi-sensor fusion is not as good as pure vision. "Lidar and HIGH-precision maps are the two crutches of smart cars, high-precision maps are responsible for unfolding the 'battlefield fog' in advance, and lidar is responsible for avoiding obstacles and filling blindness in complex scenarios, but with the evolution of CMOS and computing power and algorithms, pure vision must be the only choice for Wei Lipeng in the end." The industry's big V "Blood Flag" publicly stated on Weibo.
If Tesla wins, there is no lidar position in the ultimate solution, which means that lidar is just a transitional product, Wei Xiaoli will have to correct the route in the future, and the lidar market will also face a revaluation.
The industry generally supports Wei Xiaoli's multi-sensor fusion route. Many people in the industry have expressed a view on the deep way: vision, millimeter wave, lidar, the three are to be used together for a long time, to achieve high-level automatic driving, lidar is indispensable.
He Yuhua pointed out that the visual scheme is flawed in many scenes and will soon reach its ability ceiling. For example, in the case of assisted driving in and out of the tunnel, the optics have changed greatly, the visual solution is basically invalid, even Tesla, can only maintain 20% of the ability.
He believes that in the ultimate solution, between the major sensing components of assisted driving, there must be a relationship of timely scheduling and redundancy, rather than a relationship of primary and secondary. The system will dynamically call different sensors according to the changes in the scene, and constantly do the allocation of parameter contribution values. "The future must be a comprehensive perception scheme, and these sensors need to confirm each other, which is indispensable."
However, the multi-perception scheme greatly increases the difficulty of the algorithm and puts forward higher requirements for the software capabilities and hardware adaptation capabilities of car companies. "In addition, the wiring harness layout and data transmission volume in the car will change, and the computing power of the platform will increase accordingly." He Yuhua said.
Therefore, in the future, the real gap between driving companies is not whose lidar is more advanced, but depends on the algorithm, computing power and system of the vehicle. Lidar, can not fight alone.
The accelerated development of lidar is the result of the joint promotion of players from all sides of the industry. It is powerful and promising, and should not be ignored, but it should not be mythologized either.
Any industry needs to be ahead of the curve. For innovation, we should be more tolerant and rational. Only by looking at lidar rationally and objectively can the industry develop in a healthy and orderly manner.