Head image source: Stills from "Kung Fu"
On the road of car intelligence, the autopilot system led by Tesla is undoubtedly the focus of global attention. Musk once said that Tesla's artificial intelligence is the most advanced in the world.
However, in the Chinese market, car companies and head autonomous driving companies have expressed dissatisfaction. For Tesla, they put on a posture of catching up later, localization advantages easy to defend and difficult to attack.
"In the urban NOH, the product force of the end is ahead of Tesla's performance in China." A few days ago, Gu Weihao, CEO of Zhixing, announced at a press conference.
He Xiaopeng also said at the 2022 China Electric Vehicle 100 Forum that "the city-level NGP effect pushed out during this year will be quite better than the Tesla FSD effect." In fact, He Xiaopeng shouted tesla 2 years ago, "You must have the mind to be prepared to be beaten by us." ”
WM recently released its L2-level advanced driver assistance system Living Pilot, and Shen Hui, chairman of WM Motor, believes that this system understands Chinese better than Tesla, which has achieved full automatic driving.
Previously, Liu Tao, co-CEO of Zhiji Automobile, also said in an interview with Auto Market Story, "We admit that Tesla is leading the world in the field of automatic driving, but Tesla is arrogant, and we are confident to create a better intelligent driving experience than Tesla." ”
Xiao Yong, deputy general manager of GAC New Energy Automobile Co., Ltd., declared that in terms of intelligent driving, GAC New Energy products have surpassed Tesla. He has mentioned on many occasions, "Tesla is not a myth that cannot be broken, we can use it as a benchmark to surpass." ”
In the field of autonomous driving, Tesla has long become a radical representative. Back in April 2019, Tesla launched FSD (Full Self-Driving) on top of autopilot's assisted driving feature. Musk recently stressed that "the most urgent project in the short term is complete automatic driving, so that autonomous driving operations can solidly exceed the safety level of human driving." It can be done this year. ”
Why is it that Tesla, which started earlier, has stronger research and development strength, and has a higher level of automatic driving, but is frequently "hanging" in China? Is the Tesla FSD really "hanging"?
#01
Tesla VS Other:
Factional arguments that are difficult to win or lose
One of the debates between Tesla and a number of Chinese car companies on autonomous driving is the debate between the two major factions of pure visual perception and multi-sensor fusion perception.
Tesla is a major proponent of visual perception schemes. Tesla believes that the data collected by the camera is most similar to the real world perceived by the human eye, closer to the human driving pattern. When visual perception is powerful enough, other sensors are meaningless.
Domestic car companies basically take the multi-sensor fusion perception route. They believe that cameras are susceptible to light, weather and other factors, detection accuracy may be reduced, and lidar can be used as a redundancy, which is essential for high-level automatic driving.
A few days ago, Li Xiang, the founder of Ideal Auto, expressed his views on the design scheme of lidar on social platforms;
Subsequently, Xia Yiping, CEO of Jidu Automobile, and He Xiaopeng, chairman of Xiaopeng Automobile, also joined in. The dialogue between the three new car manufacturers has brought lidar to the front of the stage and become the focus of the industry's attention.
Up to now, smart car brands including Xiaopeng, Weilai, Ideal, Zhiji and other intelligent car brands have released mass production models equipped with lidar. The head of the mech of the New Energy Salon under the Great Wall even shouted "How many lidar do you have?" If you have less than 4 pieces, please do not talk. ”
When tesla companies discuss several lidars, Tesla has even eliminated all radar technology, which means that Tesla's models will rely only on a visual system composed of 8 cameras to activate autonomous driving. This is simply incredible in the eyes of other car companies.
The reason why Tesla loves visual solutions, in addition to having a powerful set of pure visual ranging methods to replace millimeter-wave radar, cost issues are a major consideration for Tesla to build cars. It is reported that the average cost of a single camera currently used by Tesla is about 150 yuan, and the minimum lidar sought after by new domestic forces is about 200 US dollars.
Despite this, an autonomous driving practitioner told The Story of the Car Market, "In theory, it may be possible to collect data with only cameras, but there are many unknowns that cannot be avoided, and if you want to be 100% confident that the system's judgment is correct, it is best to integrate other sensing to assist, such as lidar." ”
Earlier, Su Zhen, the former head of Huawei's autopilot business, publicly said, "The data quality and dimension of Tesla's pure visual perception scheme may not be enough, and the room for progress of the matching system and algorithm will be limited." ”
In order to make automatic driving safer, most of the new domestic power brands are based on the use of vision + multi-sensor + high-precision map, they are ahead of the hardware configuration, because the hardware redundancy is stronger than Tesla FSD, it looks more secure.
Judging from the planned or launched car series configurations in 2022, 3 car companies have invariably configured 1-2 lidar. The Xiaopeng P5 has 4 millimeter wave radars and 2 lidars more than the Tesla Model 3.
Xiaopeng P5 lidar
The increase in sensors has also produced a huge amount of image data to be processed, taking NIO as an example, the NAD system is equipped with many camera sensors, which can generate up to 8GB of image data per second.
Due to the high degree of openness of Nvidia, the new forces and new car series are equipped with the latest chip Orin, and the computing power has been greatly improved on the basis of the original, and the computing power of xiaopeng G9 has increased from the 30TOPS of xiaopeng P5 to 508TOPS; the hash rate of WEILAI ET7/ET5 is the highest, reaching 1016TOPS; the ideal X01 is relatively low, 254TOPS.
The head of Baidu's autonomous driving research and development department said that the road environment in Beijing is more than 15 times more complex than that of California, and Beijing is only a microcosm of China's complex road environment, and if it comes to Chongqing in 3D, it is another kind of "complexity".
In the face of these complex problems, China's new car-making forces do have to spend a lot of time to overcome, and multi-sensor coordination is needed to achieve more redundancy. "What does Tesla do when it doesn't have radar sensors when it 'doesn't know' objects or situations?"
Of course, the advantages and disadvantages of automatic driving are not only based on the simple discussion of hardware, but also require continuous optimization and iteration of perception algorithms, planning algorithms, control algorithms and other aspects of car companies.
Recently, Toyota Motor's self-driving subsidiary Woven Planet also announced that it will explore vision-based autonomous driving solutions. Michael Benisch, vice president of Woven Planet, believes that as technology evolves, camera-based autonomous driving technology has the potential to surpass some self-driving technologies that use sensors such as lidar and millimeter-wave radar. At the same time, he stressed, "at the moment we cannot judge how long it will take to achieve this.".
#02
Tesla FSD:
The United States is easy to use, but China is not necessarily
The Tesla FSD beta version was opened up for testing in small batches very early, and now the iteration speed is gradually accelerating. In the first quarter of 2022, Tesla FSD was updated with 7 versions, and the current trial users are about 100,000. Tesla expects to push the FSD Beta version to all users in the United States on a large scale during the year.
However, even if FSD is good to use in the United States, it may not be easy to use in China. In stark contrast to Tesla's hot sales, Tesla's core technology, FSD, is very few. According to statistics, the FSD selection rate has declined year by year, with only 11% of global car owners buying, while in China, Tesla FSD purchase rate in China is only 1%-2%.
The low availability of FSD is largely due to its still in the futures concept. What do users get for $64,000?
First, the high-speed pilot function NOA (Navigate on Autopilot): In the basic function of ACC+LCC, automatically drive in and out of the highway ramp or overpass fork, exceeding the slow-moving vehicles; automatically assist in changing lanes on the highway. At present, competing with Tesla NOA are Weilai NOP, Xiaopeng NGP and so on.
Second, fully automated driving (FSD) may be realized in the future. Since 2020, Musk has more than once promised in public that he will achieve full autonomous driving within that year, but has not been able to achieve it.
Whether the existing NOA is easy to use affects the expectation of FSD. According to the analysis of the autohome iV-RATING intelligent car evaluation system, Tesla, the benchmark of autonomous driving in the traditional impression, is inferior to Chinese brands, and the comprehensive ranking is only eighth.
Although the Model Y scored the highest in the high-speed scene, when the speed was greatly improved, the automatic lane change function had a low false positive rate, the logic was crisp and the execution action was accurate, but it was not seen in the top five in the urban scene, parking scene and emergency scene.
In the domestic market, Xiaopeng Automobile is recognized as a car company with a fast layout of automatic driving, and it is developing an NGP function that can be applied to urban roads. This function enables road identification, automatic overtaking and lane change, speed limit adjustment, automatic entry and exit ramps, takeover vehicle reminders, and more.
He Xiaopeng introduced that after obtaining the approval of the relevant competent authorities, the urban NGP function will be launched in the first batch of cities at the end of the second quarter of 2022. He said that the urban NGP test results carried by the Xiaopeng P5 show that the number of takeovers per 100 kilometers is close to high-speed NGP.
Consistent with the rhythm of Xiaopeng City's NGP function landing planning, there is also the ultimate Zhixing. In November 2021, the Great Wall Wei brand Mocha NOH Smart Pilot Edition was officially launched at the Guangzhou Auto Show, which is the first mass-produced passenger car equipped with the NoH Pilot AssistEd Driving System.
According to the plan, Zhixing will deliver the city NOH in the first half of 2022, and the full-scenario NOH will be delivered in the second half of 2022. In the full-scene pilot auxiliary driving ability landing, the goal of The End Zhixing is half a year faster than That of Xiaopeng Motors.
The pilot auxiliary driving that took the lead in launching the urban scene is also the confidence of domestic car companies to dare to call Tesla. There is no clear timeline for the large-scale push of Tesla FSD in the United States, and when it can be realized in China seems even more distant.
Since Tesla does not have a R&D team in China to localize autonomous driving, FSD requires much more localization work than NOA.
"Tesla is currently the most core intelligent driving, in this regard we may be a latecomer, but China's traffic and driving habits, and the United States is completely different, our understanding of Chinese traffic must be deeper than theirs, this is the space we want to play," Liu Tao said of the car market in an interview, "on the other hand, such as policy requirements, such as high-precision maps." Relatively speaking, high-precision maps of highways and cities, Chinese electric vehicle companies will take some advantages than Tesla, and there are some dividends in terms of policies and regulations. ”
Starting from the design of user habits and special local scenes, Chinese car companies understand China's road conditions better. Local companies can use China's complex scenarios to train algorithms so that users feel like the system is smart and more like a human driver.
#03
Tesla is trapped in the "tight curse" of data collection?
Considering the sensitivity of the data, in 2021, Tesla announced that it has established a data center in The country, carried out desensitization processing, and all the data generated by vehicles sold in the Chinese mainland market will be stored in the country.
But today, there are still many scenes where Tesla private cars are explicitly required not to enter.
In terms of localization of autonomous driving, it may not only be Tesla that "suffers losses", but the problems that traditional overseas car companies will face.
For example, GM's autonomous driving technology capabilities are in the first echelon. As early as 2017, GM's SuperCruise system achieved highway let-off driving in the United States, and was one of the first companies to achieve L2-level autonomous driving that could be released; in terms of L4-level autonomous driving, GM also acquired cruise, a well-known self-driving startup, which has been considered second only to Waymo's self-driving company in the field of autonomous driving in the United States.
However, in areas without HIGH-definition maps, SuperCruise does not work well in China. In order to make autonomous driving better land in the Chinese market, in 2021, GM announced that it will invest $300 million in domestic autonomous driving startup Momenta.
Coincidentally, in China, Toyota has also invested in self-driving startups Xiaoma Zhixing and Momenta; earlier this year, overseas media also reported that Volkswagen will cooperate with Huawei to jointly build an intelligent driving company.
According to a staff member of Volkswagen's software division, Cariad, a member of the German magazine Manager Magazine, said, "We need systems that are different from those in the United States and Europe, otherwise, the risk of the group getting into a trade dispute between China and the United States is too great – China may not let vehicles enter its market." ”
Not only that, the localization of research and development in the era of automatic driving is far from being comparable to the era of traditional fuel vehicles. Overseas car companies not only carry out in-depth cooperation with domestic autonomous driving solution providers, but also increase the proportion of localization research and development in China.
Volkswagen Group's CARIAD software division will set up a branch in China to support the functions and capabilities of advanced intelligent interconnection, advanced intelligent driver assistance systems in the future; the GM China Forward-looking Design Center and Forward-looking Technology Research Center in Shanghai are rapidly expanding, and the number of software engineers and designers will double to accelerate the development and application of General Driving Technology in China.
The localization of intelligent driving systems and solutions by these car companies can be regarded as a precaution for the development of China.
Chinese car companies rely on high-precision maps, and the effect of high-level automatic driving in urban scenes at this stage may be better than Tesla FSD, but it is not absolute.
If the car travels to a place that is not covered by a high-precision map, including Xiaopeng NGP and other systems will actively exit, if it drives to some roads under construction, the map has not been updated in time, and the system cannot be realized.
On the contrary, this time can better reflect the universality of Tesla FSD. The data shows that tesla has driven 4.8 billion kilometers after activating Autopilot, which is equivalent to running 120,000 laps around the earth. Tesla's vision solution can adapt to the changing environment and complete intelligent driving through independent learning.
"In the era of intelligence, all car companies are learning Tesla, and the era is what he splits, which is beyond doubt." Only in local research and development, involving data security and homeland security, the effect of Tesla FSD is not as good as that of local companies, such as traffic signs, ghost probe scenes of complex road conditions, etc. An autonomous driving practitioner confessed to the car market story.
In the short term, autonomous driving relies on local conditions, and Chinese car companies can rely on radar and high-precision maps to "overtake in curves" for Tesla FSDs, but in the long run, it is not a worry-free time. After all, there is more and more news that heralds the imminent arrival of Tesla's new generation of FSDs.
It is reported that the current generation of Tesla FSD (HW3.0), the image sensor is ON Semiconductor's AR0136AT, with only 1.23 million pixels. In the new generation FSD (HW4.0), the image sensor became Sony's IMX490, and the pixels increased to 5.43 million. It is expected to be the first to be installed on the Cybertruck, which is constantly delaying mass production.
This means that for every doubling of pixels, the computing power should be increased by 3 times, and the computing power of the FSD chip may break through 1000Tops. You know, at present, in addition to Tesla, no car company can develop its own AI chip, which is also 1000Tops, and the effect is completely different. Musk has said that Tesla's computer vision software running on Nvidia hardware processes about 200 frames per second, but Tesla's specially designed chip can achieve 2,000 frames per second.
Tesla's "wolf" is growing rapidly. At that time, will the Chinese car companies with vision + radar + high-precision map scheme be the best in chess, or Tesla, which adheres to the "first principles", will win?
For more data and analysis, please refer to the full text of the 2022 China Intelligent Vehicle Development Trend Insight Report
End
Add friends please note the company and position