Autonomous driving removing charm: from dimension reduction to diminishing dimension reduction

At a public event in 2020, Cao Xudong, the founder of Momenta, the self -driving startup company, was asked by the host why he took a different strategy with head companies like Waymo. The "connotation" of the founder Han Xu, the two sides started a round of lip -tongue sword.

Regarding Momenta's two -line combat player of L2 and L4, Han Xu moved out of the old saying: Bad two rabbits, not a rabbit [1], emphasized that some things need to be focused, and then came out: "We look at the top of the mountain top Gold rushed up, and the broken silver on the side of the road would not pick up, so he cut the thorns, opened the road every mountain, and met the water bridge.

After hearing it, Cao Xudong was a little unhappy and said back: "Operating millions of ROBOTAXI in China's first- and second -tier cities (unmanned taxis), this matter is quite challenging." Instead of the border of Hua Hui.

Two years ago, companies that do L2 and L4 -level autonomous driving are like two parallel worlds. Between you and me, it is clear, and even a little bit of it, but now, this boundary is being broken. Many L4 L4 are L4. Many L4 L4 The company began to "reduce dimension" into the L2 field.

Among them, it is dismissive of shattered pills and believes that in 2024-2025 Robotaxi can run on the street.

In May of this year, Wenyuan Zhixing took Bosch's investment. The two parties will jointly develop L2-L3 autonomous driving. Coincidentally, Liangzhou Zhihang, who mainly did L4-level Robobus, also launched a high price as low as 10,000 yuan at the same time during the same period. Step autonomous driving solutions, trace forward, Huawei and Baidu also use the technical capabilities of L4 levels to "reduce dimension" into some car companies.

In just two years, "dimension reduction" seems to have become the keywords in the autonomous driving industry. These changes seem to confirm the words of Su Jing, the person in charge of Huawei's front autonomous driving: "At this stage, companies that do ROBOTAXI are finished."

01

From dimension reduction to diminishing dimensions to survival

The autonomous driving has been divided into two forks from the beginning. One is the step -by -step form of technology from low to high, representing Tesla, the other is a one -step formula for directly achieving the ultimate goal, representing Waymo.

Autonomous driving grades are L3 as the watershed. Below L3 is generally called ADAS

L3 above is high-level autonomous driving, Robotaxi belongs to L4-L5

In 2016, Google's driverless team independently for Waymo, which officially opened the curtain of this industry. A group of Chinese engineers with a large Internet factory and the background of academic circles also chose to go to the sea. In the following two years, Xiaoma Zhixing and Wenyuan Zhixing (then at the time (then at that time Named Corporation), Momenta, Autox, Roadstar.ai and other companies have been established one after another. These companies have chosen Robotaxi, the most difficult, but also the most commercial potential track.

In 2018, Morgan Stanley gave Waymo a valuation of $ 175 billion, an increase of $ 100 billion over the previous round, of which Robotaxi's business was worth $ 80 billion, Robotruck was worth $ 90 billion, and the software authorization business was worth $ 7 billion. Essence

McKinsey predicts that in China alone, Robotaxi's order amount will reach US $ 260 billion in 2030, which will exceed the sales amount of autonomous driving passenger cars [3].

Capital sees a possibility: autonomous driving will allow drivers to "laid off". These startups that grip and diminish anti -dimensional strike technology will grow into Uber and Didi in the future. , Ending the business model of selling cars, this story sounds sexy than that of electric vehicles.

Under such expectations, in 2018, there was a wave of investment in autonomous driving. At that time, the industry's financing was 16.2 billion. Xiao Ma Zhixing, Wenyuan Zhixing, Momenta, and Roadstar.ai all completed hundreds of millions of dollars in financing. [4] Robotaxi has become a golden nest that bred a unicorn.

However, the situation quickly turned down.

At the end of 2018, former Waymo CEO John Krafcik took the initiative to stand out and poke the bubbles when the commercialization pilot progresses slowly, indicating that there will never be a completely unmanned car. [5] will never appear. The matter was actually running on a track, thinking that 90%of the problem was solved, but when I looked back, 90%were not resolved.

A similar statement is tantamount to a pot of cold water poured the capital to cool.

In China, affected by the Sino -US trade war and the new rules of asset management, startups' financing is no longer as good as before. The amount of financing of the autonomous driving industry was reduced by one -third in 2019. After the cord of the capital was cut off, Robotaxi Company I have to try new ways to make self -hematopoietic, including entering driverless trucks, unmanned bars and other tracks, and even start trying to build cars.

In theory, small subdivision scenes are easier to realize more commercialization than ROBOTAXI, but each has problems: although the route of unmanned minibus is fixed and technically easier to land, the market size is too small; the size of driverless trucks is large enough enough However, the safety and regulations and policies are equally severe; car building is a good story, but the threshold of funds is 20 billion. An industry person has commented on this, "Everyone has a clear commercial path, and it will not be charged by various tracks."

02

Multi -constraint

If the process of human beings realizing autonomous driving is climbing, then ROBOTAXI is obviously Mount Everest, and the problem is that no one can say exactly how far is Robotaxi from the peak.

After 5 or 6 years of establishment, all ROBOTAXI companies will be questioned: only financing or mass production, are you in entrepreneurial or in a pian?

The problem faced by Robotaxi is that the top hardware, advanced algorithms and massive data are indispensable. Some companies can only meet one indicator, such as Uber and Didi. They have a large amount of driving data, but the dimensions are relatively single, the accuracy is also accurate, and the accuracy is also not enough.

Some companies can meet two indicators, such as Waymo, with the top hardware configuration and algorithm, but the collected data is relatively limited. As of the end of 2021, Waymo's cumulative mileage is 32 million kilometers.

Rand is expected that the self -driving technology needs to test 10 billion miles to prove that it exceeds humans in terms of security. If a hundred teams are formed, even if you test it for 24 hours, you have to run for almost 200 years.

Even if all three conditions can be met, it does not mean that everything is ready. To achieve the commercialization of high -level autonomous driving, it must not only rely on the data closed loop, but also pay attention to the engineering ability. It is necessary to meet the requirements of vehicle regulations, such as safety and reliability, and also meet the needs of low power consumption and low cost.

For the early Robotaxi company, it is almost impossible to achieve a perfect balance in three dimensions of technology advanced, low cost, and rapid development.

In order to obtain more dimensions of data, ensure absolute security during driving, and reduce the frequency of manual intervention, Robotaxi must use high -performance hardware, including large computing chips, laser radar, redundant execution agencies, etc. When the chain is immature, they have seriously pushed the cost of the autonomous driving system.

Two years ago, the cost of a fully armed Robotaxi usually exceeded 1 million yuan. In 2021, Baidu launched a Robotaxi system mass production kit of 480,000 yuan. The peers Auto X immediately scolded, such a car "definitely dare not sit [5]."

Didi Robotaxi team, the price of bicycles exceeds one million

But the more critical issue is that the government allows passengers not to sit.

Today, although Beijing, Shenzhen, Changsha and other places all allow Robotaxi to operate the commercialization of Robotaxi in name, they are all trial operation in the government -oriented demonstration area. They need to get a license through testing and the team size is limited.

In 2021, AUTO X seized first in the business operations of Robotaxi. He was urgently stopped by the local traffic management bureau to send the team to pick up the passenger without permission. [6]

Zhou Guang, CEO of Yuan Rong Qixing, believes that to achieve completely unmanned driving, facing classic chicken eggs and eggs -raw chicken problems -if you want to let the policy be released, the government must master sufficient data reports to judge; but if the policy is policy, if the policy Without letting go first, it is difficult to accumulate enough data to persuade the government in a limited market underground.

At the same time, ROBOTAXI cannot be commercialized, but it is difficult to accumulate cash flow, and it cannot further expand the scale of operations to achieve positive circulation. Even super rich second generations like Waymo are unsustainable. In 2021, Waymo's valuation shrinks to $ 30 billion, and John Krafcik also exhausted.

03

It is easy to reduce dimension without imagination

When the leapfrogs of autonomous driving are trapped in the muddy mud, the road gradually goes more and more smoothly.

In the past ten years, Tesla has sold more than 3 million vehicles around the world, most of which are equipped with AUTOPILOT with auxiliary driving system. Through the shadow mode, Tesla collected billions of miles of road conditions and driving data.

In 2021, based on these data, Tesla comprehensively reconstructed its software algorithm for the auxiliary driving system to make it more or more autonomous on the technical framework, and pushed a new autonomous driving function FSD Beta, towards no one towards no one Driving further.

Although Musk's driverless commitment is always jumped, it has indeed explored a path of implementation. Many car companies also realize that what they currently need is not the L4 of the driver. The L2 and L3 driving, which makes the autonomous driving company always thinking about the curve overtaking, has to review its strategic choices and re -add the gradual route.

Since the second half of last year, Yuan Rong Qixing and Qingzhou Zhihang, who were originally engaged in high -level autonomous driving, have successively launched a mass production kit for low -cost autonomous driving systems. The common feature of the two sets of solutions is to "reduce dimension" of Robotaxi's technical solution to multiplication to multiplication On the auxiliary driving system of the car, it is sold to the OEM to achieve the ability of high -end intelligent driving (industry, also known as L2+). The changes in hardware and software system driven by car companies' more intensive intelligent driving competitions have provided them with opportunities.

Before 2020, because the sensors used by L4 and L2 were different, for example, the former would use expensive lidar, and the latter could not afford it at all. The differences in L4 algorithm and L2 are very different. If the company that allows L4 is allowed to provide car companies with L2 solutions, it means that the algorithm needs to be rewritten, and the workload is very large.

But after 2020, the decline in hardware prices was close to the distance between L2 and L4.

On the one hand, the performance and stability of lidar are constantly improving, but the price has dropped from tens of thousands of dollars to thousands of dollars. On the CES in 2020, Bosch, Huawei, DJI, Velodyne and many domestic startups have launched them The price of the product is as low as $ 100, which greatly speeds up the speed of laser radar.

The actual situation is the same. After 2021, a large gang of domestic new forces have launched models with lidar, including Xiaopeng P5, which sells less than 200,000. In addition, the in -depth cooperation description of Huawei and Jihu, the price of the computing platform+L4 -level sensor can already be reduced to less than 200,000 yuan [7].

On the other hand, the emergence of high computing power SOC chips has also played an important role, especially Nvidia's Orin.

Zhou Guang, CEO of Yuan Rong Qixing, said that the algorithm model of Robotaxi is very complicated. The test car is often used to operate with industrial control machines and high -computing graphics cards, but this cannot be used for mass production cars. The Orin chip has achieved a balance between high computing power and power consumption, which can drive complex algorithm models, and highly integrates, which can be arranged in the car.

At the software algorithm level, traditional L2 auxiliary driving is usually stacked by multiple sets of algorithms provided by different suppliers, with a low capacity limit. In order to achieve high -end intelligent driving, car companies have introduced more advanced algorithm frameworks, which are the fields that L4 companies are good at.

However, the computing power required by the native ROBOTAXI algorithm may be as high as hundreds or even thousands, and the production vehicles considering the costs, space, safety redundancy and other issues, and the computing resources are still relatively limited. In this regard, some L4 companies choose to streamline or compress algorithm models, and others choose to develop their own "reasoning engine". The purpose is to improve the operating efficiency of the algorithm, reduce the needs of computing power, and run in mass production cars.

It is precisely because of the maturity of the hardware system and the improvement of the software algorithm that the technical migration between L4 and L2 may be possible, and the adaptation efficiency is higher, but even so, the L4 "dimension reduction" L2 will still encounter in reality. many problems.

MINIEYE, a domestic ADAS startup of BYD, Dongfeng and other car companies, said that the vice president Zheng Wei said that the car company and the supplier of autonomous driving systems cooperate with the most important thing about "no shortcomings". experience.

The characteristics of ROBOTAXI are powerful, but insufficient production experience, lack of cost control capabilities, and product reliability are yet to be verified. Motor companies need to build a longer cycle. The traditional Tier 1 Bosch is willing to become the cause of green leaves.

On the other hand, "the talents from the Internet and AI background are actually incompatible with the style of the OEMs. Can they sink down to do dirty and lively?" An insider of ADAS startups questioned this.

Reference

[1] Great migration of autonomous driving: The maintenance of dimensions from L4 to L2, Jiazi Guangnian

[2] Li Kaifu: Buying a car is the worst investment of 96%of his life.

[3] The next five years of the travel market: driving force, winning factor, operating mode and future competition pattern, PricewaterhouseCoopers

[4] Inventory of the year 丨 Autonomous driving in 2019: After the bubble breaks, enter the "silent period", Youou.com

[5] AUTOX Xiao Jianxiong: 480,000 unmanned taxis dare not sit, the safest car company lives to the end, AI Finance News Agency

[6] Exclusive 丨 Shenzhen Traffic Management Bureau: It has not been allowed to be unmanned on the road, daily car telecommunications

[7] ROBOTAXI is mass -produced in front of L2, opportunities and challenges coexist, 9 chapters of Zhili

Edit: Luo Songsong

Visual design: Shu Rui

Editor in charge: Li Motian

Research support: Wang Lei

- END -