Keywords of autonomous driving 3.0: big model, mass production and gradual route

In 2022, self -driving continues to speed up shift.

Tesla's 2022 Q2 financial report shows that the FSD Beta version of 100,000 test users have a total of about 42 million miles. Musk expects if the number of testers will increase to 1 million at the end of this year, and the cumulative mileage of FSD Beta will soon exceed 100 million miles. FSD is called Full Self-Drive, which is "fully autonomous driving". Musk has made it clear that FSD is Tesla's future and Tesla's most important asset. However, Tesla FSD is considered by many people as Musk "drawing cakes" because the L5 -level autonomous driving cannot be landed in the short term.

In the Chinese market, the "progressive landing" of autonomous driving shows stronger vitality. Data from the Ministry of Industry and Information Technology showed that in the first half of this year, the market penetration rate of the new L2 -level assisted driving passenger car reached 30%, an increase of 12.7%year -on -year. In many people's impression, only "new forces" such as Wei Xiaoli have intelligent driving capabilities. In fact, traditional car companies do not hesitate. For example, Great Wall Motors sold 88226 new cars in August, of which the proportion of intelligent models has increased to 85.59% Essence

In addition to the acceleration of high -level intelligent driving mass production, China's autonomous driving has made breakthroughs in the dimensions of laws and regulations, testing roads, and business scenarios. So far According to the news of the 2022 World Intelligent Connected Auto Conference, the nationwide open test roads at all levels exceed 7,000 kilometers, and the actual road test mileage exceeds 15 million kilometers. Multi -scenario demonstration applications such as distribution are developed.

Autonomous driving technology has a long history, but it did not enter the public's field of vision until 2010. At that time, the track of technological giants such as Google and Baidu entered the game, setting off a wave of technology to this day.而从技术演进路线来看，在全球范围内自动驾驶技术都呈现出一个新的趋势：不再是硬件或者软件驱动，不再是测试道路下的数据驱动，而是以真实道路行驶场景为核心Data driver, this is a sign of self -driving 3.0.

Autonomous driving 3.0 era comes

Today, looking at the technical strength of a self -driving enterprise, we must first look at the operating mileage. Just as the chip depends on the wafer, the storage depends on the capacity, the reason is that the autonomous driving is essentially AI technology, and the core logic of AI operation is: high will be high: it will be high Quality data is constantly feeding to AI. AI is continuously evolved through self -learning. The more data obtained by a company's operating mileage, the higher the degree of intelligence, but also shows that it has the strength of "long -term real scene test".

At present, there are two companies that have the opportunity to impact the 100 million kilometers of operating mileage clubs worldwide: one is Tesla, based on the participation of millions of car owners, its FSD cumulative mileage will rush to 100 million miles; the other is to be; Scholarship. At the 6th HAOMO AI Day, Zhang Kai, chairman of Zhixing, revealed that the mileage of the Smart Bank of Smart Bank of Smart Bank has exceeded 17 million kilometers, ranking first in the Chinese autonomous driving company. The carrier rate will exceed 70%.

Gu Weiyi, CEO of Zhixing, proposed on AI Day that the industry is entering a new era with data -driven as the core: the 3.0 era of autonomous driving. So, what exactly is autonomous driving 3.0?

Hardware driver 1.0 era: In the first generation of autonomous driving 1.0 era of the initial self -driving vehicle of technology giants such as Google, hardware is the core driver. The unmanned vehicles modified by traditional models are full of hardware such as laser radar. The host and other equipment responsible for AI computing. At this stage, the hardware is the capacity of the ability to get autonomous driving, especially the radar sensor. The larger the number and the stronger the performance, the higher the corresponding degree of autonomous driving intelligence.

The problem of self -driving the hardware driver is significant. One is that the cost of the vehicle is high, and the cost of transformation is millions; the other is the low degree of intelligence, because the hardware iteration requires a long cycle. The unmanned vehicle at this stage is only only the unmanned vehicle. It can be explored in a small range and failed to move towards large -scale commercial or mass production, and the mileage is within 1 million kilometers.

Software -driven 2.0 era: After AlphaGo defeated Li Shishi in 2016, AI technologies based on deep learning are popular, and self -driving driving through many years of hardware driver enters the software driver era. Large computing central computing chips get on the car. Outside the laser radar sensor, new perception modes such as machine vision are realized based on multi -channel cameras. As edge computing development, clouds are also involved in various forms. These have greatly reduced bicycles. The cost has improved the intelligent effect, and the mileage of autonomous driving has gradually increased to tens of millions of kilometers.

Data -driven 3.0 era: It can be regarded as the continuation of the 2.0 era. AI is still core technology. The core difference is that the requirements of AI have undergone qualitative changes in data, and "big models" have become new technical cornerstones.

In 2021, Microsoft, Nvidia, Google, domestic waves, Huawei and Ali ... more and more technology giants are layout AI models. In the past ten years of deep learning, AI has entered the stage of industrialization. Its support is more widely universal. It is necessary to support greater and more complicated AI computing needs, and to upgrade from weak artificial intelligence to strong artificial intelligence. It is difficult to satisfy. Ai models with the characteristics of "huge amounts of data, huge computing power, and huge algorithms" are born, and their essence is the "enhanced version" of deep learning. By "feeding" big data to the model Self -learning ability, which has a stronger level of intelligence. Microsoft CEO Nadella said: "Deep learning has made great progress in the past 20 or 10 years, and the big model will be the next big event." AI models support "AI pre -training" and support "greedy training" through stacking data sets, which is very suitable for applications on autonomous driving, including perception, cognition, and decision -making. The core technology with a large model AI, the autonomous driving 3.0 driven by the real road data will be different:

1. Size: Around the real road scene, the data scale is larger and more diverse. The mileage will enter the 100 million kilometers.

2. Perception: Based on the large model AI, the radar, vision and other sensors work jointly, and the multi -mode state outputs the result.

3. Cognition: Increase driving comfort in various scenarios, combined with human driving common sense decision -making.

4. Model: No longer the deep learning model of manual supervision, strong planning, and intensive intervention. Instead, AI based on big data and large models, massive big data self -training, based on data channels and computing centers to achieve more efficient accumulation data. Translate data into knowledge.

In short, the era of autonomous driving 3.0 is still based on AI technology, but the connotation has changed: one is that the big model has replaced traditional deep learning to become a new training mode, and the other is the mileage of autonomous driving to a new level. The data is becoming more and more "big", which has created conditions for self -training autonomous driving driven by large models, which greatly accelerates the processes of autonomous driving. When the above, the amount of quantity will also cause qualitative changes.

3.0 will be a minority game

Compared with the 1.0 era and the 2.0 era, the core logic of autonomous driving in the 3.0 era has changed, and players in the center of the stage will be greatly reduced. The reason is as follows:

First, autonomous driving 3.0 is a data driver, and it is "massive big data driver in real scenarios".

As mentioned earlier, the basis of autonomous driving 3.0 is a large AI model, which requires enough "big" data, including mileage scale and diversity. An important detail is: Smart Travel, Tesla published "operating mileage/driving mileage", and some companies announced "test mileage". This "mileage" is not the "mileage". The mileage data is higher than the bottom, Tesla's operating mileage/driving mileage, but the value is far from each other.

In the past few years, autonomous driving companies generally announced test mileage data because their vehicles can only run on some open test roads. The more the data is, the smaller the value, because the autonomous driving technology has a "long tail effect", just as the Ministry of Transportation Ministry of Transportation, just as the Ministry of Transportation Department Zhou Wei, director of the Central Academy of Highway Science Research, said: "Autonomous driving also has a long tail effect. For example, the test of autonomous driving smart vehicles can be found in 50 days. . From the test mileage, 99.9%of the problem can be found when testing at 150,000 kilometers, and the problem of 0.1%may not be found and solved in 1.5 billion kilometers. "

To break through the long -tail effect of autonomous driving, it is necessary to continuously expand road driving scenes. The most ideal situation is open roads, and autonomous driving can arrive in places where cars can get. The "Graduation Line" turns this point into reality. The form of autonomous driving technology on this route is a high -level auxiliary driving and can run on an open road. Of course, this has another front condition: enough car participation Come in.

The premise of Tesla FSD can move to a mileage of 100 million miles is to participate in the test of millions of car owners. It has worked hard for decades to achieve a million+production vehicles offline. It's far from each other.

The confidence of Zhixing is "China's first place in mass production auxiliary driving", which adopts a unique iron triangle development model: "Scenic user experience design, AI artificial intelligence technology, and technical engineering capabilities "Cooperation", it developed three -generation smart driving system HPILOT for two years, relying on the Great Wall to achieve more than 10 different platform vehicles mass production and new models, such as We brand Mocha and tank 500. HPILOT 3.0 will officially land in 2022, and will become China's first high -level auxiliary driving product to truly mass -produce cities.

In short, at the end, Zhixing relying on the capacity of mass production and landing, mastering the entrance of the user scenario, with a large amount of high -quality and diversified big data in a large number of real road scenarios, and then grasped the entry vouchers of the 3.0 era of autonomous driving. It can be asserted that the ultimate winner of autonomous driving technology must be players with real road scenes. This is why some head autonomous driving technology players need huge sums of money to build cars.

Second, the real scene of autonomous driving 3.0 running must take "a gradual route from auxiliary driving to autonomous driving."

There are two types of autonomous driving routes:

One is the top -level design facing the form of autonomous driving, a leap forward line of "one step in place". It is light to make high -precision maps for autonomous driving to "see clear". Rebuilding and assembly -related vehicle coordination equipment on the road side, the advantage of this route is that it can directly enter the high -level autonomous driving stage, but the construction cost is high, the construction cycle is long, and the maintenance cost is not small. At present The test road is like a high -speed one to the ground.

The other is the auxiliary driving technology facing the existing road network, and it shares a gradual route with a traditional car with traditional cars. This route does not require infrastructure such as roads or even high -precision maps. The cost is lower, the threshold is lower, and the difficulty is smaller. It is more conducive to large -scale landing. Essence

"Let a part of the car be completely intelligent" or "let the car first intelligent"? The choice of Tesla and Zhixing is the second point, which will be more conducive to participating in the 3.0 competition of autonomous driving, because the "data -driven" of 3.0 requires "massive big data in real scenarios". To obtain the corresponding data There must be a user entrance under mass production capacity, but also the real driving ability of the corresponding road, and the progressive route is the best path for data accumulation. On the contrary, players who adopt the "leap forward line" are limited by open test road mileage and the intelligent transformation process of the road, etc., will be widely opened in the mileage.

Zhang Kai had previously made it clear that the worldview at the end was that from the beginning, it was determined that it was necessary to follow the progressive development route. On AI DAY, its further clarification of auxiliary driving is the only way to autonomous driving, because "the mass production time of the progressive route is earlier, it can quickly form a scale, and accumulate enough data from the actual use scenarios of users." Compared to compared to In terms of the directional collection data of Yuejin routes, the cost of collected data is lower and the quality is higher. In the practice of Zhixing, the cost and quality of autonomous driving product capabilities and scale data have formed a positive circulation effect.

Perhaps the idealists who may "make a part of the car are completely intelligent" can launch a more perfect autonomous driving solution, but the chance is small and hopeless in the short term. At least the players who "let the car are smarter" have now achieved mass production and commercial use, and they have evolved quietly in the continuous accumulation of data. The maturity of technology and the return of business are wrong, which is more in line with the laws of business itself.

Third, autonomous driving 3.0 is a strong artificial intelligence and requires a new algorithm and computing power infrastructure.

Autonomous driving 3.0 relies on AI models, while large models have three characteristics of "huge amount of data, huge amounts of algorithms, huge amounts of computing power". Each feature means that the AI ​​model is a technical competition with high thresholds. For any company, any company Including giants, it is not easy to build a big model. It needs to collect massive data, need to buy massive computing power, and need a large amount of research and development. Almost all companies with large AI models are strong financial and powerful giants -Microsoft even claims that it has used a supercomputers worth $ 1 billion to train its large AI model.

At the huge amount of data, players with commercial mass production capabilities, user scenarios, and real road scenes have significant advantages. This is obvious.

At the level of huge algorithms, Zhixing is already preparing. It released China's first data intelligent system MANA as early as the AI ​​Day last December, which can be deeply excavated on massive data. In the application of large AI models to autonomous driving, Smart Xing has also explored for a long time. According to Gu Weizheng, the reassessment of the TRANSFORMER model was launched as early as June 2021. The transformation and upgrading of the training platform, the preparation of data specifications and labeling methods, and exploring the model details of the model of perception and cognition of specific tasks, which laid a solid foundation in the urban navigation auxiliary driving scenarios.

On the AI ​​Day, based on the technical practice of "emphasis and light map", Zhixing launched the first urban auxiliary driving solution that was perceived. Smart driving problems such as road space and "diverse urban environment".

At the level of huge computing power, technology giants undoubtedly have innate advantages, but Tesla and the end are also accelerating the layout. Tesla released a supercomputers DOJO last year. The Smart Strip's solution was a supercomputing center. It found that with the application of the ATTENTION model, the demand for self -driving for computing power has far exceeded Moore's law, which has led to large model training training. The cost is very high, and it is particularly difficult to land on the terminal equipment. In response to this low -carbon supercarbing, it has been announced in December last year to build its own supercomputing center. In addition, it has also improved the AI ​​model by improving the car end model, chip design and data organization. Cost landing.

At the same time, there are "massive big data drivers in real scenarios", "step -by -auxiliary driving to autonomous driving", "enough powerful algorithms and computing infrastructure" players are rare At the moment, Tesla and Zhixing are more relevant players. However, even if you enter the mileage stage of 100 million kilometers, autonomous driving is still long -term. The realization of autonomous driving 3.0 faces many technical challenges, as Gu Weiyi summarized on AI Day:

How to apply large models in the field of autonomous driving;

How to make the data play more value;

How to use heavy perception technology to solve the problem of real space understanding;

How to use the interactive interface of the human world;

How to make the simulation more real;

How to make the autonomous driving system move more like people.

Overcoming these problems will also be the heavy responsibility of the mainstream players such as autonomous driving 3.0.

Written at the end:

Autonomous driving has become the consensus of technology, automobiles, transportation and other industries in the future, but how do you achieve it? What posture will autonomous driving go to the public? The industry has not had a unified answer. Over the years, players from all walks of life have touched the stones to cross the river, explore different routes, and jointly promote the evolution of autonomous driving technology.

So, what is the end of autonomous driving? Now it seems that whether Tesla's first "data closed loop" or the "data -driven 3.0 era" that Zhixing took the lead in shouting, it was gradually making the dispute gradually becoming consensus, and the future picture of autonomous driving was clearer: Driven by the large amount of big data of the real road scene, the new technologies such as large models are the cornerstone, so that the algorithm is continuously evolved by self -training, and then the car thinks like a driver, so that the fully autonomous driving is no longer far away. Based on this, whether to take a gradual route or a leap forward route, whether to adopt the technical solution of "strong perception, light map" or "perception fusion + high -precision map", the industry will soon form a greater consensus.

- END -