Technical anxiety of traditional car companies

Since 2023, the self-driving head company Waymo has laid off more than 200 people. Embark Trucks, an autonomous driving startup with a market value of more than $5 billion, now puts the dissolution business in the option, and Argo AI supported by Volkswagen and Ford also announced its closure last year.

In China, some companies that insist on L4 autonomous driving are not as high-profile as before, but more and more news about fixed-point cooperation with mass production car companies comes out. Even those new brands that hold high the banner of intelligent driving seem to be out of date in the market.

If autonomous driving triumphed a year or two ago, now the industry seems to be at a contradictory crossroads.

On the one hand, it is the "commercial eve" that has been called, and all cities have NOA video screens. If you don't talk about autonomous driving ability, it doesn't seem worthwhile to sell a car.

On the other hand, no company dares to show the optional rate of advanced assisted driving in the terminal market, while BYD, which has always been considered to be weak in intelligent assisted driving, has entered the quartet.

What are the trends behind complex industry changes?

Yu Kai, founder and CEO of Horizon, made such a sound at the 2023 committee of 100, an electric vehicle in China. After experiencing the peak of rapid technological development, the industry began to need to return to the commercial essence. And what is the essence of business? In fact, enterprises create value, transmit value and gain value, that is, some people are willing to pay for products.

Returning to the essence of business: advanced assisted driving creates value. Yu Kai believes that autonomous driving should return to the commercial essence, and the value of advanced assisted driving is being recognized by more and more users.

A set of data shows that the core demand of consumers for intelligent driving technology at this stage is to alleviate driving fatigue, and the most important factor for users to recognize high-speed driving to assist driving is to "make driving easier".

Not long ago, Horizon discussed a topic at the NOA test drive meeting of Journey 5: What is the perception of computing power and chip iteration for users?

The ideal car owner "Big D has an attitude" means: "The most intuitive perception is that the display interface is clearer and can tell me how many things the car can perceive around it. On the other hand, when dealing with unexpected problems, it is best to take the initiative to change lanes. "

The former is to give users more confidence, and the latter is to make users more willing to use the assisted driving function.

Yu Kai also talked about his own experience: "Last year, I once returned to Beijing from Shandong and stood in line for five hours. At that time, I was driving a Li ONE, and the whole journey was very easy, so I didn't have to step on the accelerator all the time. This is different from the feeling when there was no assisted driving before. "

Therefore, from a commercial point of view, the core significance of autonomous driving at this stage is to make it easier for users to drive and meet the value needs of users in order to obtain commercial returns.

This is exactly what Horizon is trying to do.

Not long ago, Auto Company experienced the high-speed NOA function of the ideal L8 equipped with Horizon Journey 5. It can be said responsibly that compared with the early driver assistance system, both the overall driving efficiency, the logic of changing lanes, and the more subjective confidence and security for drivers have been greatly improved.

According to the data released by Xpeng Automobile, during the period of Spring Festival travel rush this year, the proportion of high-speed NGP mileage to the total mileage of users reached 3.25%, which was greatly improved compared with 1.38% during the National Day four months ago. Many new power brand users also expressed their recognition of advanced assisted driving functions on social media.

Compared with Japan and Europe, assisted driving functions are more driven by safety regulations, such as AEB, ACC, lane keeping, emergency automatic braking and other functions are playing an active and safe role.

The expectation of assisted driving in China market is based on safety and driven by better user experience and user value. Intelligent electric vehicles are increasingly replacing traditional BBA luxury cars in the market, and high-speed NOA is also rapidly entering mass production, which is proof.

In Yu Kai's view, this is an opportunity for China's automobile industry to change lanes and overtake.

Without anxiety, how should the NOA autopilot industry respond to such opportunities and needs?

Yu Kai believes that last year was the first year of mass production of L2+ high-speed NOA, and some head enterprises will start to launch NOA in urban areas this year, but from the technical point of view, it will take about three years to make a good progress.

In fact, the current advanced assisted driving arms race seems to be out of proportion to the improvement of user experience. Just like for most users, the assisted driving based on camera perception seems to be not much different from the laser radar blessing system.

Therefore, Yu Kai called on the industry not to be too anxious about autonomous driving, because the industry is not developing so fast. By 2025, at a reasonable cost performance, high-speed NOA will achieve a silky smooth experience, and at the same time, it will require considerable investment to make urban NOA available.

In fact, the development of technology and the improvement of user experience have always complemented each other and promoted each other, as has the development of autonomous driving.

When NGP was first popularized in Tucki, a researcher once said that "more users can use it" was a very important consideration in developing the system.

This principle remains the same today, just like the beginning of the "flywheel effect". ?

NOA has a long way to go in the development of urban areas, but it is of great value in data collection, scene collection, feedback research and development, and the continuous expansion of user usage scenarios will also give more users the opportunity and conditions to use advanced assisted driving functions.

Doing high-speed NOA is to make the user experience close to the extreme, let users build trust and dependence on advanced assisted driving, and at the same time better understand the value and use boundary of advanced assisted driving.

Horizon: Optimize with algorithm, approaching the upper limit of user experience. In Yu Kai's view, at present, dozens to hundreds of TOPs or even 1000Tops are high-speed NOA, and the difference is not too big, so there is still a lot of work to be done in the industry.

This is also the direction of horizon's efforts.

Horizon believes that under the given computing power, it is necessary to constantly optimize software algorithms and approach the upper limit of user experience with more data. "I believe that in a few years, we will reach this level, about a few hundred, and NOA in urban areas can also do well." Yu Kai said.

Not long ago, at the world's top artificial intelligence computer vision CVPR conference, Horizon, as the first author, proposed an end-to-end autopilot algorithm framework based on Transformer. This paper stood out from more than 9,000 papers and was selected as the TOP 12 paper.

This is the prospective research and development of horizon algorithm. For the first time, the neural network is used to monitor, track and predict the trajectory of the arrow end to end, and it is completed with a simple structure from front to back. The traditional way is to package them, cut them into many modules and then separate them.

This means that Horizon may train the whole autopilot system with end-to-end large-scale data.

Based on the understanding of the algorithm, Horizon injected the forward-looking research of the algorithm into the design and development of the chip architecture, which is also the core computing architecture developed by Horizon-BPU (Brain Processing Unit).

In Journey 5, the third generation BPU architecture-Bayesian architecture is used, which can efficiently support transformer calculation. Compared with competing chips, Horizon can complete more efficient calculations with lower power consumption.

Higher computing power and efficiency enable Horizon to learn human driving attempts unsupervised with larger data and larger models, "just like you learn from a large number of unsupervised, unlabeled and overwhelming natural texts." Yu Kai said.

Specifically, by giving the current traffic environment, navigation map and a driver's driving behavior history, let the machine learn to predict his next driving action.

The experimental data also prove that when the parameter scale of the model increases, the test loss will decrease accordingly, that is, the more parameters, the smarter the system.

Relying on a strong ecology, Horizon can help OEMs realize their dreams. Horizon has always been committed to building a software and hardware ecology of intelligent driving and robot computing. In the industry, canoe Zhihang has obtained the mass production project of a well-known automobile factory, which is based on the urban NOA scheme constructed by single journey 5, while autopilot companies such as Ma Xiao Zhihang and Wenyuan Zhixing have also developed software schemes on the horizon chip.

The time window of the chip industry is very narrow. In the window period, we don't choose a chip technology pioneer company, but choose a fastest-growing ecology, just like Android and Apple survived, but windows phone lost. Therefore, the purpose of Horizon Ecological Construction is to involve more partners, make the cake bigger, and realize the smooth operation, upgrading and prosperity of the whole industrial chain with upstream and downstream partners through the soft and hard collaborative technology platform and flexible and open business cooperation mode.

The commercial progress made by Horizon is also visible to the naked eye. At present, more than 50 models have been mass-produced, with more than 120 front mounting points, and nearly 3 million self-driving chips with vehicle specifications have been shipped. Zhengtu 5 has also become the only mass-produced chip in the industry with computing power exceeding 100 tops, and has been loaded and delivered on the ideal L8, and has also reached a strategic cooperation with Audi ag.

From the automobile company's point of view, in Yu Kai's view, the development of autonomous driving in the next 10 year is like "people riding a horse", which can realize autonomous driving, but people and vehicles should also cooperate, and the ultimate responsibility lies with the driver.

In fact, although some car companies have already landed L3-level automatic driving function, they can also see their very strict conditions of use, which is a natural contradiction with "letting more people use it".

But no matter what kind of trade-offs and choices, people-oriented should be the core.

Man is the ultimate beneficiary of all technological developments, and also the payer of these technological developments. The road ahead of autonomous driving is not achieved overnight. In this process, how to balance the relationship between technology and market demand and how to make them promote each other may be a topic that the whole industry should constantly think about.

At present, focusing on user value may be a more pragmatic and feasible choice for the autonomous driving industry.

This article is proud of the author automobile company, and the copyright belongs to the author. Please contact the author for any form of reprint. The content only represents the author's point of view and has nothing to do with the car reform.