Shady screens, robots, mobile vehicles...This afternoon, under the background of the curtain of flowing technology and future atmosphere, The Last Wisdom Walk officially held the "HAOMO AI DAY" online, a strong technical wind.
At the event site, Zhang Kai, chairman of Zhixing, and Gu Weihao, CEO of Zhixing, jointly counted their eye-catching report cards in the past year:
1. As of December, the user's mileage was nearly 4 million kilometers; the smart pilot auxiliary driving system NOH was officially launched in the fourth quarter of this year.
2. As of now, the intelligent driving assistance system has been installed on 5 mass production vehicles, and has reached cooperation with Meituan and Ali Damo Institute on the terminal unmanned logistics vehicle;
3. On the 22nd of this month, Zhixing completed the A round of financing of nearly 1 billion yuan and was promoted to a new unicorn company;
4. Expose the city NOH test video and the listing time in 2022;
5. Prepare for the construction of an autonomous driving supercomputing center.
The most important thing is the release of MANA (Snow Lake) of the autonomous driving data intelligence system.
According to Zhixing, the product is the first in China and hopes to challenge Tesla to achieve data intelligence leading the industry.
Gu Weihao said a few months ago: "The future of intelligent networking is that products are king and scale wins. ”
So what is the ability of mana, which has high hopes, to support the ambition of Zhixing to fight Tesla and maintain a technological leading industry?
Data Intelligence System: MANA
Road data is the root cause of technological development for autonomous driving companies, but the process of claiming this resource is destined to be long and difficult.
An evaluation of self-driving cars by the RAND think tank shows that self-driving systems need to undergo at least 11 billion miles of road verification if they meet the conditions for mass production applications. For self-driving companies with small fleet sizes and poor funding, it will take years, if not decades, to accomplish this goal.
In this case, each autonomous driving player has explored different road data acquisition methods to continuously iteratively optimize the automatic driving technology, such as through road measurement, or through virtual simulation, or the collection of passenger car driving data.
However, the road test of each self-driving car will produce terabytes of data on average every day, and how to extract high-quality and effective data for technical iteration in massive data is an industry problem that is difficult for many players to solve.
Mana, an autonomous driving data intelligence system released by Zhixing, aims at this problem by highly integrating perception, cognition, labeling, simulation, calculation and other links.
Specifically, MANA can be divided into four modules: VENUS, LUCAS, TARS and BASE, which correspond to the data visualization system, algorithm application scenario practice, core algorithm prototype practice and data general capability.
By linking these functions closely, MANA can effectively use data and ultimately improve autonomous driving technology.
In order to complete the acquisition and calculation of ZB-level data, Snow Lake uses the computing power of the mini-magic box of the Zhixing Domain Controller 3.0 with a computing power of up to 360T to obtain a large amount of accurate image data. After the data is processed, the domain controller mounted on the vehicle side will immediately exert the data effect and ensure the timeliness of data processing.
In addition, The Little Magic Box 3.0 can support more than 10 cameras and 3 lidars, providing multi-sensor data. However, in the process of data acquisition, the Team often encounters the problem that multi-sensor targets are truncated and cannot be used effectively.
To do this, they first mapped the intermediate results of the perception of the camera and lidar into Tensor Space through transformer multimodal fusion. Then add the characteristics of the timing, use RNN and optical flow SLAM for spatiotemporal fusion, and perform multi-Head calculations on this basis, and the final result can improve the accuracy and accuracy of target recognition.
Efficient identification of scenarios is only part of it, how to achieve decision-making planning in line with human driving habits in different scenarios is the goal pursued by many autonomous driving companies.
In the process of self-driving cars, the scene may include rain and snow, dark and white, whether the road boundaries are clear, and a variety of traffic participants. In the same road environment, even if one of these elements changes, it is a new challenge for the autonomous driving system, which requires the automatic driving system to maintain a high level of stability and safety.
In addition, due to differences in habits, different drivers will take different ways of dealing with the same scenario, and if the decisions of the automatic driving system cannot be different from person to person, it will bring a bad experience to users with different driving habits.
In order to ensure product safety and overall road traffic safety, Zhixing designed a dual reasoning system. One of the inference systems learns the laws and methods of human driving through a large amount of data in the scene library, and then guides the algorithm iteration.
On the other hand, in the complex scenes of the city, human driving rules cannot be exhausted, and there is no clear standard for driving style. In order to be as close as possible to the needs of different drivers, Zhixing will subdivide the data of different driving styles, and realize the "thousand faces" of the automatic driving system as much as possible through deep learning, providing a more comfortable driving experience.
Through these innovations, MANA can simultaneously achieve large-scale processing of ultra-big data, adaptability to a wider range of scenarios, and higher cognitive capabilities.
The data intelligence closed loop has been formed
The importance of road data for autonomous driving companies is self-evident, but very few companies have invested a lot of resources to build data systems like Zhixing.
So, why is the act of wisdom so much trouble?
At the third brand open day of TheLmed Wisdom Bank, Gu Weihao once said that the success of AI autonomous driving technology depends on two key links: model and data, and enough and good enough data is the premise of making a good model. Because of this, Zhixing has invested a lot of resources in the past few months to build its own data intelligence closed loop.
The so-called data intelligence, which is interpreted by Zhixing as data collection, value mining, and value application. At present, Zhixing has completed a comprehensive data layout around these three points.
In terms of data accumulation, through the "windmill strategy" approach, through large-scale product landing, It obtains a large amount of data from passenger cars and terminal unmanned delivery vehicles. Compared to other self-driving startups, the cost of obtaining road data is even lower.
Specifically, through the large number of mass-produced models sold by Great Wall Motors, Miller Chi Heng can collect different types of road data extensively, and the source is stable and sustainable, which is an important reason why the Miller Smart Mobility Assisted Driving System has traveled more than 4 million kilometers in a very short period of time.
In addition, Zhixing is actively developing terminal logistics unmanned vehicles and regards it as one of the most commercially commercialized products of the current autonomous driving technology. Since April this year, it has cooperated with Meituan and Ali Damo Academy, and has mass-produced more than 1,000 terminal logistics unmanned vehicles off the production line. The continuous improvement of the scale of terminal logistics unmanned vehicles will accumulate richer road data for The End of Zhixing.
In terms of value mining, MANA is an important part of this.
At present, MANA has integrated data perception, cognition, annotation, simulation, calculation and other links, saving a lot of costs while efficiently utilizing data.
First, MANA uses clear system failure signals and more powerful models to diagnose data from hundreds of millions of kilometers of data that has an improvement in current capabilities.
Second, MANA uses unsupervised learning to vectorize images into eigenvectors and cluster similar images through "spectral clustering". After obtaining the clustering results, MANA divides the results into positive and negative samples based on "whether they are the same as the problem scene category", and selects the data near the "class center" and "class boundary" to improve the labeling efficiency. In this way, you can effectively improve the effect of the final model.
In the third step, MANA will efficiently mine data in different spaces and times through a dual inference system.
Finally, MANA adopts a hybrid scheme in which data and model are parallel at the same time, and optimizes the forward calculation of the model, which can greatly improve the efficiency of algorithm training while reducing the overall video memory usage.
Through the above four steps, MANA can achieve efficient value mining.
In terms of value application, backed by the big tree of Great Wall Motors, the solutions and products of Zhixing have no lack of place.
It expects that in 2022, it will undertake 34 Great Wall models, accounting for about 80% of the overall models to be launched. In the next three years, the number of Great Wall passenger cars equipped with The Ultimate Wisdom Driving Assistance Product will exceed 1 million units.
MANA builds a data intelligence system around five major capabilities: perceptual intelligence, cognitive intelligence, labeling, simulation, and computing.
The starting point for turning the tide: Snow Lake
Just as the name of Mo Mo Zhi Xing is derived from the "Lao Tzu Tao Te Ching Chapter 64" "The Wood of The Embrace, Born in the Mo Mo", mana's Chinese name "Snow Lake" also has a deep meaning.
In Liu Cixin's science fiction novel "The Three-Body Problem", the alien advanced civilization Trisolaran tris to invade the earth because the mother star is on the verge of collapse, and Luo Ji, a wallfacer selected by the United Nations "Wallfacer Project", lives in seclusion in the mountains and forests to seek a way to fight the Trisolaran. After experiencing multiple reality strikes, he accidentally fell into the cold waters of the lake, but he had an epiphany about the law of the "Dark Forest" and used it to suppress the Trisolaran population for centuries.
The data intelligence system "Snow Lake" is taken from this, carrying the dream of Zhixing's desire to break through the data bottleneck and lead to automatic driving.
In a way, current self-driving players are in the same situation as Earth in The Three-Body Problem.
In the book, human beings live in fear that they may be hit by trisolarans at any time, and now various self-driving companies are worried about survival because of the unopened commercial realization.
At this time, the data intelligence system "Snow Lake" appeared, or the key weapon of Zhixing to commercialization.
At present, Zhixing already has a variety of sustainable road data sources. In addition to Weipai Mocha, Tank 300 and other models, its auxiliary driving system will also be applied to weipai machiduo, Weipai latte, Haval divine beast three mass production models.
Compared with other self-driving startups, this innate advantage will help Zhixing accelerate the acquisition of road data, and after the launch of "Snow Lake", it can transform natural resource advantages into technical advantages.
In addition, "Snow Lake" is also a springboard for Zhixing to try to achieve technological leadership.
At the scene, Hemo showed his ambition to compete with Tesla's AI technology.
At present, Tesla has collected data extensively in different road environments around the world through the shadow mode of millions of production models sold, and has improved autonomous driving technology through various means.
In contrast, although the number of mass-produced models currently on board is small, the number of its cooperative brands is increasing, and the future road data may double due to the increase in the number of models. In addition, the data from the terminal unmanned logistics vehicle will also be used by The End of Zhixing.
Zhang Kai, chairman of Zhixing, once pointed out that 2022 will be a watershed year for the commercialization of AI autonomous driving, for startups in the autonomous driving industry, forming a relatively stable business model is the basis for 2022, and sustainable and low-cost high-quality data will promote technological development and provide opportunities for commercialization.
In response to the data that may grow exponentially in the future, Gu Weihao said that Zhixing is preparing to build a supercomputing center to support massive video or image data processing and improve autonomous driving technology.
If The Supercomputing Center is successfully launched and cooperated with the "Snow Lake", the data intelligence system of the Wisdom Of the Miller will be more complete, and its data advantages will become more and more significant.
Three battles in 2022
In the "HAOMO AI DAY" activity, Gu Weihao made a prediction for the development of autonomous driving in 2022. He believes that in 2022, Zhixing must win the three battles of data intelligence technology, assisted driving city scenarios and terminal logistics unmanned vehicle scale.
In order to win the battle, he expects that the passenger car assistance driving project will be expanded by 7 times in 2022, and the terminal logistics unmanned vehicle project will be expanded by 3 times.
Today, less than ten days after Zhang Kai called the "watershed year of the commercialization of AI autonomous driving", The release of MANA by Millima Zhixing as the campaign approaches, an important part of the data intelligence system, seems to reflect that the research and development rhythm of Momo Zhixing is completely in accordance with its market expectations.
What's more noteworthy is that in just a few months, Zhixing has launched products including noH, a driver assistance system, and ICU 3.0, a mass-produced autonomous driving computing platform, and the speed of product development of startups has rarely been so rapid.
At present, the data intelligence system that is regarded as the key to victory by Zhixing has been completely established. Through a variety of mass-produced models of Great Wall Motors, Zhixing will obtain a steady stream of high-quality data.
Recently, Zhixing completed a round of nearly 1 billion yuan of financing, which will help accelerate the construction of its data intelligence system and automatic driving supercomputing center and further maintain the existing technological advantages.
As its high-quality data advantage snowballs rapidly, It may win three battles in 2022 and become the final player in the autonomous driving arena.
Leifeng networkLeifeng networkLeifeng network