A hundred years later, people at that time will surely admire, even envy, the times we live in.
For the automotive industry, this is an era of great change once in a century; for the Chinese nation, this is an era from backwardness to great rejuvenation; for all mankind, this is a century after the electrical revolution, the era of intelligent revolution, and the third industrial revolution.
Cao Xudong wrote in a New Year letter to all members that, facing the century-old opportunities, he chanted the slogan of Autopilot Brain and led Momenta into the fast lane.
Following the roadmap of incremental and mass-produced auto-driving technology, Momenta positioned itself as a supplier in the entire automotive industry chain. Tier1 or Tier2, the main service OEM, and OEM jointly combined the brain and body of auto-driving to achieve auto-driving.
At present, Momenta announces that it has financed a total of $200 million, valued at more than $1 billion, ranking first in the domestic autopilot industry.
Cao Xudong, the companys helmsman, is known as technology cancer. From Tsinghua University to Microsoft Asia Research Institute, from Shangtang Technology to self-employment, what is the story of this absolute technology bull?
Road to growth
Cao Xudong graduated from Tsinghua University in 2008, majoring in engineering mechanics. After graduation, Cao Xudong got the opportunity of direct blog of Tsinghua University because of his excellent performance. But out of his strong interest in artificial intelligence, he chose to give up. In an interview with Together, Cao Xudong said, I have a strong interest in statistics, but I am deeply involved in computer-related artificial intelligence. Courses, even their own classes, reading, literature are enjoyed.
Figure: How-old, Microsoft Artificial Intelligence Program
Cao Xudong joined the Computer Vision Group of Microsoft Asia Research Institute after three unsuccessful retainments. He liked the research atmosphere here very much and served as a researcher. His academic interests included machine learning and computer vision. Face recognition algorithms developed by Cao Xudong were used in Microsoft Xbox, How-old and other products. HowOld.net has hundreds of millions of users, and he is also an amateur researcher. The Babel Fish Project is a chat robot project dedicated to English learning.
During his study and research, Cao Xudong also published more than ten papers at the top conference of computer vision, such as CVPR/ICCV/ECCV, among which three CVPR papers and two ICCV papers were honored by oral report (the acceptance rate was less than 5%).
He has participated in the National Data Science Bowl competition, competing with more than 1,000 teams from around the world, and won the second best result in the world.
Picture: Cao Xudong during his schooling
In 2015, Cao Xudong felt that AI technology needed to be put into practice in order to play its real value. He also jumped from Microsoft Asia Research Institute to Shangtang Technology, an AI startup company that was still unknown at that time. He was the executive director of research and development and was responsible for technology research and development and the preparation of the Beijing team.
After more than a year of Shangtang Science and Technology, Cao Xudong chose to join the wave of entrepreneurship, combined with his own advantages and experience, chose the autopilot track. In September 2016, he announced the establishment of Momenta, an artificial intelligence company in the field of autopilot.
He used to know that many Tsinghua students regard life as a marathon, running in the same crowded track, towards the desired goals of the public, inertia.
Momenta received $5 million in Angel round financing from Blue Lake Capital Leadership, Innovation Works and True Fund at the beginning of its founding; A round financing from Shunbei Capital Leadership at the beginning of 2017; A round of $146 million in July 2017, led by Ulai Capital, Daimler Group, Shunbei Capital, Innovation Works and Jiuhe Venture Capital and Investment; and B 2 round in October 2017. Financing, led by Kaihuis Sino-French Innovation Fund, GGV Jiyuan Capital.
In October 2018, they announced a new round of financing of more than $200 million. Investors include Tencent and other institutions. State-owned background investors include China Merchants Bureau Venture Capital, Guoxin Capital, Suzhou Yuanhe Capital and Jianyin International.
So far, Momenta is valued at more than $1 billion, making it one of the Unicorn AI start-ups.
Cao Xudong divides AI start-up companies into three categories: the first category, technology, no application scenarios; the second category, there are certain scenarios but business model is not innovative; the third category is the perfect state in his eyes, technology brings demand scenarios, thus triggering business model innovation and industrial pattern changes.
To this end, they formulated a three-step plan: the first stage is the construction of the underlying infrastructure platform; the second stage is the establishment of a series of software algorithms based on the underlying platform, such as environmental perception, high-precision map and positioning, driving decision-making planning; the third stage is the formation of autonomous parking, highway and urban ring roads and other different scenarios and levels of automation. Driving solutions.
It is understood that Momenta provides core technologies based on in-depth learning, such as environmental awareness, high-precision semantic maps, driving decision-making, and other products, including different levels of automatic driving programs, as well as derived large data services.
In October 2018, Xia Yan, Director of Momenta R&D, said that after the completion of the infrastructure platform construction, the team has developed algorithms related to environmental awareness, high-precision maps and driving decisions, and has further developed a number of different levels of auto-driving products and solutions, including rear-mounted safety-assisted driving products, L3-level auto-driving for expressways and urban loops. Driving solution, L4 level autonomous parking solution and L4 level urban road auto-driving solution.
On the ground, they signed a strategic cooperation agreement with Suzhou in April last year. They will set up a large-scale test fleet in Suzhou to promote L4 automatic driving landing.
In October of the same year, Momenta obtained the road test license of automobile issued by Shanghai Municipal Government. We noticed that Jiading District of Shanghai was open 11.1 kilometers and the port-vicinity area was open 26.1 kilometers. The test road of Jiading District will focus on promoting the research and development of automobile driving technology for passengers, and the test road of Lingang District will focus on promoting the research and development of automobile driving technology for commercial vehicles. Hair test.
In addition, as a sub-area of Momentas deep cultivation, the team has invested a lot of energy. At present, there are three main problems in high-precision maps, which are too expensive. Traditional mapping methods based on lidar are expensive and difficult to spread on a large scale. Finally, the cost of mapping will be transferred to OEM, which has a high threshold to use; they dare not use it, which is different from traditional navigation maps. Precision maps serve machines, and any information errors may lead to fatal accidents, so it is particularly important to ensure that the information provided by high-precision maps is safe and reliable; no use, the development of high-precision maps is still at an early stage, the industry has limited experience in the real application of high-precision maps, and most users have not yet formed a very clear demand for high-precision map products. u3002
Based on these painful points, Momenta adopts a vision-based solution, which has low cost, can support large-scale and crowdsourcing deployment, and can also achieve map creation and large-scale updating on the premise of ensuring quality.
Then, Momenta realizes the complete process from discovering information changes to quality verification and testing through visual crowdsourcing, which can quickly and large-scale update high-precision maps, thus ensuring the security of high-precision maps.
In view of the limited experience of most high-precision map users in real application of maps, they have broken through the whole chain from perception to high-precision map, and then from decision-making planning to control through a number of different levels of automatic driving solutions. In this process, they can accurately define the requirements for high-precision map in different automatic driving scenarios.
Obviously, for a start-up company that has only been established for more than two years, it is far from enough to achieve extraordinary success. Cao Xudong once said, I believe in the productivity and freedom of the future world, which will come from big data and intelligence. He firmly determined the direction of self-driving, and that is why he chose to start his own business.
Road to search
Cao Xudong once said that the greatest advantage of a team is its ability to learn algorithms in depth.
Secondly, features acquired by in-depth learning have a strong ability to transfer. Thirdly, the cost of Engineering development, optimization and maintenance is low. The deep learning algorithm is mainly convolution and matrix multiplication. For this kind of optimization, all deep learning algorithms can improve the performance.
Figure: Uber autopilot fatal accident that shocked the industry
In terms of algorithm research and thinking, Cao Xudong said frankly, the most important place is the safety of autopilot. He repeatedly mentioned on many occasions that for human drivers, a fatal accident will occur on average 100 million kilometers. Automatic driving system to achieve human safety level, it will take about 10 billion kilometers of data, if safety is more than human driving safety. Sex is an order of magnitude, which requires 100 billion kilometers of data.
So, how to achieve 100 billion kilometers? According to the calculation, 1 million cars can run 100 billion kilometers a year, but if the cost of each car is calculated in terms of 1 million RMB, it will cost about 1 trillion RMB, which is a very amazing number.
He believes that crowdsourcing can solve the problem of data accumulation very well. The first point is that the hardware cost and operation cost are low. Travelers can help the system collect all kinds of data on the way to commute. The other point is that crowdsourcing can collect shadow data to test the accuracy and safety of the automatic driving brain. Shadow testing does not need direct control. Compared with the real drivers driving behavior and the predictive behavior of the system, the safety of the automatic driving brain can be judged.
Cao Xudong also said that if safety is not guaranteed, automatic driving cannot be commercialized on a large scale.
But it is undeniable that such challenges also mean greater demand for automatic driving in China. If popularized on a large scale, it will bring about changes in Chinas transportation, Cao Xudong said.
Faced with the downward pressure of the economy, the challenge of being an entrepreneur is also enormous. Before the business model can not be verified, everything is full of variables.
Data citation and reference:
Together, Momenta Cao Xudong: Can he be a breaker in automatic driving?
Leifeng. com, Cao Xudong 7000 Words Analysis: Is end-to-end learning reliable? u300b Euro 100 million, etc.