Lead: The prosperity of Silicon Valley’s autonomous driving is not only due to California’s open policy on autonomous driving, but also because of the rich technical talents and excellent capital environment of Silicon Valley.
Leiphone: As of April 17, 2017, California Department of Motor Vehicles has issued 30 automatic driving test licenses. The world’s most fast-growing and ambitious Internet giants, car companies and their suppliers in the field of automated driving, and almost all start-up companies have chosen to conduct autonomous driving tests in California. This is not only due to California's open policy on autonomous driving, but also supported by the rich technical talents in Silicon Valley and its excellent capital environment.
Of the 30 companies on the DMV form today, in addition to car companies, suppliers and internet giants, there were only fewer than 10 start-up companies. PlusAI is the latest startup company to get a test license. In the past year, they quickly began the research and development of autonomous driving, built a technical framework that they believe can be gradually expanded, and won cooperation orders from two car companies. In March, Leiphone.com chatted with PlusAI CEO David Liu (Liu Wanqian). They summarized this past year’s exploration and told us
The following is David Liu's self-reporting.
From 2011 to 2015, my company is doing games. At that time, the main business was in China. My family was in Silicon Valley. I spent most of my time travelling between Beijing and Silicon Valley. The game company was good in the early days, but in the next two years, the industry competition was fierce, and the cost of development and promotion was also high.
Since 2015, my friends and I have started to look more into the AI field. Although I have not directly engaged in technical work myself, after graduating from Stanford with a Ph.D degree in electrical engineering and working for many years, I recognize the importance of having a keen insights into the trend of science and technology. In the beginning, we wanted to start with some investment project. At that time, we thought that some small startup companies would have very promising prospects. Among them, there are many projects that are leaning toward BI or business intelligence, such as IoT and Marketing. Only a few companies were doing intelligent transportation, and there were not many companies working on end-to-end solutions. At the beginning of 2016, we felt the opportunity was very promising. We decided not to wait and began the project ourselves.
My co-founder is also a fellow student at Stanford, who used to be the founder and Chief Architect of Yahoo Beijing Global R&D Center. Our team now has about 20 people who are familiar with machine learning and have experience working on large-scale software system. We are able to write, read scientific research paper, and write code. It is a required skill to read dozens of papers when stepping into a new field. From today's point of view, what’s lack the most in the field of AI and autonomous driving are experts. The big players in elite Internet companies and top laboratories have often become leaders in this field.
Our team members have more than 15 years of work experience on average, most of which came from big-name Internet companies. They are capable of doing things that involve AI and DL. One of the main tasks in 2016 was research and experimentation. We visited and consulted a lot of people who previously did autonomous driving, including many companies and scholars in the US. and China.
Autonomous driving has its special characteristics. It needs very deep technology. In addition to technology, it has high requirements for capital and resources. Even today, only a few startups can successfully conducted autonomous driving. California DMV autonomous driving test issued a total of 30 copies. Besides Google, Uber, Baidu, Tesla, and the traditional automotive industry tycoons, there were only fewer than ten startups.
What we did was Level 4 automatic unmanned driving, which involved precision mapping, deep learning, path planning and control, and etc.. Companies that are doing autonomous driving are using different paths to implement these modules and apply them to different business scenarios.
Our products ultimately have to face the Chinese market because the company’s main founders are Chinese, and they have worked in China for some years. Thus, it is natural that half of the early investors are also Chinese. Autonomous driving is divided into commercial and passenger vehicles. We will focus more on commercial scenarios. We hope to improve efficiency in the area of truck logistics.
Nowadays, mainly two types of people are doing autonomous driving, which represents two kinds of mainstream thought modes. One kind come from a background of doing robotics, and the other has a background doing machine learning and computer vision. The representatives of the robotics direction are Google, as they come from DARPA at the very beginning. Zoox and Otto also belongs to this school. They have a deep accumulative understanding in system engineering. Representatives of machine learning such as Drive.ai and Tesla, and also us PlusAI, are leaning towards the innovation of deep learning. Traditional robotics engineers don’t usually do machine learning, and those who do machine learning don’t do robotics. In the blueprint of autonomous driving, there are intersections between the two technology routes.
Our team integrates deep learning and systems engineering. Deep learning has made breakthrough progress in some areas recently. However, due to the high security requirements of autonomous driving, the construction of systems engineering cannot be ignored. After setting the direction, in 2016 we took a standard DARPR Stack, which allowed the prototype to run in a limited number of blocks and also explored better ways to solve some problems.
Take traffic light recognition as an example. Traffic light recognition has been in the CV field for decades, and the traditional approach is to do feature extraction. If you can determine the location of the traffic lights, then you can use algorithms to determine the status of the traffic lights and thus to determine the status of the intersection. However, accurate mapping out each traffic light is very expensive. Who can do this thing? Google can do it because it has a very complete data set, it can capture the height of the traffic lights through its feature data collection. However, if the location of traffic lights deviates, the problem will be very complicated. Deep Learning approach is based on a completely different way of thinking, judging the status of intersections from the perspective of visual cognition. It also includes finding the position of traffic lights and judging the status of each light. It is not a judgment of a single condition.
The technological development of autonomous driving is still relatively new, and the technical route is not completely established. Although there are no major differences among the mainstream teams’ high-level framework, the specific technical solutions for each level will be very different. The application scenario that everyone considers is not the same, and the problems to be solved will be very different, such as commercial scenarios or ride scenarios, long distance or short-distance, high-speed road conditions or urban road conditions, and so one. Then, the solutions used to solve these different problems will be naturally different.
This year in March, we got automatic driving test license in California at the same time as Uber. Obtaining a license plate is also a solid indicator in autonomous driving industry. Another indicator is to look at the scalability of infrastructure design and technical solutions at all levels. We also cooperated with universities such as Stanford on some research project, mainly comparing with previous research projects, such as how to improve the accuracy, performance and data usage efficiency of deep learning.
Deep learning is the key technology in autonomous driving that enables cars to self-improve. There are a group relatively concentrated experts in deep learning in Silicon Valley. At the same time, there are relatively clear rules and a more transparent and fair environment. Domestic auto OEMs cannot be compared with the pace of developing autonomous driving at Silicon Valley as technologies in China can be in a status that’s 2-3 years early. This year, we have finalized strategic cooperation with two OEMS. Ultimately, driverless technology depends on reliability and safety.