R&D projects that uses artificial intelligence for autonomous driving is becoming the solution for next generation. Comma.ai, Drive.ai, and momenta that we previously reported are companies that use artificial intelligence and deep learning to make L3 and L4 solutions.
PlusAI, a company that I just recently contacted, is also a company that uses deep learning as an autonomous driving solution. It was established in 2016 and has R&D centers in Silicon Valley, Beijing, and Xi’an.. PlusAI cooperates with universities including Stanford and Xi'an Jiaotong University. PlusAI has now obtained a test license for autonomous driving in California, which is relatively rare in startups. At the same time, its autonomous driving prototype has been tested in California, and has been tested for tens of thousands of miles. PlusAI has already absorbed the investment of a first-tier dollar fund and individual business leaders, including a map company, and has established cooperation with two other vehicle companies. In addition, PlusAI is also negotiating with the local government in terms of industrial development.
PlusAI CEO Liu Wanqian told the 36kr that the team is currently focusing on technological breakthroughs in deep learning and enhanced learning, and hopes to complete L4-level unmanned technology R&D in all weather conditions in the next 1-3 years.
The specific advantages of PlusAI are as following:
At present, there are many companies that use deep learning to do autonomous driving, including comma.ai, drive.ai, momenta, etc. The commonality is to do end-to-end deep learning, taking signals such as images (camera acquisition) as input, and output as control of the vehicle to turn left, turn right, acceleration and deceleration. This is an idealized end-to-end deep learning solution with a simpler construction. Complicated operations and processing are done in the channel between the end and end. But at the same time, once the system makes wrong judgments, it will be a difficult problem to adjust the data. This is often called the “black box” of deep learning.
Regarding this, PlusAI's solution is to divide the end-to-end learning process into many segments. For example, in the perception stage, the image is disassembled after the image is put in, the driveable lanes and obstacles are classified, and then this data is transferred out. At the same time, Liu Wanqian emphasized that PlusAI will also correct the output of each step, that is, if the information obtained in the previous phase is correct, it will then perform the next step of judgment.
As mentioned above, PlusAI will combine deep learning with enhanced learning. In fact, as early as the 1960s, there existed two different schools. One idea was the expert system, thinking that people (machines can refer to design) think through a tree structure to recognize the world by rules; another idea is pattern recognition, which is learning through neurons. The former belongs to reinforcement learning, and the latter belongs to deep learning.
For Liu Wanqian, these two solutions need to be combined. For example, at the camera sensing level, a computer vision solution is adopted, and lane recognition and tracking, passable area recognition, and recognition of 2D/3D scene streams are performed through deep learning. In terms of planning, reinforcement is often used. The way of learning is more through rules, reasoning, and induction, such as imitating the old driver through reinforcement learning. At 100 kilometers per hour, you can use the light and line, overtaking, and then back to the original lane. That is, whether or not to overtake and the line is determined based on the speed of the preceding vehicle.
At the commercialization level, PlusAI has established cooperation with the two companies, but Liu Wanqian frankly stated that social acceptance of laws and regulations are not matured yet. PlusPlus is also promoting the development of L4 automatic driving. Therefore, it takes time for the commercialization of L4. PlusAI and automotive companies are still in the stage of providing technical interfaces.
However, in the cooperation with automotive companies, PlusAI has also tapped its new needs, namely to provide L2, L3 level automatic assisted driving solutions for commercial logistics automotive companies, Liu Wanqian told 36kr, this is a short term to see the path of commercialization, the fastest at the end of this year, early next year will have products landed, the two sides will develop a prototype out.
Since PlusAI uses auto-driving with deep learning, its L2 and L3 solutions are more like the simplified version of its L4 solution. The technical framework is similar, except that the sensor configuration is different. For example L4 level will use laser + millimeter wave radar + camera, and L2, L3 only need to configure a simple sensor on it, which is radar + camera. In the primary and secondary relationship of the sensor, Liu Wanqian stated that PlusAI had previously used a completely laser-based and completely vision-based solution. At present, taking into account the cost of commercialization and the actual results, PlusAI uses a vision-based approach. Laser is an auxiliary solution because vision is a passive perception solution. The light comes from the light or the sunlight. The amount of information is larger than the active light emitted by the laser itself.
In addition, the above mentioned that PlusAI will focus on solving big data collection and automatic calibration problems. In this regard, Liu Wanqian stated that PlusAI uses sensor fusion to do data labeling. Currently, it has been possible to do synchronous labeling under automatic labeling, as it can handle one hour of data within one hour. The GPUs used by PlusAI includes self-built networks and Amazon's engine room.
The recent focus of PlusAI is to provide a L2 solution for commercial logistics vehicle enterprises that can run on high speed and on the trunk line. In terms of positioning and guidance, Liu Wanqian indicated that it will adopt the method of GPS+ medium precision map, which is the so-called medium accuracy map than the standard accuracy. With one lane-level data on the map, PlusAI will obtain real-time data through the sensor module and combine it with the map to obtain accurate positioning. This map maker that had previously invested in PlusAi can help with maps, acquisition and accumulation of data.
Dr. Liu Wanqian, who is currently the CEO, graduated from Stanford University and was selected as the “Ten Thousand Talents Program” of the Group Department in 2012. He was thus hired as a national specialist. Another founder, Mr. Hao Zheng, is the founder of Yahoo Beijing Global R&D center and served as Yahoo's chief architect for mobile search. In the past decade or more, they have successively founded technology companies in multiple fields, all of which have been successfully acquired or listed. PlusAI team members are from world-class technology companies and institutions.