Ali'S Automatic Driving Road Test Platform L5 Technology Landing Is Too Early.

After Google and Baidu, Alibaba also joined the camp of simulation road test platform.

In April 22nd, the hospital launched the world's first automated driving hybrid simulation test platform. This platform adopts the simulation technology combining virtual reality with reality, introduces real road survey scenes and cloud trainers, simulates an extreme scene for only 30 seconds, and the system daily virtual test mileage can exceed 8 million kilometers, which can greatly improve the training efficiency of AI model.

For automatic driving, simulation is the core of training algorithm.

In an interview with the economic news reporters in twenty-first Century, Ao ran, a senior technical expert at the Damo hospital, revealed that the hybrid simulation platform changed the way of automatic driving. On this platform, the cost of scenario construction is almost zero, so the scene variables can be arbitrarily increased according to requirements. With the introduction of human driving behavior intervention, we can accelerate the maturity of accelerated autopilot technology.

Simulation road test can greatly improve the efficiency of training algorithm. According to the RAND think tank, an automatic driving system requires more than 17 billion kilometers of data to be accumulated. It requires a fleet of 100 vehicles to test the motorcade. It runs 500 kilometers throughout the day at a speed of 40 km / h. But if we build the simulation road test based on the real road test data, the training efficiency can be increased by the order of magnitude.

In addition, extreme conditions can not be reduced in real road survey, such as bad weather, traffic accidents, etc. However, algorithms need to be trained in the simulation environment to cope with these situations. Therefore, judging from the current test situation, L5 technology still has a long time to go to landing, and vehicle development both on platform and autopilot is at a very early stage.

Training efficiency improvement

Road test has always been the core part of autopilot landing. Research shows that autopilot needs to accumulate 17 billion 700 million kilometers of test data to ensure automatic driving perception, decision-making, and control the security of the entire link. The traditional virtual simulation test platform can quickly run the mileage of the automatic driving road, but still faces the key problem of the low efficiency of extreme scene training. The extreme scene data is insufficient, and it can not restore the uncertainty of the real road condition, so the system can not deal with the sudden situation of the real road accurately, and it is difficult to achieve further breakthroughs in automatic driving.

Therefore, a number of technology companies have set their sights on this market. At the Shanghai motor show in April 2019, HUAWEI automatic driving cloud service Octopus was first exhibited. Simulation testing is one of its service capabilities. HUAWEI believes that the rapid development and listing of autopilot and functional iteration will be the key for car companies to win the market in the future. But in the process, the challenges faced by autopilot developers are also very obvious.

It is understood that if we want to solve the problem of virtual simulation test, the processing of massive data is the first pass. Through cloud services processing massive data, automatic mining and tagging can save more than 70% of manpower cost for testing enterprises.

In addition, the problem of insufficient testing data in extreme scenarios is being solved by the Damo hospital. The platform has opened the gap between the online virtual fixed environment and the real road uncertainty under the line. Traditional simulation platform is difficult to simulate human random intervention by algorithm. However, on the platform of the damoyuan hospital, not only can we use the real road test data to automatically generate simulation scenes, but also we can simulate the scene of acceleration, sharp turn and emergency stop before and after real time by artificial random intervention, and increase the difficulty of obstacle avoidance training for self driving vehicles.

In view of the shortage of extreme scenario data, the platform can arbitrarily increase extreme road scene variables. In the actual road test, it may take 1 months to reproduce the takeover of an extreme scenario. However, the platform can complete the construction and test of special scenes such as rain and snow weather and nighttime lighting conditions within 30 seconds, and the number of scenes constructed per day can reach millions.

In the simulation test, automatic driving vehicles encounter traffic accidents, which can provide an opportunity for algorithm improvement for autopilot. So this platform, to some extent, can increase the frequency and construction cost of accident scenes through new technologies, thereby improving the efficiency of automatic driving training. This is for less accidents in the future. " Ao ran further said.

Test mileage increased by nearly 6 times.

Industry experts pointed out that this platform has solved the problem of reproducing extreme scenes in a large scale, making the training efficiency of these key scenes increase millions of times, and will accelerate the acceleration of automatic driving to the L5 stage.

The hot market of autopilot also makes the simulation platform become a new battleground for the giants. According to the China autopilot Technology Research Report (2019), it is estimated that in the next 5 years, the global market of simulation will reach about ten billion US dollars.

For automated driving enterprises, the establishment of simulation platform is the key competitiveness. However, the starting point of each platform is different. For Alibaba, Tencent and other companies, the essence is based on its cloud computing business to find a wider landing scene. At the same time, the demand for offline road test is also increasing significantly.

? ? In March 2nd, the Beijing auto driving vehicle industry innovation center of the third party service organization of automatic driving vehicle road test released in Beijing reported that as of December 31, 2019, Baidu, Wei Lai, Beiqi new energy and Daimler 13 enterprises, including 6 Internet companies, 6 host plants and 1 map manufacturers, totaled 77 vehicles, and participated in the Beijing automatic driving vehicle. The total road mileage of the annual road test is 886 thousand and 600 km, an increase of 577% over the previous year. Among them, Baidu Apollo launched a total of 52 autopilot cars for road test, accounting for 71% of the total automatic driving test vehicle in Beijing, and the mileage of 754 thousand kilometers.

In the future, the automatic driving simulation test will complement the actual road test to promote the further development of the automatic driving industry. However, the automatic driving simulation technology will always serve the laws and regulations. Through simulation, we evaluate the legal liability of traffic accidents, help to manage and supervise traffic behavior, and carry out technical evaluation of traffic rules.

In the view of the industry, automatic driving simulation technology will serve product certification, provide a scientific and comprehensive method of product testing and examination through simulation methods, and also need to open up a nationwide database. At present, the domestic autopilot simulation industry is still in its infancy, and it is too early to discuss the landing of L5 technology on this basis.

?