On April 22, Alibaba DAMO Academy, the Chinese Internet giant, released the world’s first autonomous driving “comprehensive simulation test platform”. The system’s daily virtual test mileage can exceed 8 million kilometers, which will promote the acceleration of autonomous driving to the L5 stage. Ma Yun once said in the establishment of the Alibaba DAMO Academy in 2017, “The DAMO Academy was established to affect at least 2 billion people worldwide.” Today, this “prophecy” has begun to take shape.
The study found that to prove that the mortality rate of traffic accidents caused by autonomous vehicles is lower than that of human driving, at least 8 billion kilometers of actual road tests are required. In other words, assuming that the average speed of the test vehicle during the actual road test is 40km / h, the test fleet is composed of 100 vehicles. It runs 365 days a year, 24 hours a day, and it takes about 230 years to complete this mileage. The “mileage nightmare” is a major obstacle to restricting large-scale commercial applications of autonomous driving.
The traditional pure virtual simulation test platform can quickly complete the mileage of autonomous driving road tests, but it still faces the key problem of low training efficiency in extreme scenarios: the extreme scene data is insufficient, the uncertainty of the real road conditions cannot be restored, and the system cannot accurately deal with the real road conditions. In an emergency situation, it is difficult to achieve further breakthroughs in autonomous driving.
Alibaba DAMO Academy’s self-driving simulation test platform solves this problem. On the platform, not only can the real drive test data be used to automatically generate simulation scenes, but also increase obstacle avoidance training for autonomous vehicles difficulty through artificial random intervention, real-time simulation of front and rear vehicle acceleration, sharp turns, emergency parking and other scenarios to.
In response to the problem of insufficient data for extreme scenes, the platform can arbitrarily increase extreme drive test scene variables. In the actual road test, it may take 1 month to reproduce an extreme scene takeover, but the platform can complete the construction and testing of special scenes such as rain and snow weather and nighttime lighting conditions in less than 30 seconds. The number of supported scene constructions reaches millions. Ali pointed out that this platform solves the problem of recurring extreme scenes on a large scale, making the training efficiency of these key scenes millions of times higher, which will promote the development of unmanned driving.
