The rapid development of automatic driving technology makes the dream of cars entering the "no man's land" no longer remote. However, to truly realize the large-scale commercial landing of autonomous vehicle, safety testing and verification has become the top priority for the further development of the industry. Is autonomous vehicle a "novice" or an "old driver"? Recently, Feng Shuo, an assistant professor in the intelligent transportation research team of the Department of Automation at Tsinghua University, and Liu Xianghong, director of Mcity at the University of Michigan in the United States, collaborated to develop a new safety testing system, tailoring a set of "driving test questions" for "AI drivers". Recently, the research achievement was published in the regular issue of Nature under the title of "Safety Test of autonomous vehicle Based on Intensive Reinforcement Learning", and was on the cover of the current period. With the development of automatic driving technology, when the driving level is more and more close to human drivers, the safety performance test of autonomous vehicle becomes more and more important, but also more difficult to carry out. Feng Shuo told reporters: "Currently, there is a preliminary consensus in the industry on this: it is urgent to solve the problem of 'billions of kilometers'." What is the' billions of kilometers' problem? The researchers introduced that before autonomous vehicle are put into large-scale application, large-scale road tests are needed to verify the safety of autonomous vehicle from a statistical perspective. It is estimated that the scale of this test should reach at least ten billion kilometers. From the perspective of time, resources, and cost, it is clearly difficult to achieve testing on the actual road. This is one of the most challenging issues facing the current development of autonomous driving technology. What we are doing is hoping to accelerate this process and replace the billions of kilometers of field testing with as little testing mileage as possible, "said Feng Shuo. How to find out the safety problems of autonomous vehicle with the lowest cost and the highest efficiency? Feng Shuo led the team to search for research entry points from a statistical perspective. Experienced drivers who encounter unexpected situations while driving will combine road conditions and feedback from nearby vehicles, rely on intuition to make judgments and quickly respond. So, how should artificial intelligence make decisions when facing the same situation? This is essentially an expectation estimation problem for small probability events in ultra-high dimensional space, "explained Feng Shuo, "The complexity of human-computer interaction and the complexity of road traffic state determine that autonomous vehicle need to deal with various situations that occur in ultra-high dimensional space, which is a 'dimensional disaster' we face. In order to verify safety in testing, we need autonomous vehicle to learn to deal with traffic events in various dangerous situations. Because dangerous situations are often small probability events, we will also face a 'sparse disaster'." After transforming practical problems from a statistical perspective into academic problems, Feng Shuo and team members sought breakthroughs from a theoretical perspective and creatively proposed intensive reinforcement learning methods. By identifying and deleting non security critical states, connecting security critical states, and training neural networks in the edited Markov process, they solved the "sparsity disaster". At the same time, the intensive reinforcement learning method is used to train the background vehicles in the traffic environment, and an intelligent test environment composed of autonomous vehicle and background vehicles is constructed, so that the simulation environment can replace the actual road