1.1 The commercialization of autonomous driving faces the challenge of lack of road test data
Before self-driving cars are truly commercialized, they need to go through a lot of road tests to meet commercial requirements. However, as an emerging thing, autonomous vehicles still face a large number of problems that need to be overcome, such as the time cost of road testing, the tolerance of laws and regulations in various countries for autonomous driving, the test safety of extreme scenarios and dangerous working conditions, and the road traffic environment and habits of various countries. Various issues have brought many difficulties to the development and testing of autonomous driving systems.
The time and cost of using road tests to optimize the autonomous driving algorithm is too high. Autonomous driving belongs to the category of artificial intelligence and is still in the stage of continuous development. According to research conducted by the American RAND Corporation, if an autonomous driving algorithm wants to reach the level of a human driver, at least 17.7 billion kilometers of driving data must be accumulated to perfect the algorithm. If a fleet of 100 self-driving test vehicles is configured and the road test is performed 24 hours a day, the average speed is 25 miles (40 kilometers) per hour. It will take more than 500 years to complete the target mileage. The cost is unbearable.
The lack of corresponding traffic laws and insurance claims mechanism for autonomous driving restricts the large-scale development of autonomous vehicle road tests. Since autonomous vehicles cannot yet guarantee absolute safety, the Chinese government maintains a cautious attitude towards open road tests for autonomous driving. It is difficult to meet the prosperous testing needs of autonomous vehicles by relying only on partially open roads and intelligent networked test areas for road testing. Mainly reflected in the following four aspects:
◆The road test of autonomous vehicles lacks legal basis. At present, there is still no legal basis for testing autonomous vehicles on most public roads, especially highways, which hinders the progress of the test.
◆Auto-driving road test vehicles are forbidden to carry people and goods, resulting in incomplete testing. The current regulations clearly prohibit passengers or goods that are not related to the test during the test process, which prevents the test subject from carrying out more abundant automated driving technical tests.
◆There is a lack of legal basis for the division of responsibility for autonomous vehicle accidents. Since the driving body of an autonomous vehicle is an autonomous driving system or an autonomous driving service provider, it is very different from the current traffic law system with human drivers as the main body. The liability system for motor vehicle accidents in the Tort Liability Law, the Road Traffic Safety Law and other regulations will no longer be suitable, leading to an unreliable situation in the current legal disputes related to autonomous vehicles.
◆Autonomous vehicles lack a corresponding insurance claim mechanism. Self-driving cars have broken through the regulations on motor vehicle insurance, making the current self-driving cars "no risk to invest" and increasing the risks of test companies and other traffic participants.
It is difficult to reproduce extreme traffic conditions and dangerous scenes, and there are hidden dangers to test safety. When autonomous vehicles are driving on actual roads, extreme traffic conditions and dangerous scenes can be unacceptable, and safety issues are also a major concern. According to the statistics of the National Highway Traffic Safety Administration (NHTSA), only one accident occurs when a car travels an average of 436,000 miles (700,000 kilometers), and about one person is killed when an average of 100 million miles (160 million kilometers) is driven. In addition, the self-driving car testing industry still has not reached an agreement on the safety and other standards of the test, which restricts the development and testing of autonomous driving.
It is difficult to form a globally recognized autonomous driving industry chain system. As the social and economic environments of countries in the world are very different, the road environment and traffic habits in various regions are also very different. Express delivery, food delivery, and pedestrian mixed traffic are common on China's urban roads, which puts forward higher requirements for the perception and decision-making capabilities of autonomous vehicles. Moreover, irregularities in the setting of road traffic signs and markings in China are common, and there are differences between different regions. There are also differences in the color and text description of traffic signs between domestic and foreign countries, which are difficult to change in the short term. The above-mentioned problems have caused the global development and technical exchanges of the autonomous driving industry chain to face many practical problems.
1.2 The simulation test based on the scene library has become the key to the research and development of autonomous driving
At present, the simulation test based on the scene library is an important route to solve the lack of road test data for autonomous driving. The simulation test mainly realizes the closed-loop simulation test of algorithms such as autonomous driving perception, decision planning, and control by constructing a virtual scene library to meet the requirements of automatic driving test. The scene library is the basis of the automated driving simulation test. The higher the coverage of the scene library to the real world, the more realistic the simulation test result will be. In addition, different stages of autonomous vehicle development have different requirements for the scene library, which requires the scene library to implement different test functions.
1.2.1 The composition of the scene
According to the classification of China Automotive Technology Research Center, autonomous driving test scenes can be divided into four categories: natural driving scenes, dangerous working conditions scenes, standard and regulatory scenes, and parameter reorganization scenes. The four types of scenes together form a scene library.
The natural driving scene comes from the real natural driving state of the car, and it is the most basic data source in the construction of the autonomous driving test scene. Since the natural driving scene contains all-round information such as the person-car-environment-task where the self-driving car is located, such as vehicle data, driver behavior, road environment and other multi-dimensional information, it can well reflect the randomness and complexity of the test , Typical regional characteristics, which belong to the fully tested scenarios of autonomous vehicles, and the purpose is to meet the development and verification of the most basic functions of autonomous vehicles.
Dangerous working conditions mainly include a large number of severe weather environments, complex road traffic, and typical traffic accidents. Dangerous working condition scenarios are a key part of automatic driving control strategy verification during the test of autonomous vehicles. Verifying the collision avoidance ability of autonomous vehicles in dangerous working conditions is the core of the entire autonomous driving safety test, which is the test to verify the effectiveness of automatic driving. The necessary test scenarios for sex, the purpose is to verify the safety and reliability of autonomous vehicles.
The standard regulatory scenario is a basic test scenario for verifying the effectiveness of autonomous driving. Currently, there are multiple standards and evaluation procedures such as ISO, NHTSA, ENCAP, CNCAP, etc., which provide test requirements for existing autonomous driving functions. Standard regulatory scenarios are to build test scenarios based on existing standards and evaluation procedures, with the purpose of testing the basic capabilities that autonomous vehicles should possess.
The parameter reorganization scenario is to parameterize the existing simulation scenario and complete the random generation or automatic reorganization of the simulation scenario, which has the characteristics of infinity, scalability, batching, automation and so on. The purpose of parameter reorganization scenarios is to supplement the uncovered unknown scenarios such as natural driving scenarios, standard and regulatory scenarios, and dangerous working conditions, and effectively cover the blind areas of automatic driving function testing.
1.2.2 Features of Scene Library
The research and development of autonomous vehicles includes development verification, test evaluation, and inspection certification. Due to the different purpose of each stage, the requirements of the simulation test on the scene library are also different.
1) In the development and verification stage, the scene library is to verify the functions of the autonomous vehicle, and realize the adjustment and rapid iteration of the functions. The scene library is required to have the following characteristics:
◆The scene library should cover all functional tests as much as possible to verify the safety of various functions in various scenarios, and some unnecessary functions can be eliminated;
◆The test scenario can be implemented in the real world to verify the safety of the function, and various elements in the scenario can be flexibly adjusted according to the test requirements;
◆Scenarios can be deployed to model-in-the-loop (MIL), software-in-the-loop (SIL), hardware-in-the-loop (HIL), vehicle-in-the-loop (VIL), etc. for complete in-loop testing.
2) In the test and evaluation stage, the scene library is used to evaluate the performance of autonomous vehicles in different dimensions and different aspects. The scene library is required to have the following characteristics:
◆In order to evaluate the performance of autonomous vehicles in a targeted manner, the test scenario should be highly correlated with the evaluation index;
◆In order to make the self-driving car consistent with the test evaluation results in the real world, the characteristics and indicators of the test scene elements are required to be consistent with reality;
◆In order to accurately evaluate the performance of autonomous vehicles in unknown scenes, it is necessary to perform supplementary tests on the scenes in the case of parameter reorganization. Various elements of the scene are required to be quantified to facilitate manual editing, and the scene element indicators in different scenes must be consistent for easy implementation Storage of scene data.
3) In the detection and certification stage, the scene library is to examine the safety and reliability of various functions of the autonomous vehicle under various traffic behaviors, and prepare for the final road. The scene library is required to have the following characteristics:
◆Before automatic driving is on the road, it needs to have high safety and reliability in various scenarios, so the credibility of the test and certification results must reach a very high level;
◆In order to promote the testing standards of autonomous vehicles nationwide, the testing standards should be unified, and the testing scenarios should be repeatable and consistent.
The construction of the scene library should be based on the development stage of the autonomous vehicle for targeted or modular development, and the user cost should be reduced as much as possible on the premise of ensuring the test requirements.
1.3 The combination of simulation test and road test promotes the research and development of autonomous driving
In the development process of autonomous driving, the development process of pure model simulation—software-in-the-loop simulation—semi-physical simulation—closed field road test—open road test is the most economical and efficient development process.
At present, autonomous driving simulation has been widely accepted by the industry. For example, Carcraft, a simulation platform owned by Waymo, a leading US autonomous driving company, drives about 20 million miles on virtual roads every day, which is equivalent to driving in the real world for 10 years. As of May 2020, Waymo has simulated driving 15 billion miles, compared with 10 billion miles in June last year. In addition to Waymo, GM's Cruise, AutoX, Xiaoma Zhixing and other domestic and foreign autonomous driving solution providers are also conducting a large number of simulation tests to improve their own autonomous driving systems. Simulation tests have become the most important test for autonomous driving commercial use.
In simulation scenarios, the automatic driving algorithms in common scenarios have been relatively complete, and the difficulty lies in some extreme scenarios (corner cases). Since extreme scenes are unavoidable in reality and can be easily generated using a simulation platform, the industry consensus is to increase the proportion of simulation tests in automated driving tests. At present, about 90% of the automatic driving algorithm testing is completed through the simulation platform, 9% is completed at the test site, and 1% is completed through the actual road test. The simulation test results can be tested and certified in a closed field. In addition, dangerous scenarios are summarized on the basis of road tests and fed back to the simulation test and closed field test to finally form the evaluation results. The evaluation criteria and test scenario library are gradually improved to realize the simulation test. , Closed field test, road test closed loop test, promote the iterative upgrade of technology.
With the improvement of the level of simulation technology and the popularization of applications, the industry aims to achieve 99.9% of the test volume through the simulation platform, 0.09% of the closed field test, and finally 0.01% to the actual road to complete the research and development of autonomous vehicles. A more efficient and economical state.
