The autopilot simulation test is over.

Recently, China's first L4-level autonomous driving open test base project with a total investment of 654.38 billion yuan started in Yongchuan District, Chongqing, and Baidu Apollo autonomous driving test operation center will be established to build the full chain test service capability of "virtual simulation+closed test+open test".

Image source: Yongchuan High-tech Zone?

In addition to the Chongqing Experimental Zone, many intelligent networked test bases in China are also actively adopting the "simulation+actual road test" mode to promote the landing of autonomous driving technology. The representative is the National Intelligent Networked Automobile (Changsha) Experimental Zone located in Hunan. The experimental area not only takes the autopilot simulation laboratory as one of the development priorities, but also cooperates with Tencent's autopilot team.

Then, in addition to the actual road test, why do industries "play" simulation?

Simulation is just needed.

When autonomous driving enters the open road test stage represented by Robotaxi, how to improve test efficiency, improve technology and speed up landing has become an urgent problem for the industry.

Rand Corporation, a famous American think tank, once estimated that a L5-class self-driving vehicle needs 1 1 100 million miles on the road. This means that even a self-driving fleet with 65,438+000 test vehicles will take about 500 years to continuously test at an average speed of 25 miles (40 kilometers) per hour of 7×24 hours. From the enterprise level, even Waymo, the head enterprise, has only completed 20 million miles of self-driving road test by the beginning of 2020, and there is still a big gap from the goal of 1 1 100 million miles.

It can be seen that it is unrealistic to promote self-driving vehicles on the road only by relying on actual road tests. At this time, the importance of automatic driving simulation is highlighted.

For engineers, the algorithm and sensor configuration needed to develop the related functions of autonomous driving can not produce enough valuable data even after the road test of 654.38+million miles, but the simulated driving mileage can reach 654.38+million miles a day. There is no need to test on the actual road, no oil, no bridge toll, and more importantly, no risk. For players in the field of autonomous driving, the importance is self-evident.

At present, major car companies all over the world are studying to replace part of the actual road test mileage with simulated test mileage, that is, 90% of the automatic driving test is completed through the simulation platform, 9% through the test field, and 1% through the actual road test. Online simulation test has become an important means to accelerate the development of autonomous driving technology and test landing.

As early as 20 17, James Stott, a senior software engineer of Waymo, created Carcraft, a virtual reality world similar to The Matrix. Many situations encountered by Waymo's self-driving vehicles in the field road test can be directly blurred in Carcraft. Programmers can superimpose various situations and create various extreme situations, and the data obtained in the simulator can be fed back to the test vehicles in the real world.

At present, Carcraft's virtual fleet has 25,000 cars, and the vehicles of these virtual fleets run in Google's data center 24 hours a day. Thanks to this data closed loop, in July 2065438+2009, Waymo announced its latest milestone: the total mileage of simulated autopilot test has reached 65.438+000 billion miles, and it supports Waymo models for large-scale testing.

Image source: Waymo official website

In addition, at the end of last year, Waymo? Also acquired Latent, a British analog technology company? Logic, the AI technology developed by the latter, can build realistic human driving behavior through "imitation learning" and help Waymo realize more realistic simulation technology.

Not only Waymo, but Uber, which has been sluggish because of fatal incidents, obviously can't let itself fall behind too much.

At the end of last year, Uber plans to acquire Foresight, a simulation software developer, and has started in-depth negotiations with the latter. In fact, this is not the first time Uber has acquired a technology company related to autonomous driving. Last June, Uber planned to acquire Seattle startup Mighty? AI, focusing on developing training data for computer vision models.

Domestic enterprises accelerate research and development.

Relevant domestic enterprises also realize that the simulation test is very important to speed up the automatic driving landing.

In fact, at present, many mainstream simulation test systems are developed according to game engines, such as Microsoft's AirSim and Intel's Carla…… .......................................................................................................................................... Tencent, which also has the right to speak in the game field, is certainly not to be outdone.

Based on its powerful game engine, Tencent developed TADSim autopilot simulation test software. As a technology company with rich game development experience and technical reserves, Tencent combines game engine with industrial vehicle dynamics model, virtual and real traffic flow and other technologies to create an online simulation environment that infinitely approaches the real world scene.

Image source: Tencent official website

Combined with the collected traffic flow data and the simulation of extreme traffic scenes, TADSim can carry out various aggressive driving and extreme automatic driving tests. At the same time, the high-precision map built in TADSim can complete the closed-loop simulation verification of all modules in the real vehicle, such as perception, decision-making, control algorithm and so on. The software can also simulate all kinds of weather and light, which greatly improves the efficiency of automatic driving test.

As a "BAT", Baidu is certainly not to be outdone.

As early as the end of 20 18, Baidu and Unity? Technologies has established a partnership to jointly develop a real-time simulation product, which will create a virtual environment for developers to test autonomous vehicles in a realistic simulation environment.

Not satisfied with the cooperation with the real-time 3D R&D platform, Baidu spent a lot of energy to develop its own autopilot simulation system.

2065438+March 2009, Baidu's paper was published in the sub-issue of Science magazine, Science? In robotics, this paper proposes a new automatic driving simulation system: Augmented Reality Automatic Driving Simulation System (AADS). AADS system includes a newly developed traffic flow simulation framework based on data driving and a new scene picture synthesis framework based on image rendering. Through AADS system, the test effect of autopilot system can be improved and the autopilot algorithm can be improved continuously. In July of the same year, the Apollo platform was upgraded to version 5.0, and the vehicle dynamics model was added. This technology reduces the error of traditional modeling results by 80%.

Image source: Baidu official website

In addition to Baidu and Tencent, which entered the game earlier, as a latecomer in the automotive field, Huawei also has its own logic in empowering autonomous driving cloud services.

Huawei's logic is to build a full-stack cloud platform that can be obtained on demand through Octopus, which can not only cross the mountain of data processing, but also cover the whole life cycle business of autonomous driving model, training, simulation and labeling. Therefore, "Octopus" can provide users with the following core capabilities: processing massive data, automatically mining and labeling; Accelerate the construction of software and hardware, and improve the efficiency of training and simulation; Rich simulation scenes; Clouds cooperate with each other, and Che Yun is seamlessly connected.

Image source: Huawei official website

Simulation scenarios provide more than 10000 simulation scenarios by integrating scenario design and data-driven methods. The daily virtual test mileage of the system can exceed 5 million kilometers, supporting the concurrent test of 3000 instances. It can be seen that the simulation ability is only one link in its data closed loop, and this link can only play its greatest combination advantage if it cooperates with other links.

Huawei said that it will integrate high-precision maps, 5G and V2X technologies in Octopus in the future. As a result, Huawei, which does not build cars, has a grand layout in the simulation market.

At present, the industry has reached an understanding that the simulation platform is the next battlefield for military strategists. Not only industry giants, but almost all self-driving startups represented by AutoX, Wen Yuan Zhixing, Ma Xiao Zhixing and Sidi Zhijia also independently develop simulation environments according to their own needs. At the same time, the industry has gradually cultivated unicorns with expertise in the autonomous driving simulation test platform, such as Senma Technology and 5 1VR.

According to experts' prediction, by 2023, the total international market size of simulation software and testing will be about10 billion dollars, and the simulation technology suitable for autonomous driving system will be widely used in the research and development process of automobile products. Although the simulation test can't really replace the actual test, at least on the road of mass production of self-driving cars, enterprises can take fewer detours and accelerate the arrival of business.

This article comes from car home, the author of the car manufacturer, and does not represent car home's position.