Bao Shiqiang (51VR’s expert): Automated driving simulation not only focuses on technical details but also builds an ecosystem toolchain
Recently, at the 6th International Congress of Intelligent and Connected Vehicles Technology (CICV 2019), Bao Shiqiang, R&D Director of 51VR Intelligent Vehicle and Transportation Division, delivered a speech on “Virtual Simulation Test Methods and Cases”, introducing current development processes and application cases of the automatic driving simulation.
It is essential for the current automated driving industry to develop virtual simulations. Testing and training autonomous vehicles in a virtual scene can save money, labor costs, and is safer and more efficient.
Automatic driving simulation software is currently thriving, China’s original automatic driving simulation software is also catching up. In this speech, Bao Shiqiang shared his thoughts on the simulation test method, the technical details of the simulation test process, and what level of the original domestic simulation test software has reached.
Bao Shiqiang revealed that 51VR intends to cooperate with the whole ecosystem of autonomous driving. From major OEMs, suppliers, autopilot startups, and various demonstration zones and testing agencies, 51VR is working hard to unite all parts of the industry to create a self-driving supply-chain complex. I believe that in the near future, China’s autonomous driving industry will usher in its new development.
The following is a record of Bao Shiqiang’s speech:
dSPACE and NVDIA have just shared the simulation test very well. Different from them, in the automotive industry, we are still a rookie. There are many places to learn from precedents, but we also have some unique features of our own. I hope to share with you today.
We are 51VR, founded in 2015, our vision is to create a real, complete and permanent virtual world — 51World, specifically to our Intelligent Vihecle and Transportation division, we hope to develop through visualization, simulation and prediction technology and traffic simulation software with completely independent intellectual property rights, helping you to build smarter cars, build smarter roads, and achieve safer, more efficient and better travel in the future.
With the development of the automotive industry towards intelligence, networking, and software, virtual simulation testing is becoming more widely used and more important. The meaning of virtual simulation testing at different stages is also very different.
The earliest simulation tests basically refer to the simulation of dynamics and mechanical structures. The more famous software includes ADAMS and CarSim. With the development of ADAS, there has been simulation software like PreScan that can set up test scenarios, add sensors for testing, and companies that offer HIL test solutions like dSPACE.
In recent years, with the automatic driving to the L3-L4 level, there have been a number of automated driving simulation solutions featuring high-precision restoration of actual test roads, using intelligent traffic models, deployed in the cloud, such as Cognata, Parallel Domain, rFpro, and 51VR (China).
New forms of simulation tests do not replace traditional simulations. On the contrary, they together make a larger ecology. A large number of automatic driving simulation tests raise the requirements on traditional dynamic simulation and HIL test.
Now the simulation technology develops very fast. The main engine factories, algorithm companies, the inspections, and regulatory agencies have different understandings of the simulation and have many arguments.
For example, is it better to use road test data playback to re-simulation, or just develop an intelligent traffic model? What ratio is appropriate between the simulation test and the actual drive test? Is it proper to largely use target-obstacle simulation, or a high-quality simulation camera and lidar? These issues have different considerations depending on different needs.
51Sim-One is our self-developed automated driving simulation platform, covering the whole process of automatic driving simulation tests, including static and dynamic data imports, test scene cases editing, sensor simulations, simple dynamic simulations, tests and playbacks, virtual dataset generations, etc. All these functions are highly precise and realistic.
At the same time, 51Sim-One uses a flexible and scalable simulation structure, which can be deployed in a single machine, private clouds, public clouds, flexible interfaces, and easy access to autopilot systems or integrations with existing simulation processes.
The automatic driving algorithm virtual test basically contains the following modules.
The first is the generation of the static world. We must have the ability to collect data of real road environments, generate high-precision maps, and then reconstruct the entire scene based on the high-precision map.
The next step is to build a dynamic world. We not only access the actual road tests or traffic data to the system but also develop an intelligent traffic model. These two different ways are to provide dynamic test scenarios close to actual complex traffic conditions.
Then we have to model vehicles, including the dynamics, sensors, and we have to dock interfaces for the autopilot systems which are being tested.
Other important modules are the storages, organization management and deployment of test cases, which determine the scale of the autopilot simulation test.
Eventually, comprehensive data analysis and reports are needed to evaluate the effectiveness of tests.
Let’s talk about the construction process of static scenes in virtual test in detail.
Our static scene creation is centered on OpenDrive.
To this end, we have developed an OpenDrive editing tool that provides a variety of ways to generate the OpenDrive we need. With this tool, we can edit through actual mapping data and satellite imagery, or create scenes with road templates.
After that, based on OpenDrive, we can call pre-prepared road resource libraries with Chinese characteristics to automatically generate simulation scenes including pavements, lane lines, street lights, and traffic signs.
The meaning of OpenDrive is not just to provide accuracy and guidance during static scene reconstructions. At runtime, both the simulation systems and the autopilot systems need to perform many queries.
The simulation is not just to create pretty virtual scenes. The precision, structured logical road relationships, and semantic information are all very important.
There are two main ways to generate dynamic scenes.
One is the data-driven method. The length of this test is relatively short, but the test data is huge, and data sources are also multi-faceted. It can be the actual road test recording data from traffic camera recognition.
The tracks can also be certain conditions by editors. The data collected from the actual collection needs to be cleaned, processed and stored.
An important development is the ability to re-edit data, to properly edit and expand data of real scenes.
Another way of dynamic simulation is to parameterize the traffic model. Compared with the pure data-driven approach, this method can quickly add random traffic including motor vehicles, non-motor vehicles, and pedestrians.
In addition to the need to define a set of traffic rules, including signal lights, intersection priority, and other mechanisms. For large-scale continuous testing, parametric traffic model simulation has great potential for development, but it is very high to simulate some real-world anomalies and chaos.
How the actual collected data be used to extract the behavior of different traffic situations becomes a challenging issue.
In terms of the analog aspect of sensors, the simulation of the camera involves the realism of scenes and camera models.
With the current mature game engine technologies, it is possible to restore real-world models, lighting, and materials with higher quality.
More detailed simulation involves measuring the real-world light intensity, color balance, and camera distortion parameters.
The image sequences generated by camera simulations can be directly connected to the automatic driving sensing system by means of software, or HIL tests performed by video injections or black boxes. In addition, features of all elements in virtual scenes can be utilized to generate an annotated deep learning virtual training data set at a very low cost.
Compared with road tests, an important advantage of the simulation test is that it can increase the scale of tests quickly by increasing power.
We (51VR) developed a distributed parallel acceleration simulation structure that can be easily deployed in private clouds and public cloud environments.
We also provide front-end-based case editings and management platforms, which multiple users can submit different simulation test tasks. The backend system can break down instruction to many nodes to complete every simulation tasks according to requirements.
The same simulation hardware can simultaneously accelerate the case of running a large number of decision system tests, as well as synchronous simulation of multiple sensors.
The development of virtual simulations is inseparable from the support of major OEMs, suppliers, autopilot startups, demonstration zones, and testing organizations. It is inseparable from different components of the entire ecosystem, such as HIL testings and site testings, dynamics simulations.
With the actual road test data developments and the support of intelligent transportation big data, 51VR is willing to work closely with everyone to open up the entire industry ecology and jointly promote the development of virtual simulation testing.
BTW: Welcome to leave a message below, Bao Shiqiang (51VR’s expert) will reply to all your questions.