Torsten Kluge, Group Manager Engineering, dSPACE GmbH
You will have to drive your autonomous prototype vehicle hundreds of millions of miles to test it against the seemingly endless number of traffic scenarios and requirements that are possible. Trying to accomplish that level of testing on the road is not realistic, but there is a better solution.
You can validate advanced driver assistance systems (ADAS) and applications for highly autonomous driving by using virtual test drives, where algorithms and models are used to replicate any imaginable driving scenario. Real driving is replaced with virtual driving to deliver valuable test results and data to validate functions for ADAS and applications for autonomous driving.
dSPACE offers a complete solution for creating virtual test drives.
By using the dSPACE Automotive Simulation Model (ASM) tool suite, it is possible to create complex virtual scenarios with an unlimited number of detectable objects. Elements of the environment, such as roads, intersections, buildings, traffic signs, construction zones, traffic vehicles, pedestrians, and conditions such as weather, road surface, etc., are all recreated virtually to provide a comprehensive and very realistic test environment.
The process of creating a virtual test environment begins with modeling. dSPACE ASM models are open Simulink® models that are used to simulate everything from individual vehicle components and systems to road networks, traffic objects, traffic sensors, and complete virtual traffic scenarios.
The models can be used for model-in-the-loop (MIL), software-in-the-loop (SIL), and hardware-in-the-loop (HIL) platforms.
“The ASM models are the most flexible and easy-to-use models for the whole model-based development process, cluster simulation as well as real-time simulation, and they are complete, open, and customizable,” says Torsten Kluge, Group Manager, Engineering, dSPACE GmbH.
The dSPACE ASM tool suite can be used as part of a simulation workbench to provide realistic models for function design and controller testing in the model-based development processes.
For ADAS applications, dSPACE offers a collection of ASM models to simulate vehicles, traffic objects, road networks, and sensors.
MotionDesk recreates your models as realistic 3-D animations.
Road descriptions can be very complex. To meet requirements for testing, it is critical that road descriptions are based on real-world data. Within the dSPACE tool chain, you can create artificial or virtual roads (i.e., road networks, traffic objects, etc.) by using the Road Generator in ModelDesk together with ASM models.
These tools let you define lanes, lane widths, height and surface definitions, curvatures, slopes, surfaces, guard rails, lane detection lines, etc., in great detail, making it possible to test driver assistance systems such as lane departure and lane-keeping systems.
To incorporate real-world data when creating your road networks, you can import and export data from route data coordinate measurements from sources such as Google Earth, OpenStreetMap, and the global navigation satellite system (GNSS).
Once you have built your road network in ModelDesk, it is synced with MotionDesk to get a more realistic and accurate 3-D visual of your simulated roads.
After building the road models, the next step is to define your test vehicle and other vehicles (moving and parked) that will be on the road, i.e., where and how they will be maneuvering around other vehicles and objects. Maneuvers can be simple or very complex, based on conditions and trigger events that you set up (e.g., entering an intersection or changing a lane).
Furthermore, maneuvers can also include data such as acceleration, velocity, steering, pedal actuation, etc., defined over time or distance.
Along with vehicles, traffic objects and sensors need to be factored into the virtual test environment to provide the most realistic traffic scenario possible.
Using the Traffic Editor in ModelDesk, you can import detectable and dynamic objects such as pedestrians, buildings, road signs, traffic lights, crossings, trees, etc. Individual details for object characteristics can be defined, including object motion characteristics (i.e., position, velocity, lane selection, etc.), to enhance models for testing.
“The definition of traffic vehicles and traffic objects can be complex and time-consuming,” says Kluge. “Therefore, and for easy reuse, we are working together with OEMs and suppliers to define the Open Scenario format. We will support it soon in dSPACE ASM and ModelDesk.”
In addition to traffic objects, sensor models need to be incorporated into the vehicle under test.
Models for all kinds of sensors common to ADAS applications and autonomous driving (e.g., lane detection sensors, cameras, and radar sensors) can be brought into the simulation environment for testing, and the sensor models can be parameterized. For example, you can test for the detection of other predefined objects (e.g., traffic vehicles and traffic signs) within a given range.
Sensor models can run either directly on the main processor or on the graphics card. Simulation on the graphics card allows for a very detailed simulation of camera, radar and laser sensors, including physical effects. The sensor models on the main processor are called ground truth sensors, because they provide exactly the information from the simulation ground, without any no disturbance effects, etc.
dSPACE ground truth sensor models use a geometrical approach to support sensors such as radar, lidar, and cameras. Supported sensors include traffic sign sensors, 2-D sensors, 3-D sensors, and a customizable sensor model for future kinds of sensors.
To realistically simulate traffic flow within the virtual test environment, traffic situations (for simple and complex scenarios) are constructed by defining such things as vehicle movement, lane driving and lane changing, intersections, oncoming and crossing traffic, etc. These characteristics are set by entering data details such as velocity, acceleration, distance to another vehicle, distance to the center of the road, relative distance to another vehicle, etc.
Models can be connected to third-party traffic flow simulation programs, such as Simulation of Urban Mobility (SUMO – an open-source product) and PTV Vissim®, to realistically simulate traffic flow for any number of vehicles. The traffic flow information from these sources can be used with ASM for testing ADAS functions.
Last but not least, data from real accidents can be imported to recreate scenes in ASM. With this data, tests can be run to determine how the accident could have been avoided with the use of emergency controllers.
The German In-Depth Accident Study (GIDAS) database, for example, documents accidents in detail so they can be reconstructed in a simulation program. Variables such as braking deceleration, starting speed, collision speed, and angle-changes are all included in encoded parameters.
Once you have finished modeling your virtual test environment (including vehicles, vehicle maneuvers, static and moving objects, sensors, the environment, traffic scenarios, etc.), you can begin testing the functionalities and observe system behaviors in your simulated environment.
To run the thousands of tests that will be required, dSPACE offers cluster simulation. Models created in the ASM tool suite are used in combination with the dSPACE VEOS simulation platform to perform automated testing of software (virtual test drives) directly on a PC instead of on the actual road.
Using PC-based simulation with VEOS, you can execute tests faster than in real time. This makes it possible to complete millions of test drive miles each day. And any test drive that fails can be reproduced and debugged in detail. Extensive testing can be carried out for several stages of development before starting HIL tests or real test drives for functions for highly automated driving. Traffic scenarios can be easily modified and immediately simulated, so you can run more tests without having to generate code again.
With all of the options that are available in a simulated test environment, it is not necessary to drive hundreds of millions of miles on the real road to complete autonomous vehicle testing. Instead, you can run thousands of virtual test drives right at your desktop or in the safety of your lab.