Model-Based Testing for Self-Driving Car Navigation

1 Model-Based Testing for Self-Driving Car Navigation Levi Starrett Cortland Starrett One Fact Inc. 10412 US HWY 52 S. Clarks Hill, Indiana 47930 USA. Abstract Self-Driving cars pose challenges to Testing . The Recording sensor input is an efficient means of gathering a quantity and magnitude of sensor inputs create an enormous test portion of the data required for automated Testing of the self - vector space. Traditional manual Testing methods are inadequate driving car. to cover this multi-dimensional vector space with test cases. II. TEST GENERATION FOR THE Navigation APPLICATION.

2 To address this challenge Model-Based techniques are applied When an application is modeled at an abstract level, its to manage the Testing complexity. Abstraction and automation increase the productivity of test engineering. Model-Based test behavior is captured in a model that can be executed and engineering allows generation of test cases, test data and tested. This model predicts correct behavior based on correct predicted behavior. Sensor test generation, automated input. The same model can be used to detect and grade Navigation Testing and turbulence induced Testing are elements of incorrect input.

3 Since the model does not physically crash like a strategy to improve the efficiency of Testing for Self-Driving a vehicle in the real world, it can be used to explore erroneous cars . and noisy data. It can be employed to develop recovery from bad sensor input. A case study is provided as an example. Conclusions are then summarized. The same model can be manipulated to generate its own input. Given a route, a Navigation application can Keywords Self-Driving cars , Model-Based Testing , Executable automatically produce the sensor input that would cause the UML, GPS Self-Driving car to follow the route.

4 This capability can be used to automatically produce a massive quantity of test vector data I. SENSOR TEST GENERATION needed to exercise a production system. Self-Driving cars employ many sensors as input to the III. AUTOMATIC GENERATION OF TURBULENCE. Navigation . Global positioning (GPS), enhanced GPS, LiDAR, accelerometers and cameras are just a few of the sensors used The above elements can now be combined to produce test in Navigation . A Model-Based engineering approach abstracts cases and test data that contain measured amounts of the sensors with models.

5 By modeling the sensors, automated turbulence (noise). With simple, abstract models of sensors and test vector generation is enabled. A component represents the application under test, we can automate the generation of functionality exposed through explicit interfaces. Thus, the tests including prescribed levels of noise. The ability to internal behavior of the component can be abstracted and introduce turbulence into one or more sensors allows modeled using a variety of techniques including C coding, exploration of test scenarios involving faulty or noisy data Ruby scripting, Executable UML modeling or other methods coming from a combination of sensors.

6 The test cases and test of encoding behavior. data are used to test the Navigation application in a simulation environment first and then in production. The interface to a sensor component provides a straight- forward way of providing input data to the application The model of the Navigation application and its sensors Navigation component. The modeled interface between the provide for the automatic generation of correct behavioral Navigation application and the sensor components enables sensor input. This represents one body of test cases. Recorded injection of file- based test input data.

7 A standardized method input from actual sensors provides another body of test of reading pre-recorded or pre-calculated sensor data has been stimulation. With the additional capability to inject turbulence modeled and is part of the test bench modeling configuration. into both bodies of input, a large percentage of the test vector space can be explored in an automated or semi-automated Carefully pre-calculated and simulated sensor data is fashion. This combination can provide substantial coverage in important for exploring the requirements of sensors.

8 Data a controlled manner that is applicable in both a simulation recorded during actual drives is useful to stimulate the environment and in the test laboratory (and on the road). Navigation application in the laboratory environment. IV. CASE STUDY. A. Case Study Requirements A model has been built that demonstrates Model-Based Testing in the context of Self-Driving cars . The model shall be fed input GPS data. The model shall have a Navigation module that interprets GPS data and outputs verbal instructions. The design of the model shall consider a mechanism for injecting turbulence into the input GPS data.

9 B. Test Benching Concept 'Test Benching' is the concept of substituting a test bench component in place of a compatible functional component in a multi-component system. A component requires an interface. Another component with the same interface is considered compatible. Therefore, a test bench component can be substituted for any component in a system as long is it has compatible interfaces. Figure 1 shows an example application component with its corresponding test bench component and interface. Figure 2: Test benching system diagram C. Case Study Model 1) Self-Driving Car Navigation System Figure 1: Test benching component diagram In Figure 2, the application and test bench are wired together to run the test.

10 Figure 3: Self-Driving car system model Figure 3 shows the system model which is made up of a test bench component, a Navigation component, and three sensor components (LiDAR, accelerometer, GPS). This model focuses on the GPS sensor component. 2) Test Bench The test bench initiates the simulation. For each sensor, the test bench component passes a data file of simulated data and starts a global tick that moderates the frequency at which data points are consumed by the sensors. In the current case study the GPS component is the only component which is initialized uses Valhalla behind the scenes.