Safety and homologation of automated vehicles (AVs) presents a huge challenge for their market introduction. The number of regulations and standards considering safety of AVs gradually increases, but current safety standards and regulations still have to be adopted and enhanced. Moreover, systems based on artificial intelligence (AI) are becoming increasingly prevalent in the electronics systems of automated vehicles. These systems use the available data to draw their own conclusions and successively learn from every encounter of a traffic situation. The certainty that a vehicle will react correctly in traffic thus continuously increases over time. Therefore, the objective is to be able to evaluate the system's learning progress as well as to ensure that the decisions of these systems are always safe for the traffic around. Here, an enormous number of scenarios and environment parameter combinations must be considered. Confronting conventional real-world tests with this test effort is not feasible anymore, and therefore, virtualization of testing methods by means of simulation has to be emphasized. In his keynote, Dr. Houssem Abdellatif presents the current challenges regarding the safety of AI-based systems as well as how to tackle them.
Massive change is underway in the automotive industry with trends in vehicle electrification, autonomous driving, and wireless connectivity. In this talk, Andy Grace, who leads the development of products for Model-Based Design at MathWorks, shares his vision for increased usage of simulation, design automation, and artificial intelligence to accelerate these trends.
What does the mobility of tomorrow look like? The startup Emm! Solutions GmbH is tackling this difficult problem. Their approach for implementing progressive mobility solutions is the targeted use of rapid prototyping, the modularization of functionality, and the parallel use of simulation.
In this talk, Armin Müller, the founder of Emm! Solutions, discusses his vision and development methodology for the future of autonomous vehicles using a real-world use case of developing an autonomous driving function in just 12 months.
Armin Müller, Emm! Solutions
In May 2018, a group of researchers from the Technical University of Munich (TUM) won the first Roborace Human + Machine Challenge. TUM's autonomous driving software stack manages environment perception, autonomous navigation, and trajectory tracking.
The ability to virtually test the complete autonomous driving software stack, while relying on high-fidelity simulations of the vehicle and its surroundings, is of major importance whenever developing autonomous driving systems. This talk presents a hardware-in-the-loop (HIL) environment based on scalable and expandable hardware that leverages an integrated software solution.
The full autonomous driving stack is simulated in two separate hardware target computers, which mimic the real technical setup of the Robocar—an NVIDIA® Drive™ PX2 and a Speedgoat real-time target machine. The Mobile real-time target machine, designed specifically to work with Simulink Real-Time™, acts as the vehicle ECU, translating the medium-term desired trajectories into immediate commands for the vehicle actuators though real-time CAN controllers. The NVIDIA Drive PX2 is responsible for tasks such as trajectory planning and sensor processing. Communication between both units is handled by real-time UDP.
A second Speedgoat target machine is used to simulate the vehicle dynamics as a reaction to the vehicle ECU's inputs. This real-time simulator also features sensor and actuator emulators, which make the software think it is operating in the real world with realistic data streams. The physical and behavioral modeling is handled with Simulink® and Vehicle Dynamics Blockset™, with Simulink Real-Time again enabling the fast prototyping onto the real-time target.
An additional GPU server implements the environment model of the racetrack while providing a full 3D visualization. A twin representation of the real-world racetrack can be easily built using the Level Editor in the Unreal Engine® by importing track data captured from the vehicle sensors.
Thomas Herrmann, Technical University of Munich
Michael Lüthy, Speedgoat
A vehicle control unit (VCU) is intended to implement higher level powertrain functions, such as cruise control, speed limiter, or fuel-saving assistants, to release the engine control unit from complex and resource-consuming calculations. Within VDLs next-generation bus product line, a Continental VCU device is utilized for top-level management and control of the fully electrified powertrain.
Continental provides a generic VCU device as well as basic software and a MATLAB® and Simulink® based software development toolchain called MBDS to the truck and bus OEMs. Through this toolchain, VDL is able to implement the complete VCU functionality. Using the model-based development system toolchain, MBDS, the design, integration, test, and build of the VCU software is done.
MBDS supports a fully automated model test approach, which is easy to use and powerful, to increase software quality and satisfy ISO 26262 requirements. Test definition is done using Microsoft® Excel® sheets representing different test cases. Both, stimuli and expected values, including tolerances, are specified for fully automated tests. The test cases can be applied on different test levels (model-in-the-loop, software-in-the-loop, and process-in-the-loop) and the results for all levels are summarized within a HTML-based test report. The report provides all required information about the test run, including test descriptions and visualization of the test stimuli, the expected values, and tolerances. Even references to requirements are covered for analyzing traceability of all test activities. Especially for qualifying the generated C code, the test coverage information can be captured for model-in-the-loop tests.
MBDS provides a simple MATLAB command line API to open and close projects, run tests, and build and download the final target software. Through a batch tool, a remote control session of MBDS can be established for nightly test and build sessions using a Jenkins™ server or other batch job environments.
Dr. Sven Semmelrodt, Continental
Legal limits for CO2, higher level fuel cost, and more strictly legislated countries are pushing down on the OEM to find a new sustainable way to meet these needs, causing the OEM to look at hybridization further. They must discover hybrid powertrain benefits that cannot be realized by the conventional ones, such as recuperation during vehicle or pure electric traction. Hybrid vehicles include increasingly complex components that when managed correctly can lead to fuel consumption reduction, satisfy widespread customer needs, and preserve the product's distinctiveness.
The OEMs are preventing increases in cost during development, in particular during software development, by reusing features, ensuring scalability, and managing the complexity in system architecture. To do that, Christian Corvino of Lamborghini has found that Model-Based Design is the best solution for development of embedded software compared to the classic methods which are based on low-level languages and intensive prototyping activities. Model-Based Design shortens the distance between high-level control algorithms and the actual implementation, improving the product software development process in respect to quality, scalability, variant management, and reuse goals above all for the small and medium enterprise company that the market recognizes as distinctive.
In this talk, Christian discusses different applications of Model-Based Design with particular emphasis on the prototyping techniques used to speed up software development.
Dr. Christian Corvino, Lamborghini
Over the past 15 years, the number of fatalities in road traffic in Germany decreased continuously from around 6800 to just under 3200. In the same period, the number of traffic accidents recorded by the police remained roughly constant. In addition to infrastructural changes, this is primarily a result of improved passive and active vehicle safety, to which traffic accident investigation and analyses have made a crucial contribution. Therefore, the Institute for Traffic Accident Research at Dresden University of Technology (VUFO) is investigating about 1000 traffic accidents per year as a part of the German In-Depth Accident Study (GIDAS) since 1999 and records critical situations in a Naturalistic Driving Study (NDS).
In order to continue this tendency and meet the increasing requirements from the development and assurance of future driver assistance systems to highly and fully automated driving, it is important to use statistical and simulative analysis methods efficiently and to develop and combine new approaches.
This talk presents applications in traffic accident research that use MATLAB® and Simulink® as well as appropriate toolboxes to improve vehicle safety and the assurance of highly automated driving.
This includes, among other things:
- Statistical accident analyses and mathematical models with Database Toolbox™ as well as Statistics and Machine Learning Toolbox™
- Traffic accident simulation and creation of a pre-crash matrix (PCM) with MATLAB and Simulink (external solver)
- Effectiveness analyses and efficiency assessment of advanced driver assistance systems (ADAS) with MATLAB and Simulink
- Algorithm-based processing of real driving scenarios (NDS data) for simulative applications using Computer Vision System Toolbox™, Automated Driving System Toolbox™, Signal Processing Toolbox™, and Mapping Toolbox™
Florian Spitzhüttl, Institute for Traffic Accident Research
Today, the problem scientists are plagued with is no longer how to acquire data or to store it. In a world where even your kettle has internet access, getting data from A to B is no longer a challenge. Transmitting data from remote locations has been made easy by using the “XG” (3G, 4G, and 5G) networks. Storing data is easy, too. With single drive capacities now hitting 14 TB, you are unlikely to ever run out of space, even with video data. There are also plenty of tools and databases, each with their own advantages, to process and store your data into. This has resulted in three problems: 1.) What to do with all this data we are recording?, 2.) How to analyze it?, and 3.) How can we know that the data is actually valid? This talk concerns the latter problem: strategies to make sure that you can focus on former two later on.
Data often comes from sensors and third-party devices you have no control over. For the most part, they work as expected, at least it seems so. Intermittent bugs like non-unique timestamps and changing names and values with (minor) offsets are hard to catch unless you already suspect them. Documentation is no help and software updates are likely to introduce new problems. In most cases, you end up staring at raw data searching for the needle in a haystack. MATLAB® allows you to write your whole validation and analysis toolchain in a single programming language that is easy to understand and powerful enough to process large amounts of (raw) data.
This talk describes the process and steps used during project AMPERE, a fleet study where 56 vehicles were equipped with data loggers collecting data for a period of more than 12 months, and how MATLAB was used to manage and analyze all the data. It covers an approach that has since been used in other projects as well.
Jan Grüner, Technical University of Berlin
The validation of automated driving functions is currently a pressing need in the automotive industry, which can only be solved by massive, multidisciplinary simulation approaches. In this talk, Dr. Andreas Kuhn of ANDATA explains how MATLAB® is a superb environment for such complex functional integration.
Dr. Andreas Kuhn, ANDATA
System simulation is a crucial factor in mastering the increasing complexity during the vehicle development process. Different quality assurance measures must be taken into account to ensure that the system simulation is a reliable partner. Users and developers consider the main quality characteristics to be maintainability and efficiency when focusing on a simulation framework. To fulfill these quality characteristics, Groupe PSA developed a cross-domain simulation framework named AXIOM (automotive X-in-the-loop, object-oriented model framework). AXIOM boosts the capability of reuse of component models and tools across all simulation domains like model-in-the-loop, software-in-the-loop, or hardware-in-the-loop, and makes use of Git™ version management to enable collaborative development.
To achieve increased readability and consistency, a tool named Model Configurator was developed to support the integration process for building up application models. A modular approach enables the reusability of models and simplifies the maintenance.
Agile principles have been established to work collaboratively and efficiently with cross-functional development teams. The developers and the customers who use the application models can report bugs and request features via Jira. The team will discuss the feature requests and the workload will be estimated. This feature request will be planned in frequently repeated sprints. If there is a bug, a hotfix will be performed. In regular intervals, releases are created and distributed to the customer.
To improve the release quality, a continuous deployment approach is used. Correspondingly, for physical vehicles, there are quality measurements at the end of the production line. In the virtual factory, Groupe PSA uses quality measurements to check continuously if the model is working correctly. All information is collected with Logstash, analyzed with Elasticsearch, and visualized in Kibana, the so-called ELK stack.
Andreas Erbes, Groupe PSA
Deep learning provides tools and methods to address common problems of the automotive industry in new and revolutionary ways. In addition, simulation and virtual models allow flexibility and speed-up in the development and testing process of ECU functions.
In this presentation, an ECU development process is described, which uses data-driven engine temperature models trained with nonlinear autoregressive neural networks with external input (NARX) and MATLAB® in the cloud to overcome the obstacles of traditional physical modeling approaches.
Along with XCU controller code compiled by Simulink Real-Time™ as executable for virtual calibration on a desktop computer, calibration and test engineers can use the previously generated data-driven models and typical engineering products like ETAS® INCA to tune and validate that XCU controller code on a completely virtual environment, saving costs, increasing agility, and accelerating the development and validation process of ECU functions.
Michael Wutz, Continental
Adaptive AUTOSAR is a modern software framework intended for high-performance, on-board computers often used in autonomous systems. Based on POSIX and C++, it supports dynamic and updatable software as well as services-oriented communications and has extensions for safety and security.
In this talk, MathWorks introduces you to Adaptive AUTOSAR concepts and showcases how the Simulink® product family offers support for Adaptive AUTOSAR including:
- Modeling and simulation of Adaptive software components using service-oriented communications
- Support for AUTOSAR Adaptive schema 18-10
- C++ production code generation with Adaptive middleware interfaces (ara::com), and AUTOSAR XML export
Richard Thompson, MathWorks
The ISO 26262 standard for functional safety provides guidance on the development of automotive electronics and electrical system, including embedded software. A common challenge is to determine the strategy, software architecture, and design patterns upfront in a project to achieve standard compliance and avoid mid-project changes to these foundational areas.
In this presentation, MathWorks engineers address the following topics based on their experiences applying Simulink® to production programs that require ISO 26262 compliance:
- Key considerations for model architecture for ISO 26262 compliance
- Modeling constructs required to meet freedom from interference
- Applying the above best practices to meet AUTOSAR requirements at the same time
Dr. Tjorben Groß, MathWorks
As part of the Women in Tech initiative, MathWorks will be hosting a Women in Tech lunch during this year’s MathWorks Automotive Conference, which is intended for female delegates and presenters. Join the lunch to hear from leading technical experts and to discuss your experiences. Use this opportunity to meet and network with other female industry peers.
Eva Pelster, MathWorks
In this talk, you will learn about using MATLAB® and Simulink® for perception, planning, controls, deep learning, and systems development through examples that ship in the newest release, including:
- Perception: Design LIDAR, vision, radar, and sensor fusion algorithms with recorded and live data
- Planning: Visualize street maps, design path planners, and generate C/C++ code
- Controls: Design a model predictive controller for traffic jam assist, test with synthetic scenes and sensors, and generate C/C++ code
- Deep learning: Label data, train networks, and generate GPU code
- Systems development: Simulate perception and control algorithms as well as integrate and test hand code
Marco Roggero, MathWorks
Artificial intelligence has produced better-than-human accuracy and saved time through automation in many industries. This talk discusses how deep learning and machine learning can be applied to image, signal, and text data for applications such as extracting critical events, implementing automated driving, and controlling machines. In addition, you’ll see how MATLAB® uses built-in algorithms and apps to save time in key parts of the AI workflow from data handling and labeling to code generation.
Dr. Sebastian Bomberg, MathWorks
Full vehicle simulation models are needed to assess attributes such as fuel economy and performance for each candidate. At times, this requires integrating models from different engineering teams into a single system level simulation. It can be challenging to integrate these subsystems with multiple controller models or code together in a closed-loop testing environment. In this talk, you will learn how MathWorks automotive modeling tools and simulation integration platform can be used for powertrain selection studies.
Eva Pelster, MathWorks
Learn how to design a lane-following and lane-changing algorithm for an autopilot-assist feature in highway driving. See a demonstration of system-level simulation to test the decision making, path planning, and control modules in traffic scenarios.
Mark Corless, MathWorks
The effect of the ECU on drivability can be dramatic and requires time for proper calibration. Traditionally, experienced drivers test a vehicle during tip in/tip out maneuvers and provide a subjective drivability rating, iterating on ECU calibrations until the subjective feel is acceptable. Using simulation-based methods, it is possible to conduct much of this analysis up-front using objective methods. By extracting key features from the acceleration results, an objective metric can be obtained. Formal optimization methods can identify a calibration set that provides a much better drivability response for the initial in-vehicle tests, thus reducing the overall time required.
In this talk, you will learn how MathWorks tools for data analysis and vehicle modeling and calibration were applied to perform objective drivability calibration.
Dr. Jan Janse van Rensburg, MathWorks
Reinforcement learning allows you to solve control problems using deep learning without using labeled data. Instead, it uses a model of your system that captures the appropriate dynamics of the environment and learns through performing multiple simulations. This simulation data is used to train a policy which is represented by a deep neural network that would then replace your traditional controller or decision-making system.
In this talk, you will learn how to use Reinforcement Learning Toolbox™ and other MathWorks products to set up your environment models, define the policy and its various hyperparameters, and scale training through parallel computing to improve performance.
Christoph Stockhammer, MathWorks