Automotive Engineering ADAS and Autonomous Driving: How AI and Digital Twins Support Engineering

Related Vendors

dataTec AG

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

Wuxi InfiMotion Technology Co., Ltd.

Artificial intelligence and digital twins greatly facilitate automotive engineering: They can flexibly complement and sometimes even replace testing in real environments with digital scenarios. This article shows how AI and digital twins interact constructively in automotive engineering and what IT infrastructure is needed for this.

In automotive engineering, particularly for intelligent ADAS functions or autonomous driving, the focus is often on finding an optimal interplay between mechanical, electronic, and software-controlled vehicle components to ensure safe participation in road traffic on the one hand and create a modern driving experience on the other. In the era of software-defined vehicles (SDVs), the optimal use of limited resources such as energy, storage space, and computing capacity, as well as their distribution between the vehicle (edge) and the cloud, is also a key research aspect.

To efficiently achieve optimal interplay across the various levels of vehicle technology, digital twins and AI can be effectively utilized. Together, they create a digital representation of vehicle technology and driving situations, allowing the functioning of intelligent ADAS features or entire autonomous driving systems to be simulated within a given context. This significantly reduces the scope of corresponding tests with physical vehicle technology in real traffic situations. Instead, in automotive engineering, virtual test scenarios with task-specific test catalogs can be set up to reveal the effects of different settings for various parameters and components on a given scenario.

Cloud-Based Test Platform with AI and Computing Power

For the AI-based simulation of complex scenarios based on digital twins, a powerful IT platform is required (Image 1), such as the one developed, deployed, and continuously optimized by BTC AG over the past decade.

Such an engineering platform includes, on the one hand, a library of AI components, which, for example, contain algorithms for detecting and classifying lanes, objects in road traffic, or traffic signs. Additionally, the complex calculations require computing capacities capable of processing vast amounts of data in the terabyte to petabyte range. Here, cloud capacities can be sensibly used on a project basis, as they can be flexibly scaled depending on requirements. An application architecture ultimately ensures that the necessary computing resources are deployed flexibly, optimally supporting engineering processes. Based on such a technological platform, intelligent ADAS functions or complete autonomous driving systems can be systematically tested, and the engineering processes of automakers and suppliers can be supported using AI.

Scenario-Based Testing

The following demonstrates how testing for a specific scenario can be performed using BTC's AI-based test platform with an enhanced digital twin. Specifically, this involves a perceptive task, i.e., a scenario where it is tested, for example, whether an autonomous vehicle behaves correctly in an urban area in front of a crosswalk, recognizes pedestrians, and comes to a controlled stop at an appropriate distance. The perception and recognition of this environmental situation are shown in Image 2 and, somewhat schematically, simulated in six subsequent steps with the help of a digital twin and AI.

• In the first step, the problem description and question for the engineering are identified and recorded: Which intelligent assistance system is it about? What needs to be tested?
• The cloud-based engineering platform with the aforementioned components and capabilities serves as the technological basis for the test. Here, the desired parameters for the base scenario to be analyzed are recorded.
• In the example case, the automated vehicle has IoT-based sensors available for environmental perception – these could include, for instance, three cameras and a lidar sensor mounted at specific positions and angles on the vehicle, providing corresponding data streams as input. Additionally, the vehicle's speed data, external information on weather and traffic conditions, as well as map data, serve as further inputs for perception.
• To create an enhanced digital twin for this scenario, the necessary algorithms and components are selected from the AI library of the test platform and integrated into the configuration. If additional components with further specialized capabilities are required, the existing AI library can be supplemented or expanded accordingly. Typical component functions include adapters for camera and lidar input, AI components for object detection, and descriptions of object classes such as crosswalks, pedestrians, or other vehicles.
• Once the configuration of components and the required data flow between them is complete, the enhanced digital twin for this scenario is generated in the cloud with a single click using an automatic setup, the BTC AI Cloud & Application Runtime. The enhanced digital twin created in this way is then ready to receive input data, which can include real video data from test drives, synthetic AI-generated video streams, or other CAN bus data.
• The input data used is now analyzed using the selected AI components to compute a description of the current situation as a result – this is then outputted with the desired KPIs such as own position, detected road users, and speed. This situation description indicates what the system has recognized well, what it has not – where corrections are needed, or where improvements may be possible.

The outputted situation description could then be used again as input in a subsequent digital twin that, for example, simulates the vehicle's planning and control activities. Overall, in this way, a variety of digital twins can be combined and assembled into a more complex system for autonomous driving functions.

Efficient Engineering Process

The enhanced digital twin, once created for a desired test scenario, can also modify various parameters as part of a planned test series. This allows for the analysis of how these changes specifically impact the system and how further optimization could be achieved. Test series can thus be conducted systematically and efficiently. Existing datasets can be optimally utilized and further enhanced with synthetic data. Expensive physical experiments can often be significantly reduced. Analysis results obtained through the digital twin can either be transferred to physical vehicle configurations or combined into more complex simulations in additional scenarios. A powerful AI-based test platform that makes modern cloud IT infrastructure flexibly usable supports automotive engineering and makes testing, further development, and optimization of intelligent ADAS functions or autonomous vehicles extremely efficient. (se)

*Dr. Ulrich Keil is Head of AI & Cyber Security at BTC.