Engineering Artificial Intelligence in Development: Best Practice for Lower Costs

Related Vendor

Wuxi InfiMotion Technology Co., Ltd.

Artificial intelligence can help to speed up vehicle development. The challenge: integrating the algorithms instead of using them in isolation. A guest article by Jörg Grotendorst.

Artificial intelligence (AI) can revolutionize vehicle development. It speeds up the entire process from the initial idea to the market-ready model and saves costs. Instead of laboriously designing new vehicles by hand, for example, engineers can describe individual components or entire models in natural language - the AI uses this to create graphical renderings that can be quickly adapted in CAD programs. Thanks to a learning phase with previous and current vehicle generations, the algorithms are familiar with the brand's form and design language and are able to maintain and modernize it.

The virtual models can be used to assess the proportions of the vehicle, its face and contours, the play of light and shadow and many other details - without the usual hand-made clay models. If a physical model is required, it can be produced quickly and inexpensively from a 3D printer or milled by machine. This makes it possible to try out different design variants with little effort; the teams involved in the development simply exchange digital data.

Other AI models can accurately predict the flow resistance of the new body shapes. They make suggestions as to how the structures can be improved efficiently and help to try out aerodynamic elements that change the flow pattern. The CFD simulations, which are expensive and time-consuming due to the computing power required, only need to be used once a design has already been validated and has the desired c_w value.

Generate and Improve Software Code Using AI

The possibilities for the use of AI in software development are particularly extensive, not least because current vehicles have long been completely permeated by software. Even seemingly trivial systems such as windscreen wipers now contain program logic that adapts the wiping frequency to the amount of precipitation. All in all, there are now more than 100 million lines of code per car, and the number is constantly increasing.

Not only creating these volumes of software code, but also checking, testing and improving it is hardly possible for human developers. They are concentrating more and more on making important architecture and design decisions and specifying requirements for the code and software tests - AI can largely take over the implementation.

Create AI-Based Tests

AI is already very good at analyzing code and identifying deviations from best practices and detecting program errors, inefficient sections of code or even unknown security vulnerabilities. It can also generate and run dynamic test scenarios that go far beyond static, rule-based test cases. This makes it possible to create tests based on current requirements in terms of stability, performance, security and user-friendliness.

The results can be used to determine the causes of problems and make suggestions for improvement. The AI can help to evaluate the system behavior in the event of missing or faulty input signals and thus improve functional safety - and to test the so-called fallback levels of the system.

The findings from the tests not only help to improve the current code, but also flow into future code generation. In the long term, this promises better software that is stable, secure and efficient - after all, the embedded systems should not waste memory space or computing power or require more powerful chips with more complex cooling.

Especially for autonomous driving at levels 4 and 5, runtimes and resource efficiency play a major role because enormous amounts of data have to be processed in real time. With eight to twelve high-resolution cameras on the vehicle for L4 and L5 applications, several gigabits of image data can be generated per second, which must be synchronized in real time.

AI Orchestrates Virtual Test Drives

Components and control systems are usually extensively tested on "hardware in the loop" test benches before they are used in real vehicles. This allows faults and problems to be identified and rectified at an early stage, which saves a lot of time and money. After all, the later a fault is discovered in the development process, the greater the effort and cost involved in rectifying it.

This has long been standard for the drive and chassis, as there is enough data thanks to billions of test kilometers with previous models. With AI, engineers can easily create new scenarios that go far beyond the routes and situations covered in the past. This is particularly valuable for simulating scenarios that put the environment sensors and controls of ADAS systems through their paces: On the one hand, there is significantly less comparable data from previous test drives in this area.

On the other hand, circumstances can be quickly generated that would take considerable effort to find in the real world, such as unusual lighting conditions, extreme weather events or chaotic traffic conditions. And situations such as a passer-by unexpectedly stepping onto the road can be simulated without endangering people.

Variate and Reproduce Situations at Will

All of this can be varied and reproduced as required to determine how the systems react to new situations and whether they work consistently. The information gained in this way is fed into further development and helps to continuously improve the AI models for environment recognition and vehicle control. Towards the end of the development cycles, significantly fewer kilometers on test tracks and public roads will be necessary.

CES 2026

The Chat GPT Moment for Autonomous Driving

When developing and testing environment detection, ADAS and autonomous driving functions, it is important not to rely on individual sensors alone. Rather, it is important to combine data from daylight and infrared cameras, lidar, radar and, for close range, ultrasonic sensors as part of a sensor fusion in order to obtain a comprehensive picture of the surroundings and traffic situation in real time. Even if a sensor fails or does not provide reliable information due to weather conditions, the remaining sensors can be used to generate an image. The individual sensors complement each other and are used to verify the objects detected by other sensors, which increases reliability and therefore also safety.

Firmly Anchoring Artificial Intelligence in Processes

Ultimately, there are virtually no limits to the imagination for the use of AI in vehicle development - the technology can also simplify the complex requirements management for components or support the development of new alloys and composite materials. The big challenge is not to look at individual projects in isolation, but to find a holistic approach for the coordinated use of AI and thus firmly anchor AI in the processes beyond lighthouse projects.

A holistic approach ensures that promising concepts are implemented and proof-of-concepts are put into productive operation. Time-consuming activities do not have to be repeated, as teams pursue their projects independently of each other.

To achieve this, however, manufacturers need to move away from a strict focus on technology, as successful AI deployment depends on more than just models and tools. They need an overarching strategy with clear guidelines for the implementation, scaling and success monitoring of projects and comprehensive change management. After all, AI requires new skills that employees first have to develop and changes roles and decision-making processes. Experience has shown that bottom-up approaches with pilot projects cannot initiate these changes - they need to be initiated top-down.

Better to Develop with Partners Than Alone

Partnerships are at least as important as the strategic and holistic anchoring of AI in the company. Experienced specialists can take over development activities completely, but can also contribute valuable expertise to the company when firmly integrated into the internal teams and ensure a rapid, sustainable transfer of knowledge.

In addition, studies show that AI projects with partners are much more likely to be successful than AI projects where companies try to do everything on their own.

Jörg Grotendorst advises the automotive sector at HTEC