AI is starting to reshape every stage of ADAS design, development, and testing, from automating requirements management to generating training data with text prompts. The shift is moving advanced driver-assistance systems from separate, rule-based components toward integrated end-to-end models that process sensor data and output control signals directly.
David Doria, Director of Engineering for Automated Driving at Magna Electronics, outlined the changes in a technical briefing. Each stage of the ADAS pipeline is affected differently, but all are being driven by the same underlying advances in machine learning.
Design: Automating the back office
Even tasks that happen behind the scenes are being overhauled. Doria pointed to emerging tools that automate requirements management, such as identifying overlaps between requirement documents from multiple customers.
That work has historically been tedious and error-prone for humans. Automating it frees engineering teams to focus on higher-level design decisions.
ADAS enters a third era
For real-world ADAS features running in vehicles, Doria describes three distinct phases. Before 2012, systems were hand-coded for both perception and drive policy. Perception relied on computer vision engineers crafting image descriptors that could be extracted and matched. Drive policy depended on fragile rules and heuristics to control steering and acceleration.
The second era arrived with Convolutional Neural Networks (CNNs), which largely replaced manual feature extraction in perception. Drive policy remained mostly rule-based.
Now, transformer models and foundation models are pushing the industry into a third era. "Drive Policy is beginning to be replaced similarly, with early signs of integrating the subsystems into a single 'End to End Model,' where AI is trained to take in the sensor data and emit control signals directly-much like how human drivers operate," Doria said.
Simulation cuts data collection costs
Training and evaluating these models requires large volumes of data. Collecting that data on real roads with real vehicles is slow and expensive.
AI is changing that through simulation. Tools can now generate massive datasets without leaving the office, reducing costs and shortening project timelines. These simulations produce video game-like data and, with the latest models, can modify real-world data using natural language commands.
For example, a command like "add rain to the scene" can produce new datasets with water on the roads and droplets on camera lenses. Doria said the capabilities are not yet fully mature, but they are advancing quickly. Features where a command like "add more traffic" generates the desired scenario are expected soon.
For professionals in IT and development, the shift toward AI-driven simulation and end-to-end models changes what skills matter. Understanding how transformer models apply beyond text-into perception and control systems-will be increasingly valuable. An AI Learning Path for Software Engineers covers the underlying model architectures and engineering practices now being applied to ADAS work.
Why this matters for IT and development
The engineering challenges in ADAS mirror broader software trends: managing complex requirements, testing across edge cases, and reducing the cost of data pipelines. The AI techniques being applied here-automated requirements analysis, simulation-based testing, and end-to-end model training-are the same ones showing up in other software domains.
Developers who can work with these tools will be positioned to contribute as ADAS teams adopt them, and as similar patterns spread to other safety-critical software. For a structured path into these skills, the AI for IT & Development resources offer practical starting points. The pace of change is fast, and the challenge, as Doria put it, is "to keep pace with rapid developments and harness this technology to make real-world driving safer and more comfortable, despite its enormous complexity."
Your membership also unlocks:
