The automotive industry is experiencing an unprecedented technological storm. On one hand, centralized Electronic and Electrical Architectures (EEA) are accelerating their evolution toward central brains (High-Performance Computers, HPC) and zonal controllers, causing the complexity of in-vehicle software to grow geometrically. On the other hand, Generative AI (GenAI) and AI Agents are sweeping in with a disruptive posture. However, in the realm of automotive embedded software, which pursues ultimate safety, the "black box, uncertainty, traceability gaps, and quality inconsistency" of GenAI have become fatal flaws for engineering implementation.
Image source: MathWorks
Faced with the "uncertain" AI and the "absolutely certain" automotive functional safety standards (such as ISO 26262), the industry is shifting from initial frenzy to rational reflection. Recently, MathWorks, a leading developer of mathematical computation software for engineering system design, has provided the answer: using a trusted Model-Based Design (MBD) toolchain as the foundation, integrating the creative capabilities of GenAI, and utilizing "deterministic" tools to dissolve the "uncertainty" of AI.
Breaking Through Uncertainty: Building an AI Engineering Defense Line with a Trusted Foundation
In safety-critical fields such as automotive and aerospace, the threshold for code quality is extremely high. Relying purely on Large Language Models (LLM) to generate code poses huge safety risks—even with the same prompts, the code structure provided by the LLM each time may be completely different, and it lacks bidirectional traceability from requirements to code. In the eyes of ISO 26262 functional safety auditors, this is undoubtedly an unacceptable black box.
However, when GenAI is combined with MBD, the rules of the game change. MBD has a mature application background of about 50 years in the aerospace industry and 30 years in the automotive industry. Its core advantage lies in the fully controllable process from algorithm design, virtual simulation, and code generation to compliance certification. In the new integrated practice, the core logic proposed by MathWorks is: letting the large language model find and invoke the correct, reliable industrial-grade tools.
Regarding this, Tom Erkkinen, Product Manager for Embedded Software and Certification at MathWorks, stated: "The innovation and comprehension capabilities of large models are used to parse high-level requirements, but they do not directly generate the final production code. Instead, they drive Simulink® for model creation and modification, as well as code generation, verification, and compliance work."
Since the mathematical logic and built-in features behind Simulink models are unique and verified, and the Embedded Coder, which has passed the highest level of ISO 26262 certification, automatically generates C/C++ code, this perfectly eliminates uncertainty. From requirements and architecture to detailed design and embedded code, the entire workflow is tightly closed-loop through the toolchain, thoroughly establishing a trusted foundation for GenAI creation and production software.
Agentic AI Implementation: Driving a Three-Pronged Approach for Design, Code, and Verification
In the launch of the latest product lines related to MATLAB® and Simulink, this integration has seen substantive technical implementation. By introducing Agentic AI workflows, MathWorks has officially transcended traditional "chat-based assisted design" and stepped into action-oriented automated engineering.
Image source: MathWorks
-
Accelerating Design: MATLAB/Simulink Toolkits Make Large Models "More Engineering-Savvy"
Engineers can utilize the MATLAB Model Context Protocol (MCP) Core Server universal protocol to bridge MATLAB/Simulink with any commercial large model (such as OpenAI, Claude) or open-source large model. To prevent AI Agents from "going astray" (e.g., repeatedly implementing existing toolbox functions or writing verbose code), the new MATLAB Agentic Toolkit and Simulink Agentic Toolkit play the role of "expert guidance."
"These two toolkits input rich style specifications and expert-level contextual knowledge (Skills) into the Coding Agent. This enables the AI to understand, build, and edit complex models in real time with extremely low token overhead," Tom Erkkinen explained.
He further took the whole-vehicle modeling of an F1 racing car as an example. Guided by detailed prompts, the AI Agent can autonomously create architectures including the Engine Subsystem and Energy Recovery System (ERS), and use the System Composer architecture tool to define interfaces and ports early on, flexibly adapting to distributed or centralized EEA deployments from low-level smart actuators to high-level central brains (HPC).
-
Production-Ready: Integration of AI and MATLAB/Simulink Brings Ultimate Optimization and Static Defense
"In addition to model construction, GenAI has significantly accelerated model refactoring and maintenance. In a real-world case, the AI successfully simplified the structural depth of a messy Simulink model from 4 layers to 2 layers, and reduced the number of containers from 16 to 8, making the top-level interface clearer and fully explicitly wired," Tom Erkkinen mentioned in his presentation.
At the code level, the AI Agent has demonstrated the ability to write 4,000+ lines of MATLAB code in a single pass and automatically generate test cases. To prevent code defects, MathWorks has launched the new Polyspace as You Code plugin in its latest product matrix, shifting code defense left to the engineer's IDE editor, performing real-time static rule checking while writing C/C++. Coupled with the uniformly configured Polyspace desktop application, Polyspace Bug Finder custom checkers, and the software sanitization dynamic analysis capabilities of Polyspace Test, it can comprehensively ensure that security vulnerabilities and defects in AI-generated content are nowhere to hide.
-
System Verification: Rapid Closed-Loop of Multi-Domain Simulation and Test Scenarios
Simulation is the only means to verify the multi-domain interactions (including electrical, mechanical, thermal, hydraulic, and software) of GenAI components and expose faults as early as possible. New features enhance the simulation capabilities of end-to-end system behavior through products such as the Wireless Network Toolbox and the new filter design analyzer (Signal Processing Toolbox).
Engineers can now use GenAI to rapidly reverse-engineer and create 3D test scenarios directly from real-world driving logs, and use MATLAB Test to automatically generate initial tests and equivalence tests based on command history, which will greatly compress the time cost of system-level integration and testing.
Role Evolution and Future Roadmap
As the MathWorks GenAI roadmap advances, Tom Erkkinen revealed that the company will further optimize CI/CD pipelines, introduce sandbox execution to improve scalability, and expand the Agentic Toolkit to more product lines such as Polyspace and Roadrunner, maintaining two high-quality releases and rapid iterations per year.
Image source: MathWorks
Tom Erkkinen believes that in this transformation, the daily roles of automotive engineers are bound to undergo a fundamental shift. Repetitive code writing, construction of basic test cases, and preliminary model refactoring—these foundational tasks—will be handed over to AI Agents proficient in the MATLAB/Simulink tool libraries for automatic execution within a controllable scope.
Future automotive software engineers will transition into higher-level "architects" and "reviewers." Engineers within organizations will need to focus their efforts on defining system goals, designing top-level architectures, writing precise domain-specific prompts (Skills based on Domain Knowledge), and conducting strict reviews and confirmations of the models and simulation data produced by AI.
In addition, he mentioned that although international standards (such as the ISO 22440 AI tool certification standard, which is still in the draft discussion stage) are still evolving with difficulty, the integrated practice of "trusted tools + GenAI" has undoubtedly built a solid bridge to the future for automotive engineering teams, balancing quality and speed, systems and software, and trusted processes and agentic engineering.