When a future car is being run by dozens of AI agents at once, what matters isn't how smart those agents are. It's who is watching them — and by what rules. That was the message Dr. Stefan Poledna, CTO of TrustMotion, delivered in his AID 2026 keynote. Physical AI inside the car, he argued, is no longer optional. But the AI-Defined Vehicle doesn't replace the software-defined vehicle (SDV) — it's built on top of it, and it isn't safe without it. In the end, what determines safety isn't trust in the model. It's the system architecture that keeps that model bound inside the laws of physics and a hard safety boundary.

by Sang Min Han ＿ han@autoelectronics.co.kr
한글로보기







Poledna opened with a small correction. Most of the room still knew his company as TTTech Auto. The name had changed, he noted briefly; the mission hadn't. Then he moved straight into the subject: safety and determinism as the foundation of Physical AI.







AI is already spreading through the entire car. It handles path planning and object detection on the ADAS side, voice control and predictive routing on the infotainment side. And according to Poledna, it's now reaching into places drivers never think about — network intrusion detection, virtual sensors that fuse radar and LiDAR into data the car wouldn't otherwise have, battery health monitoring, even climate control. The list itself isn't the point anymore. The point is no longer where AI is being used. It's that AI has stopped behaving like a feature and has started becoming part of the vehicle's infrastructure. In Poledna's own words, that's the shift from Edge AI to Physical AI.
He described today's AI as “application-scoped, but not system-aware” — a model on the cockpit side and a model on the perception side, each isolated, each running on its own logic. Physical AI begins the moment that isolation ends.
“It's about orchestrating multiple AI systems safely, deterministically, and in real time to control the physical world. That's the core of it,” Poledna said. And that, he stressed, is exactly why the AI-Defined Vehicle needs SDV rather than replacing it. “It's important to understand the AI-Defined Vehicle is not a replacement of SDV. Rather, the SDV is a foundation for the AI-Defined Vehicle.”
You need to get data out of the car to have anything to learn from. You need over-the-air software updates to deploy what you've trained. And you need the SDV's central and zonal compute just to run the models in the first place. It isn't a choice between one or the other.







One Brain, Three Guardrails

That argument came through most forcefully in a single diagram. Every domain in the car — battery, chassis, ADAS, central compute, zone, security, cockpit, infotainment — gets its own AI agent. All of those agents report up to a single “Super AI Agent.” That agent sits inside a “Safety Monitor.” The Safety Monitor sits inside an outer layer Poledna simply labeled “Safety & Supervision.” Three concentric rings, wrapped around every piece of intelligence in the car. In many ways, it looked like a blueprint for the operating system of an autonomous machine — one built on the assumption that AI cannot be trusted on its own.
Not one brain — one brain under permanent watch. These rings don't move quickly, and that's by design. The fast loop — the constant retraining and tuning of domain agents — is only allowed to move fast because the slow loop underneath it, the outer rings, isn't moving at all.
Poledna made the same point with an example. Smarter AI-driven optimization of a drivetrain and battery system can squeeze an extra percentage point or two of range out of an EV — a gain the driver feels immediately. But that optimization is only permitted inside the safety boundary set by a physical model, a guardrail that rarely changes. The car gets smarter. The guardrail barely moves.
He was equally blunt about what happens without one. Once a car's functions are exposed as an API for agentic AI to call, you can't hand a model unrestricted access to whatever the vehicle can do. He referenced incidents in which an unsafe command chain disabled a vehicle's headlights on a dark highway and led to a crash — cases, he said, that most of the audience had likely already heard of. The driver survived. It should never have been possible in the first place. The fix isn't trusting the agent more. It's a guardrail that checks every command before the vehicle is allowed to execute it.








Building the Guardrail Into Silicon

Architecturally, this pushes middleware well past scheduling CPU workloads. It now has to arbitrate GPUs, NPUs, and AI accelerators alongside the safety models watching over them — deterministic resource management, mixed-criticality execution, end-to-end latency guarantees, treated not as best efforts but as hard requirements.
Poledna showed what that looks like in silicon with NXP's S32N7: compute organized by domain, a safety-and-security layer running underneath all of it, and an external NPU for the heavier AI workloads — a structure where safety alignment, resource routing, and failure isolation become properties of the chip itself, not of the software running on top of it.
TrustMotion's own MotionWise is being extended onto exactly this kind of hardware, as a small family of products — MW Schedule, MW Communication, MW Orchestrate, and the underlying MW Safety Middleware — built so that software developed once can move between central and zonal ECUs with less validation effort each time.







Poledna closed by turning to his own team's development process. A six-step loop — observe, learn, solve, orchestrate, deploy, measure — applied not to the vehicle but to the orchestration layer itself, continuously retuned against real execution data.
His final slide left the room with an invitation: bring your AI workloads, build on target hardware, optimize safely together, deploy with confidence.
The industry spent the last decade arguing about what a software-defined vehicle even is. The next debate may be far more difficult.
Inside a machine that can kill someone if it fails, who supervises the dozens of AI systems running inside it? Poledna's answer was unambiguous. AI doesn't replace SDV. It depends on it. And in the end, what protects the person behind the wheel isn't the model's intelligence — it's the physics underneath it.

 

AEM(오토모티브일렉트로닉스매거진)