With this statement, André Loesekrug-Pietri, Chairman of the Joint European Disruptive Initiative JEDI, captured one of the central challenges of Physical AI at The Autonomous Main Event 2026 in Vienna. His keynote was fittingly titled “Verify, Don’t Trust.”

Powerful AI models are becoming increasingly available. The harder question is how their decisions can be reliably connected to real machines operating in constantly changing physical environments.

An autonomous system may still be running exactly the same software as on day one while its physical body has already begun to change: motors age, friction shifts, sensors drift, tools wear, and loads and environmental conditions vary.

Physical AI therefore needs more than computing power and intelligent decisions. It needs a representation of its environment – and of its own physical state within it.

Europe Is Building the Physical AI Stack

The contributions in Vienna showed just how concrete this stack has already become.

With “The Future of Physical AI is Built on Chips,” Infineon placed the physical foundation centre stage: semiconductors, power electronics, motor control, sensing, safety and security. The underlying message was clear: semiconductors do not merely support Physical AI; they determine essential properties of the physical system.

In the joint panel “From Cars to Robots: Shared Foundations for Safe and Trusted Autonomy,” NXP, Infineon, Bosch, NEURA Robotics and NVIDIA discussed how compute, power, communication, actuation, safety and security can come together to enable scalable and trustworthy autonomy.

NXP connects sensing and actuation with edge compute and central AI processing. TTTech focuses on deterministic communication in distributed systems. In its panel, TTControl explored how Physical AI and autonomous machines can scale in the off-highway sector, from pilot projects to real-world deployment. Bosch brings together sensing, embedded technology, AI and industrial-scale implementation.

The software and deployment layer was also a major topic. The TrustMotion-hosted panel “Building Secure and Scalable Edge AI Frameworks for Autonomous Systems” presented an architecture in which on-device edge AI works together with large data lakes, cloud-based model development, over-the-air updates and continuous data collection. Beyond autonomy itself, this infrastructure is intended to support applications such as predictive maintenance, virtual sensors and performance monitoring.

Many of the critical building blocks for Physical AI are therefore already in place.

One layer, however, does not emerge from them automatically:

a continuously comparable representation of the overall physical system state.

How Does a Machine Represent Its Physical World?

What state is a machine in as a complete system – in this specific context?

The same motor current may be perfectly normal under heavy load and indicate a changed operating state under different conditions. Temperature, torque, vibration or motion therefore derive their meaning not from individual values alone, but from how those signals interact.

An autonomous system also never exists independently of its environment. Load, tool, surface, speed, temperature and task all influence how its signals should be interpreted.

Perception needs sensor fusion. Sensor fusion needs context. And context needs a representation in which relationships can be compared.

This is where SF2 applies Representation Learning and Semantic Folding Sensor Fusion.

Multivariate sensor and operating data are transformed into a semantic representation of the physical system state. “Semantic” does not refer to language here. It refers to meaning created through relationships and context: a signal gains significance through its relationship with other signals and with the current operating situation.

SF2 represents these states as compact, sparse binary State Fingerprints. Similar states produce similar representations. When the interaction between signals changes, drift, transitions and new states become comparable.

In simplified form:

Sensing → Sensor Fusion → Representation Learning → State Fingerprint → Comparison → Action

“Autonomy does not begin with the decision, but with the representation,” says Francisco Webber, CEO of SF2 Systems. “A machine must be able to represent its own physical state and the context in which that state arises in a form that can be compared with what is already known. That representation layer is exactly what we are building with SF2.”

Understanding Begins with Comparison

A current State Fingerprint is compared with known states.

High overlap means high similarity. Decreasing overlap shows that the system is moving away from a known state. A previously unknown pattern can be captured as a new state and recognised again later.

SF2 does not require large labelled fault datasets or conventional supervised fault training. Depending on the use case, the semantic structure can be learned from historical operating data.

Once a model has been fixed and versioned, runtime evaluation is reproducible:

same input → same transformation → same fingerprint → same comparison result.

Not Just Theory: 52 Channels Demonstrate the Principle

SF2 has already demonstrated on industrial data that this approach is more than a concept.

In a real water-pump system with 52 sensor channels and more than 220,000 data points, an analysed sequence revealed a relevant change in the overall system state around four days before a documented pump failure. Building the model from the available data took less than 15 minutes on a standard laptop.

This is not a universal prediction claim, nor is it yet evidence for robotics.

It does, however, demonstrate the underlying principle: the interaction of many signals can reveal a change in system state that cannot be readily inferred from individual measurements alone.

The next step is to validate this principle with real Physical AI and robotics data.

“Europe already has many of the decisive building blocks for Physical AI,” says Christoph Gretzmacher, Business Development at SF2 Systems. “The exciting question now is how these layers work together. We want to use real data to demonstrate the additional value a shared representation of physical state can create across sensing, edge, communication and application layers.”

Trust Is the Outcome – Trustworthiness Must Be Verifiable

This gives JEDI’s “Verify, Don’t Trust” a concrete technical dimension.

Trust itself is not a metric. What can be verified are the properties from which trustworthiness can emerge: safety, security, deterministic communication – and whether a physical system continues to behave like known reference states.

SF2 is neither a functional safety system, a security solution, nor a robot controller.

It answers a different question:

Is the physical machine still behaving as expected – or is its behaviour beginning to change?

The robotics side of the event also highlighted how important this question becomes in real-world operation. In its keynote on autonomous robots in industrial use, ANYbotics framed trust around safety, consistency, transparency and security – particularly when autonomous systems move from pilot projects into daily operation.

Infineon likewise emphasised safety as a fundamental design requirement for Physical AI.

Alongside Safety by Design, this creates a complementary perspective:

State Awareness at Runtime.

And this is where the connection to the other layers of the stack becomes particularly interesting.

Infineon’s motor-control and sensing technologies provide key signals from the physical system – precisely the data from whose interaction a state representation can be derived.

NXP brings computing power close to sensors and actuators – to the very place where compact State Fingerprints can be generated from high-resolution signals.

TTTech provides deterministic communication paths; condensed state information can form an additional information layer between distributed components.

TTControl brings edge compute, sensor fusion and autonomous functions into real mobile machines whose physical state continuously changes under varying loads and environmental conditions.

Bosch brings the experience required to integrate and scale sensing, embedded technology and AI under real industrial conditions – exactly where an additional state representation based on existing data can create practical value.

SF2 brings these physical data streams together at another layer: their interaction becomes a semantic and reproducibly comparable representation of the overall system state.

High-Frequency Data Must Become State

With Physical AI, both data volume and temporal resolution are increasing. Motor control, current and position measurement, vibration, radar and acoustic sensing can generate high-frequency, high-resolution data streams that do not need to be transported permanently through every layer of a system and evaluated centrally.

This is where edge computing and SF2 fit together structurally.

Close to the data source, multivariate signals can be transformed into a compact State Fingerprint. State Fingerprints are sparse binary representations that can be compared using efficient bitwise operations. This makes the approach fundamentally suited to resource-efficient edge processing.

SF2 therefore creates state representation close to the data source: without requiring large labelled fault datasets and without mandatory cloud infrastructure. This does not exclude cloud, data-lake or OTA architectures; it complements them with state information that can be generated locally.

Higher layers can then work with condensed information about state, similarity, drift or transitions, rather than continuously processing every raw data stream.

The more Physical AI perceives, the more important it becomes to turn data into state.

This logic can also be scaled hierarchically:

Component → Subsystem → System → Fleet

A State Fingerprint at one level can become part of the representation at the next level up. As Industrial IoT expands and 5G – and, in future, 6G – infrastructures become more interconnected, this kind of condensation becomes increasingly relevant.

Connectivity alone does not create understanding.

The Missing Layer May Not Be Another AI

Physical AI already has powerful sensing, semiconductors, actuation, edge computing, communication, safety and AI.

The open question is:

How do perception and context become a shared representation of the actual physical system state?

To address this gap, SF2 is developing a lightweight, edge-capable and reproducible Representation Layer that translates multivariate physical signals into a semantic state representation – providing an additional technical foundation for the principle JEDI describes as “Verify, Don’t Trust.”

Autonomy needs representation. SF2 represents physical state.