Engineering Physical AI: How Digital Twins and Simulation are redefining Embedded Intelligence

Artificial Intelligence is no longer confined to the digital realm. Beyond generating text, images and code, AI is increasingly operating in the physical world – perceiving, reasoning, acting and adapting in real time. This new frontier is known as Physical AI. It powers autonomous trucks that haul ore across remote Australian mines [1], forklifts that navigate warehouses alongside human workers at Walmart [2] and robots that assemble delicate components with millimeter precision.

Unlike traditional automation, Physical AI does not simply execute pre-programmed tasks. It uses sensory data – from cameras, radar, lidar, inertial measurement units, temperature sensors and more – to understand its environment, make decisions and take action dynamically. In a nutshell, Physical AI takes the power of Generative AI beyond the digital realm – into the tangible, complex and unpredictable physical world.

This shift promises higher productivity, improved safety, greater flexibility and entirely new classes of Autonomous Systems. At Rio Tinto’s Greater Nammuldi iron ore mine in Australia, for instance, more than 50 autonomous haul trucks, each capable of carrying up to 300 tonnes, operate with just one operator overseeing 25 vehicles. Coordinating and monitoring these robots is the company’s Operations Centre (OC) in Perth – located roughly 1,500 km (930 miles) to the south. Since deployment, Rio Tinto reports a ~15 % increase in safety and productivity and across 17 sites more than 360 driverless trucks now make up 84 % of its fleet – a powerful demonstration of the transformative potential of Physical AI in large-scale industrial operations [3].

Physical AI system classes like these also present significant engineering challenges like: How do we design, validate and deploy AI-enabled physical systems that must operate safely and intelligently in real-world environments?

The answer increasingly lies in Digital Twins and simulation-based system validation — technologies that Fraunhofer IESE is pioneering through Fraunhofer FERAL and Eclipse BaSyx.

From Generative AI to Physical AI: A Fundamental Shift

Most people associate AI with Large Language Models (LLMs), image generators or recommendation systems. These technologies excel at manipulating data within a digital context, relying on human input and existing in virtual spaces. But the real world is different. – It is continuous, uncertain and governed by physics. In this regard, Physical AI perceives the world through sensors, processes the input to understand what’s happening and reacts properly in real time. It doesn’t wait for carefully prepared datasets, but instead collects information directly from the environment, learning and adapting over time, improving its performance based on experience.

This is more than a semantic difference; it is a shift in how AI systems are conceived, designed and built, being ready to deal with:

• Real-world complexity: Environments are noisy, dynamic and full of surprises.
• Safety-critical decisions: Failures can cause accidents, not just errors.
• Continuous adaptation: Systems must evolve as conditions change.

These challenges require new engineering approaches, and that’s where Digital Twins and simulation come in.

Digital Twins: The Foundation of Physical AI Engineering

A Digital Twin is a virtual representation of a physical asset, behaving like their real-world counterparts, responding to inputs, interacting with their environment and evolving over time. These capabilities of Digital Twins are essential for Physical AI as it allows engineers to:

• Test algorithms before hardware exists: Predict how a perception or control system will behave under different conditions.
• Run “what-if” scenarios safely: Inject faults, simulate edge cases or explore extreme conditions without risking physical assets.
• Accelerate learning: Use simulated environments to train AI models faster and more safely than real-world testing.
• Continuously validate updates: As software evolves, simulations ensure that new versions remain safe and reliable.

The Hidden Cost of 3D Models – and a Smarter Way Forward

A common assumption is that adopting Digital Twins necessarily involves the use of 3D models. Isha Salian from NVIDIA claimed that “Physical AI development starts with the construction of high-fidelity, physically accurate 3D environments. Without these lifelike virtual environments, developers can’t train advanced physical AI systems such as humanoid robots in simulation, because the skills the robots would learn in virtual training wouldn’t translate well enough to the real world.” [4]

In some cases, complex 3D models are essential – for example when training robots to manipulate objects with precision or navigate cluttered environments. But they also come with trade-offs [6]:

• High computational demands: 3D simulations require significant GPU and cloud resources
• Increased energy consumption: They can generate substantial CO₂ emissions – up to 70 tonnes per year for high-fidelity, physics-rich digital twins.
• Longer development cycles: Complex models take time and expertise to build and maintain.

Fraunhofer IESE’s FERAL platform is a game changer when it comes to creating lightweight Digital Twins and enabling high-fidelity virtual testing and validation without unnecessary complexity and high carbon footprint. To this end, FERAL enables (i) creating non-3D digital twins of embedded assets — controllers, sensors, actuators – and (ii) integrating them into co-simulation environments. FERAL’s capabilities to deal with complex behavior, timing, communication and interactions enables continuous prediction of system behavior long before physical prototypes are built.

As Rodney Brooks of MIT famously put it: “The world is its own best model.” [5] This is the principle on which Fraunhofer FERAL grounds its capabilities towards Physical AI.

Real-World Impact: Proven Use Cases

Fraunhofer FERAL has delivered significant results across industries:

• Bosch: Virtual validation of electronic control unit (ECU) networks reduced release test time by 80% (from two weeks to two days) and sped up component tests by over 30%.
• Mercedes-Benz: Jointly developed tools to precisely characterize DRAM performance for next-generation automotive systems.
• CADFEM: Validated ADAS functions using digital twins and driving simulators like CARLA, enabling systematic fault injection and CAN-bus resilience testing.

Each example highlights the same principle: Simulation-first engineering leads to faster development, lower costs, safer systems and more resilient operations.

Conclusion: Building the Future of Intelligent Machines

Physical AI represents one of the most significant technological shifts of the next decade. It promises machines that understand, adapt and act – transforming industries from logistics and manufacturing to mobility, healthcare and infrastructure.

But unlocking that potential requires rethinking how we engineer these systems. Digital Twins, simulation-based validation and co-simulation are the keys to bridging the gap between code and reality. Here, platforms like Fraunhofer FERAL provide for building a powerful, intelligent, real-world AI without unnecessary complexity – and without waiting for physical prototypes.

The future of AI is not just about algorithms. It is about engineering intelligence into the physical world – and doing so predictively, sustainably and safely.

[1] https://im-mining.com/2025/05/09/thiess-norton-gold-fields-and-eacon-launch-autonomy-trial-in-australia
[2] https://corporate.walmart.com/news/2024/04/11/a-fork-in-the-road-walmart-bets-on-associates-automation
[3] https://www.bbc.com/news/articles/cgej7gzg8l0o
[4] https://blogs.nvidia.com/blog/physical-ai-research-siggraph-2025
[5] https://people.csail.mit.edu/brooks/papers/representation.pdf
[6] https://arxiv.org/abs/2407.14892