Engineering Physical AI

• The next frontier for AI: physics
• Re-engineering engineering from silicon to systems

Despite algorithmic wizardry and unprecedented scale, the engineering behind AI has been relatively straightforward. More data. More processing. More I/O.

But that’s changing.

With an explosion of investment and innovation in robots, drones, and autonomous vehicles, "physical AI" is making the leap from science fiction to everyday reality. And the engineering behind this leap is anything but straightforward.

No longer confined within the orderly, climate-controlled walls of data centers, physical AI must be engineered — from silicon to software to system — to navigate countless new variables.

Sudden weather shifts. A cacophony of signals and noise. And the ever-changing patterns of human behavior.

Bringing physical AI into these dynamic settings demands far more than sophisticated algorithms. It requires the intricate fusion of advanced electronics, sensors, and the principles of multiphysics — all working together to help intelligent machines perceive, interpret, and respond to the complexities of the physical world.

We have taught AI our languages and imparted it with our collective knowledge. We’ve trained it to understand our desires and respond to our requests.

But the physical world presents a host of new challenges. If you ask AI about potholes, it will tell you how they’re formed and how to repair them. But what happens when AI encounters a large pothole in foggy, low light conditions during the middle of rush hour?

Our environment is highly dynamic. But the one, unbending constant? Physics. And that’s why physics-based simulation is foundational to the development of physical AI.

For AI to function effectively in the real world, it needs finely tuned sensors — such as cameras, radar, and LiDAR — that deliver correlated environmental data, allowing physical AI systems to accurately perceive and interpret their surroundings.

Physics-based simulation allows engineers to design, test, and optimize these sensors — and the systems they support — digitally, which is significantly less expensive than physical prototypes. Answers to critical "what-if" questions can be attained, such as how varying weather conditions or material reflectivity impact performance. Through simulation, engineers can gather comprehensive and predictive insights on how their systems will respond to countless operating scenarios.

Equally important to being able to "see" our world is how well physical AI is trained to "think." In many cases, we lack the vast, diverse datasets required to properly train nascent physical AI systems on the variables they will encounter. The rapid emergence of synthetic data increasingly helps innovators bridge the gap, but accuracy has been a concern.

Exciting progress has been made on this front. Powerful development platforms — such as NVIDIA’s Omniverse — can be used to create robust virtual worlds. When integrated with precise simulation tools, these platforms enable developers to import high-fidelity physics into their scenario to generate reliable synthetic data.

Ansys Perceive EM in NVIDIA Omniverse models 5G/6G antenna signals of moving vehicles in Denver, Colorado

Design and engineering methodologies have traditionally been siloed and linear, with a set of hardware and software components being developed or purchased separately prior to assembly, test, and production.

These methodologies are no longer viable — for physical AI or other silicon-powered, software-defined products.

Consider a drone. To fly autonomously, avoid other objects, and respond to operator inputs, many things must work in concert. Advanced software, mechanical parts, sensors, custom silicon, and much more.

Achieving this level of precision — within imprecise environments — can’t be achieved with traditional methodologies. Nor can it be delivered within the timelines the market now demands.

Digitally enhanced products must be designed and developed as highly complex, multi-domain systems. Electrical engineers, mechanical engineers, software developers, and others must work in lockstep from concept to final product. And their work must accelerate to meet shrinking development cycles.

Ansys electromagnetic simulation software within a rendering of downtown San Jose in NVIDIA Omniverse with 5 cm resolution

The complexity of today’s intelligent systems demands solutions with a deeper integration of electronics and physics. Engineering solution providers are moving fast to meet this need. Our recently closed acquisition of Ansys, for example, combines the leaders in silicon design, IP, and simulation and analysis. Together, we can deliver holistic system design solutions for customers to rapidly innovate AI-powered products.

Our ways to innovate must be as multidimensional and dynamic as the world we live in, and with this, traditional engineering processes must be re-engineered. It will be our key to unleashing physical AI and enabling technologists to achieve their next breakthrough.

This article originally appeared in Electronics Weekly