Physical AI doesn't need a bigger brain. It needs a nervous system

This year, the World Cup united elite athletes from around the world in competition at the highest level.

As I watched, I kept returning to one question: What makes these athletes elite? Knowledge, fitness, or desire to win don’t fully explain it. It’s something else. It’s mastery of the task. Execution at the highest level when it matters most is what separates the elite from the rest.

I’ve been thinking about that word, “elite”, a lot lately. Physical AI is a hot topic and deservingly so as robots, drones, and autonomous machines in all sorts of shapes are emerging. If we’re going to bring all these intelligent systems into our lives, what does elite look like for them?

Excellence is not thinking harder

If you’re a football fan, you’ve seen Messi play. He handles the ball so effortlessly it looks like cheating. The moves he makes seem to anticipate what the opponent will do. We assume that kind of precision must come from thinking faster. It’s the opposite. The decision is too quick for thought at all, instinct takes over.

Look closer at that idea of “doing without thinking,” and you find it everywhere. Keeping balance, blinking, or navigating a crowded room, all of it with no deliberation and minimal energy. And yet for machines this is the hardest thing of all.

The real world changes the rules on AI

There is an old observation in robotics worth revisiting. The tasks we consider the pinnacle of human intellect like solving equations, parsing logic, or recalling facts are the ones machines can do the most easily. The tasks a toddler masters, like grasping a cup, staying upright, or stepping over obstacles are the ones that have challenged AI engineers for decades.

The lesson in that reversal is simple: Intelligence isn’t only something you scale by making the brain larger.

Most of the intelligence we’ve built so far lives in the cloud. Ask it a question and it can afford a moment to think. The physical world grants no such patience. A drone correcting for a gust of wind, a vehicle reacting to a child stepping off a curb, a robot catching a part before it hits the floor, none of these can wait for a round trip to a distant server. The response must happen where the event happens, in milliseconds, every time.

And the stakes are different in kind, not just degree. A wrong answer in software can be corrected. A wrong move in the physical world cannot. You cannot patch a collision or roll back a fall. When intelligence must act in the physical world, then latency, efficiency, and reliability stop being features and become survival requirements.

A blueprint from biology

Biology has been refining intelligence in the physical world for hundreds of millions of years, and the design it arrived at is telling: Intelligence is distributed. The cerebrum handles reasoning and conscious thought, the cerebellum handles motion control and balance, and the spinal cord handles reflexes.

Consider a stumble. By the time our conscious mind registers it, we’ve already caught ourselves. That correction never reached the part of the brain that reasons; it would have been far too slow. Reasoning, coordination, and reflex are separate, independent, and aligned systems, each placing the right intelligence in the right place with the right performance to get the job done.

That is the principle we should be designing toward. Machines are made more capable through the right kind of intelligence in the right place. That’s what a distributed, neural-axis architecture makes possible.

Take the human body as inspiration, and our own neural axis architecture focuses on three independent but highly coordinated layers. A slower, highly intelligent reasoning layer, an intermediary coordination layer, and an ultra-fast reflex layer at the edge. Each playing its part.

Picture a humanoid robot carrying a package across a warehouse. Its reasoning layer sets the destination; its coordination layer keeps it steady as it walks. Then an item falls from a shelf and strikes its arm. Within milliseconds, the reflex layer must rebalance, adjust its grip, and secure the package. Those reflex-like decisions can only happen at the edge, in real time—putting the right intelligence in the right place is what lets machines operate safely.

Making sense of the physical world

Reacting quickly is necessary, but not sufficient. A machine also must understand the world around it. Weight, momentum, friction, the way an object will topple if pushed off center. Children do not learn any of that from a manual. They internalize real world physics from trial and error and parental corrections. But a robot does not have this luxury, because in the physical world, trial and error can do real harm.

Messi didn’t learn any of that from a manual; he absorbed it over thousands of hours. A robot cannot learn the same way.

This is why the most interesting work in the field is teaching machines to anticipate rather than merely respond. Vision-Language-Action (VLA) models are a bridge between perception and action: connecting what a machine sees, what it is told, and how it moves.

That capability is difficult to create and even harder to optimize to run within the tight power and memory limits of an edge device. Closing that gap, between what is possible in a data center and what is practical in the real world, is one of the defining engineering challenges of this decade.

For the physical AI industry, overcoming this challenge means getting two things right. First, building capable and energy-efficient processors optimized for every layer of robotic intelligence, from the cerebrum and cerebellum to the distributed sensory nervous system. Second, equipping teams with tools and frameworks to make developing autonomous agentic intelligence at the edge as simple and efficient as possible.

Trust is the price of admission

Because the real world has no undo button, Physical AI will never reach its potential without trust.

We will only welcome these systems into our homes, our roads, and our factories if we believe they will behave predictably when it matters and fail gracefully when something goes wrong. That kind of dependability cannot be sprinkled on at the end. It must be designed in from the silicon up, so that when a system encounters something it never anticipated, it still does the safe thing.

And that trust doesn’t come from a bigger brain. It comes from the architecture. The same distributed design that biology settled on long ago. Reasoning, coordination, and reflex, each operating where it belongs, and each trustworthy in its own right. If that design is good enough to keep us alive in an unpredictable world, it’s good enough for the machines we will live alongside.

Physical AI will earn its place beside us when it finally gets a nervous system—a neural axis built for trust at every layer.

We've featured the best AI website builder.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit