For a stretch of the war, Israelis opened a navigation app in the center of the country and found themselves, according to the screen, apparently in Lebanon or Iran. The streets had not moved. The invisible map above them had.
It was a strange civilian side effect of something engineers and soldiers understand well. We live inside an electromagnetic environment that we almost never notice. Radio waves carry our calls, fix our position, and connect our machines, and most of the time we notice them only as a bar on a screen. As we hand more decisions to autonomous cars, drones, and robots, that invisible environment becomes part of the physical world AI has to learn to navigate.
That matters more now, because Israel has just decided that machines acting in the real world are a national priority.
Israel's new national AI plan names physical AI as a strategic niche: machines that perceive, decide, and act in the real world, across transportation, health, space, and agriculture. It calls for sandboxes and a national physical testbed, a controlled, instrumented environment where autonomous systems can be validated under conditions representative of reality before they are trusted in it.
But the plan raises an engineering question that sounds simple but is not: What counts as the physical world?
Roads count. Weather counts. Walls count. People count.
Radio waves should count too.
That last line is my argument, not the plan's. The document does not explicitly say its testbed will include the electromagnetic environment. I think it should be part of the conversation, because a wall is not only something a robot's camera sees. It is also something a radio signal has to pass through, reflect off, or bend around.
We experience Wi-Fi or cellular as an icon showing signal strength. The physics underneath is far less tidy. Radio behaves a little like invisible light with bad manners. Some surfaces block it, others reflect it, and the route from transmitter to receiver is rarely a straight line. A single signal can arrive several times over, through different paths, a fraction of a moment apart.
Engineers have even learned to turn that mess into an advantage. Modern wireless systems use multiple antennas to exploit the fact that radio takes different routes through space, turning some of that complexity into extra capacity. And that is the catch. If the technology depends on the physical environment, then testing it means understanding that environment.
To an autonomous machine, none of this is abstract telecom theory. It shapes what the machine can transmit, where it thinks it is, and which other machines it can coordinate with. And unlike a wall, the electromagnetic environment can be changed on purpose, by someone who wants your machine to fail.
Ukraine is the extreme case. Defense researchers at RUSI have documented widespread jamming of navigation and command frequencies, and an almost old-fashioned answer to it: small drones that trail a thin fiber-optic cable, trading the freedom of a wireless link for resistance to jamming.
The same research notes that even friendly radio-controlled drones can interfere with each other when too many crowd the same airspace. Interference is not only an enemy weapon. It is a property of a crowded spectrum.
The lesson is not that autonomy needs a perfect connection. Often the opposite is true. When Earth and Mars are separated by a signal delay that can reach around 20 minutes each way, NASA cannot drive a rover like a remote-control car. It has to push more of the thinking onto the machine itself. Autonomy is partly what we build for the moments the network cannot be trusted. But the network is still part of the world the mission has to understand.
The lesson is not that autonomy replaces communication. It is that the two become more important together. Returning to the example above, the rover must be intelligent enough to make decisions on its own. But a rover that explores Mars and can never send its discoveries home has lost much of the purpose of the mission. Its autonomy allows it to act; communication gives those actions meaning beyond the machine itself.
The same is true closer to Earth. A single autonomous robot can navigate, decide, and adapt alone. But put 10 robots on the same mission and communication becomes what allows these intelligent machines to behave like one intelligent system: sharing what they see, dividing tasks, warning one another, and changing plans together. Humans work the same way. Intelligence lets us solve problems; communication lets us solve problems together.
Fortunately, another part of the AI world already shows how to prepare for conditions too rare or too dangerous to meet for the first time in the field. Self-driving companies do not wait for every strange road situation to happen at the worst possible moment.
Waymo says its system has driven billions of miles in virtual worlds, and its latest simulation can produce rare scenes on demand, from severe weather to unexpected objects in the road. NASA checks rover commands against a digital twin before sending them across that long delay to Mars.
Radio deserves the same treatment. The tools already exist to model how signals travel: terrain and geometry, materials, antenna patterns, the positions of transmitters and receivers. More and more, that modelling is being tied to digital twins of real places. The next digital twin should map not only where the walls are, but what the walls do to a signal. The goal is not to replace field testing. It is to walk into the field having already failed, safely, thousands of times in simulation.
That is the opportunity sitting inside Israel's physical-AI push. If we want to be the place where physical AI proves itself, our test environments should eventually represent more than what a machine can see and touch. They should also challenge what it can hear, transmit, and locate: coverage shadows, competing transmitters, positioning that drops out, and, where it is relevant, deliberately hostile conditions. This is not a departure from the government's plan. It is the same logic, taken one step further.
For most of us, the radio world only becomes visible when something breaks. A call drops. Wi-Fi disappears. A navigation app suddenly decides that Tel Aviv is somewhere in Lebanon. Autonomous machines will not have that luxury. We are teaching AI to see the world, reason about it, and move through it. The next step is to teach and test it in the invisible world too.
Daniel Ferber is co-founder and CEO of RFix.ai, an Israeli company developing RF simulation technology for testing wireless systems and electromagnetic environments.