AI models, computer vision, reasoning, learning, autonomy... The “brain” of a robot features in almost every conversation about robotics. And that makes sense: without this intelligence, no machine can adapt to a real-world environment. But when we talk about operating a robot in a factory, hospital, mine, port or critical infrastructure environment, another variable comes into play. It receives far less attention, yet it can be even more decisive: latency.
In the digital world, a slight delay is an inconvenience. In the physical world, it can be a risk. A machine that moves with a delay, a signal that arrives late, an instruction executed with inconsistent timing, a remote operator who does not receive stable feedback... In every case, the risk is operational, and the implications are serious. These are just a few examples of why latency is essential to safety, productivity and trust.
Physical AI requires machines to sense, decide and act in real time. For me, that is precisely where its value lies: integrating intelligence into machines and operations so that factories, warehouses, cities and critical infrastructure can respond faster, more safely and at scale.
This requires us to rethink infrastructure. When we are talking about remote control, collaborative robotics or autonomous machines operating near people, we can no longer evaluate connectivity based only on capacity more bandwidth, broader coverage or higher theoretical speeds. We need networks that are stable, predictable and deliver low delay variation. In robotics, jitter can be just as important as average latency.
When every millisecond counts
This is where NTT Group’s vision for the Innovative Optical and Wireless Network (IOWN) becomes particularly relevant. In November 2022, a demonstration in Musashino connected two points more than 100 km apart using the IOWN All-Photonics Network (APN), with capacity of more than 100 Gbps per wavelength across certain segments, low latency close to the physical limit and virtually no delay variation.
Later, in February 2025, NTT and Medicaroid connected a hinotori surgical robot between two hospitals over the same network to explore remote-surgery scenarios. The key point is not simply the ability to transmit large volumes of information. It is the ability to do so with enough stability for two physically separate environments to operate as though they were integrated.
In surgery, latency is not an abstraction. Japan’s 2022 remote-surgery guidelines set an end-to-end limit of 100 ms. A five-year validation of the hinotori system, conducted between 2020 and 2025 by Kobe University, Medicaroid and NTT DOCOMO across more than 30 remote sessions, achieved approximately 100 ms round-trip latency over 5G Sub6 while simulating a distance of 800 km, with stable stereoscopic video even when 1 Gbps of network congestion was introduced.
Above that 100 ms threshold, according to a 2025 article in the Journal of Robotic Surgery, hand-eye coordination begins to deteriorate. Another 2025 article in the International Journal of Computer Assisted Radiology and Surgery found that, in the 150 ms–300 ms range, surgeons begin to perceive a loss of control.
Expertise without physical boundaries
The implications for businesses are significant. If intelligence can be distributed and processed close to where it is needed, we are talking about new operating models not just more autonomous robots.
For example, this could allow an expert to supervise critical machinery remotely, enable an operator to intervene in hazardous environments without being physically exposed to the risk, or give a hospital access to advanced capabilities without depending on specialists being physically present.
To make this possible, we need to think of the network as part of the robot. It is not simply supporting infrastructure; it is an extension of the operation’s nervous system. A machine sees, interprets, calculates, moves and responds as part of a chain that includes sensors, edge computing, cloud, models, connectivity, enterprise systems and people. If any part of that chain introduces uncertainty, autonomy becomes fragile.
The business conversation starts with the operational SLA: What response time does the process require? What margin of error is acceptable? Which tasks can operate autonomously? Which require human oversight? And what happens if connectivity fails? With Physical AI, operational design needs to account for both intelligence and the possibility of failure.
There is a human dimension too. Teleoperation, collaborative robotics and autonomous systems do not replace people. Instead, they can extend expert knowledge to places it could not reach before, enable high-risk tasks to be redesigned and help teams work with machines that do more than execute tasks they also interpret their environment.
The next frontier will be building operations where intelligence, connectivity and trust work as one capability. Beyond how much a machine can think, the future of robotics will be defined, in large part, by its ability to act at the right time.