Manufacturers have spent the last several years investing in artificial intelligence to improve operations, from predictive maintenance to quality inspection. But a new phase of AI is beginning to take shape, one that moves beyond analysis and into physical interaction.
Often referred to as physical AI, this emerging category focuses on systems that can perform real-world tasks, learn from those interactions and continuously improve over time. Instead of working with static datasets, these systems are trained through experience, using sensors and feedback from the environments in which they operate.
That shift changes not only how AI is developed, but what factories must support to make it viable.
In traditional AI models, data is collected, processed in the cloud and used to generate insights. Physical AI introduces a different loop. Machines collect data through cameras, LiDAR and other sensors, send that information to a processing system, receive updates and then repeat the process continuously.
“Now you want to train a robot to do something—carry something, put it somewhere or perform some physical activity,” said Tamer Kadous, General Manager of the XCOM RAN business unit. “That training is happening on-site, in the environment where the system is operating.”
This creates a constant exchange of data between machines and computing systems. Information must move in both directions (sensor data flowing upstream, model updates flowing back downstream) without interruption.
That requirement is placing new pressure on factory connectivity.
Looking for quick answers on assembly and manufacturing topics?
Try Ask ASM, our new smart AI search tool.
Ask ASM
“Everything here requires a communication link,” Kadous said. “You’re not going to pull wires between every sensor and a central controller.”
Wireless networks have long filled that role, but many of the systems currently in place were designed for less demanding use cases. As factories increase the number of connected devices, those networks are being pushed beyond their original limits.
In environments with high device density and real-time requirements, maintaining consistent performance becomes more difficult. Physical AI amplifies those challenges by requiring not just connectivity, but reliable, high-capacity communication that can support continuous learning loops.
Kadous described this as a shift toward more demanding performance expectations across multiple dimensions at once.
“You have communication in the uplink and in the downlink,” he said. “Sensors need to send information to the server, and the server needs to update the model and push it back. That’s a very high-quality link right there.”
Unlike traditional automation systems, where communication may be periodic or one-directional, physical AI depends on sustained, bidirectional data exchange. Any interruption in that loop can affect how quickly systems learn or how accurately they perform. That raises an important question for manufacturers: not just whether they can deploy AI, but whether their infrastructure can support it at scale. Connectivity, in this context, becomes more than a utility. It becomes an enabling layer for how AI systems function.
New approaches to private wireless networks are emerging in response. These systems are designed to handle dense deployments of sensors and machines, while maintaining consistent performance across large facilities. Rather than treating connectivity as a background system, they aim to support the specific requirements of automation and AI-driven processes.
One example of that approach is XCOM RAN, a private 5G system built around a coordinated network design rather than traditional cell-based coverage. Instead of operating radios independently, the system allows them to function together as a single, unified layer of connectivity across a facility.
That structure is intended to reduce interference and eliminate the need for handoffs as devices move through a plant, while also increasing overall network capacity. In environments where machines, sensors and robotics systems are constantly exchanging data, that consistency becomes critical.
“The real requirements [for deploying physical AI] are not necessarily yet known … Systems need to be flexible enough to adapt as those requirements become clearer.”
– Tamer Kadous
As manufacturers begin to explore physical AI, those types of architectures are gaining attention not just for coverage, but for their ability to support continuous, real-time communication between systems.
At the same time, the requirements of physical AI are still evolving. Manufacturers are experimenting with robotics, analytics and sensor-driven systems, but there is no single blueprint for how these technologies will be deployed.
“The real requirements are not necessarily yet known,” Kadous said. “Systems need to be flexible enough to adapt as those requirements become clearer.”
That uncertainty is shaping how companies think about long-term investments. Infrastructure decisions made today must support not only current applications, but future ones that may place even greater demands on performance.
Physical AI represents a shift toward more adaptive, responsive manufacturing systems. Machines are no longer just executing predefined tasks; they are learning, adjusting and improving in real time.
But that evolution depends on something less visible: the ability to move data quickly, reliably and continuously across the factory floor.
As manufacturers look ahead, the pace at which physical AI develops may depend as much on connectivity as it does on the intelligence of the systems themselves.
