Humanoid robot demos are everywhere and capital is pouring into industrial foundation models. "Let's bring robots into our plant" conversations have started. But the question that decides success with physical AI is not the robot's spec sheet.
"Is your factory a place machines can read?"
Robots can't read paper logs
Physical AI perceives, decides, and acts. But no matter how good a robot's cameras and lidar are, they cannot understand a factory on their own:
- What this line is producing right now (item, lot, work order)
- What state the neighboring equipment is in (running, stopped, alarming, overheating)
- What is forbidden in this zone (safety interlocks, work rules)
Human workers know these things through experience, paper logs and word of mouth. Robots don't. For physical AI to work, the factory's state must flow as machine-readable data. That is the real infrastructure to build before the robot arrives.
Five marks of a ready factory
1. Real-time telemetry — the layer that extends a robot's senses. A robot's onboard sensors see only its surroundings. The flow of the whole line — an upstream stoppage, conveyor speed, buffer inventory — must be supplied by the plant's data layer, continuously and at sub-second latency.
2. A machine-readable map — asset hierarchy and a unified namespace. "The packer next to charger #3" is a human address. Machines need a standard hierarchy (ISA-95: plant → line → equipment → tag) and a unified namespace (UNS) where every system finds data under the same name. Without it, every robot becomes another custom integration project.
3. A common language — standard interfaces. If the robot vendor, the AMR vendor and your legacy equipment each speak their own dialect, integration costs will exceed acquisition costs. A plant that publishes and consumes data over OPC-UA and MQTT has a socket ready for whatever physical AI shows up.
4. A virtual training ground — the digital twin. Physical AI learns by failing, and you can't let it fail on a live line. A digital twin that reflects layout, behavior and data lets you train in simulation and validate changes before they touch reality. The twin, too, is built from accumulated plant data.
5. Control and audit — how much do you allow the AI? AI that moves the physical world causes accidents when it goes wrong. Read widely, write only through approved paths — with every access logged. This is not an option to bolt on later; it must be designed into the data layer from day one.
Getting the order wrong is expensive
A plant that buys robots first and builds infrastructure later pays twice — once for the robots, and once for the custom integration work while the robots sit idle. A plant with the data infrastructure in place doesn't "install" physical AI; it plugs it in.
The good news: none of the five items above is a physical-AI-only investment. Real-time acquisition, asset hierarchy, standard interfaces, digital twins, access control — all of them serve today's OEE, quality and energy management as-is. Preparing for physical AI is simply putting your plant data infrastructure in order, and that investment pays back before any robot arrives.
Closing
The physical AI era will belong not to the companies that pick the best robot, but to the ones that made their factories machine-readable first. PlantPulse provides that foundation — 42 protocol drivers, an ISA-95 asset hierarchy with UNS, OPC-UA re-serving, digital twins, and approval-based control with full access logging — in one on-premises platform. Before the robots arrive, make your factory readable.