AI On the Screen, AI Off the Screen
For the past three years, the AI that manufacturing experienced lived mostly "on the screen." AI that summarizes documents, writes code, and queries data to produce reports — useful, but picking things up and moving them, or stopping a line, was still the job of people and traditional automation equipment.
Physical AI crosses that boundary. It is AI that perceives the physical world through cameras, lidar, and torque sensors, and acts on what it perceives — moving a robot arm, rerouting an AMR, adjusting equipment parameters. Since NVIDIA declared it "the next wave of AI" in its CES 2025 keynote, investment in humanoid robots and industrial foundation models has been exploding.
The reason manufacturing sits at the center of this current is simple: the factory is where repetitive work in the physical world is most densely concentrated.
The Three Layers That Make Up Physical AI
1. World Models — Foundation Models That Understand the Physical World
Just as an LLM predicts the next token of text, a World Foundation Model predicts the next state of the physical world. Whether an object bounces or shatters when it falls, whether a robot arm moving at this angle will collide with a part — a model that has learned physical laws and spatial relationships.
The training data for these models is not internet text but simulation and real sensor streams. The sim-to-real pipeline — running millions of trial-and-error iterations virtually on synthetic data, then calibrating with real field data — has become the standard architecture.
2. Digital Twins — The Training Ground for Physical AI
You cannot train a robot by trial and error on a live production line. The cost of a single defect or a single collision is too high. So most physical AI training happens inside a physically accurate digital twin.
What matters here is that a digital twin's "accuracy" is determined not by the visual quality of the 3D model but by the fidelity of its data. If the equipment's actual cycle times, actual vibration spectra, and actual temperature curves do not flow into the twin, the AI trained inside it collapses in the real world. A robot that learned in a fake factory cannot work in a real one.
3. The Execution Layer — Robots, AMRs, and Existing Equipment
The execution layer of physical AI is not just humanoids. In fact, the clearest near-term ROI comes from:
- Vision-based quality inspection: closed-loop quality control in which learning-based vision detects micro-defects and the results immediately drive process parameter adjustments
- Autonomous logistics (AMR/AGV): logistics robots that change their own routes and priorities by reflecting production plans, inventory, and equipment status in real time
- Adaptive robot cells: robot arms in high-mix, low-volume production that change their grasping strategy on their own, with no re-teaching, every time the item changes
- Autonomous equipment control: equipment that, when it detects signs of an anomaly, lowers its load or stops on its own, subject to human approval
See the common thread? None of these four works without a real-time flow of shop-floor data.
Physical AI's Real Bottleneck Is Not Robots — It's Data
Many companies approach physical AI as a "robot procurement project." But in real pilots, what sinks the project is not the robot hardware — it is everything upstream of it.
First, the perception gap. A robot knows the state of its own joints, but it does not know what state the equipment next to it is in, or what material just came onto the line. That information is scattered across PLCs, MES, and sensors, mostly in formats the robot cannot read.
Second, the context gap. For a signal like "abnormal vibration on press no. 3" to turn into robot action, you need an ontology (asset model) that knows which line press no. 3 belongs to, what item it is currently producing, and what the next process step is. If meaning is not linked behind the tag names, AI cannot translate signals into actions.
Third, the latency gap. A data warehouse that gathers data once a day in batches is useless to physical AI. The perceive-decide-act loop runs on milliseconds to seconds. Event streaming and CEP (complex event processing) are preconditions.
In other words, physical AI adoption starts not with buying robots but with putting the data infrastructure in order:
- Connect — collect signals from PLCs, equipment, and sensors over standard protocols (OPC-UA, MQTT/Sparkplug B)
- Add meaning — link asset and process context to signals with a UNS (Unified Namespace) and an ontology
- Go real-time — process events at millisecond granularity with streaming pipelines and CEP
- Sync the twin — stream real data into the digital twin to minimize the gap between simulation and reality
- Connect to action — put robot, AMR, and equipment control on top of it all
The Opportunity for Korean Manufacturing
Physical AI is a game that favors Korean manufacturing. World-leading robot density (no. 1 globally in industrial robots per 10,000 employees), precision manufacturing know-how, and an industry portfolio of physically complex sectors such as shipbuilding, defense, and secondary batteries — all of these are conditions under which the value of physical AI is maximized.
The deciding factor is the speed at which shop-floor data is made ready in a form AI can use. Robots and world models can be bought off the shelf; your own factory's data infrastructure cannot. A foundation that gathers equipment signals scattered across 42 protocols, gives them meaning through an ontology, and processes hundreds of thousands of events per second in real time — that is the ticket of entry to the physical AI era.
AI that steps off the screen ultimately walks on data. Turning every signal in the factory into the pulse of the business comes first.