The Dark Factory Is Not Science Fiction

Lights Out

FANUC's Oshino factory has been producing robots without human operators on the floor since the early 1990s. Xiaomi's Changping facility in Beijing produces one smartphone per second with minimal human involvement. These are not pilot projects -- they are production facilities generating revenue with the lights off.

What they have in common: narrow product scope, fixed processes, and minimal variability. The hard version of dark manufacturing -- high-mix production with frequent changeovers -- is a different problem entirely. And the gap between what exists today and that harder version is less about better robots and more about better coordination.

It Already Exists (In Narrow Forms)

The FANUC factory produces approximately 50 robots per 24-hour shift, with human workers only visiting periodically for maintenance and quality audits. NVIDIA's PCB manufacturing lines have operated in lights-out mode for years. Philips' razor factory in Drachten, the Netherlands, operates with a robot-to-human ratio of 14:1.

The common characteristic: single product or very narrow product family, on highly specialized lines. Purpose-built machines. Fixed processes. Minimal variability. When the process is stable and the product does not change, you can hard-code the coordination logic into the automation itself.

The hard version -- high-mix, low-volume production with frequent changeovers, variable quality requirements, and multi-system coordination -- is what most manufacturers actually need. And that requires something fundamentally different from hard-coded automation sequences.

The Scaling Problem

A single robotic cell running lights-out is an automation problem. An entire factory running lights-out is a coordination problem.

Consider what happens when Machine 12 on Line B detects a tool wear condition at 2:47 AM. In a staffed factory, an operator walks over, inspects the tool, decides whether to continue or swap, adjusts offsets if needed, and informs the shift supervisor if production is impacted. In a dark factory, something else needs to make all of those decisions.

Not just the immediate decision — "swap the tool" — but the cascade of downstream decisions. Which robotic arm retrieves the replacement tool from the presetter? Is the replacement tool available, or does it need to be pulled from the tool crib? While Machine 12 is down for the swap, can the upstream buffer absorb the interruption, or should Line B be paused to prevent WIP buildup? If Line B pauses, does that affect the shipping schedule for the 8 AM truck? Should the logistics AGV reroute?

Every one of these decisions requires context from a different system. The tool management system. The inventory system. The production scheduler. The logistics planner. A human operator navigates these systems intuitively, making dozens of small decisions based on experience and judgment. A dark factory needs a digital twin that holds all of this context and makes those decisions formally.

The Technical Requirements

Building a dark factory — a real one, not a demo line — requires four technical capabilities, each non-trivial.

The IoT Sensor Mesh

Every machine, every conveyor, every AGV, every environmental zone must be instrumented. Not just for monitoring — for closed-loop control. That means vibration sensors on spindles, thermal imaging on electrical panels, vision systems on quality stations, force-torque sensors on robotic grippers, and environmental sensors for temperature, humidity, and particulate count in clean environments.

A typical high-fidelity sensor mesh generates 10,000+ data points per second across a facility. The networking infrastructure alone — industrial Ethernet, TSN (Time-Sensitive Networking) for deterministic latency, edge compute nodes for local processing — is a significant engineering project. OPC-UA and MQTT are the standard protocols, but standardization at the protocol level does not mean standardization at the semantic level. A temperature reading from a Siemens PLC and a temperature reading from a Fanuc controller need to mean the same thing in the digital twin.

The Real-Time Digital Twin

The sensor mesh provides data. The digital twin provides meaning. It maintains a continuously updated model of every physical entity in the factory: machines, tools, materials, WIP, robots, AGVs. Not just their current state, but their relationships, constraints, and predicted future states.

This twin must be ontology-driven. It cannot be a flat database of sensor readings. It must understand that Machine 12 is a CNCMillingCenter, that it is currently executing Operation 47 of WorkOrder 8892, that Operation 47 requires Tool T-2201, that T-2201 has a remaining life of approximately 35 minutes based on current feed rates and material hardness, and that WorkOrder 8892 has 4 hours of operations remaining — so the tool will need replacement mid-job.

Without this semantic layer, the sensor data is noise. With it, the data becomes decisions.

Autonomous Decision Loops

A dark factory operates on three decision loops with different time horizons.

The reactive loop runs in milliseconds. A force sensor detects an anomaly during a robotic pick — the gripper adjusts pressure, re-grasps, or aborts. This happens at the PLC/controller level and is well-understood industrial automation.

The tactical loop runs in seconds to minutes. Machine 12 needs a tool change — the system selects the replacement, dispatches the tool-change robot, adjusts the buffer strategy for downstream stations, and updates the production schedule. This requires the digital twin and ontological reasoning.

The strategic loop runs in minutes to hours. Demand shifts, a supplier delays a shipment, energy prices spike during peak hours — the system re-optimizes the production plan, adjusts shift scheduling (for the machines, not the humans), and rebalances workloads across lines. This is where AI planning algorithms, fed by the digital twin's current state, generate new production strategies.

All three loops must operate without human intervention on the factory floor. The first two are achievable today with sufficient engineering. The third is where the frontier lies.

Exception Escalation

No autonomous system handles every situation. The dark factory needs a formal exception-escalation framework: a taxonomy of situations the system can handle autonomously, situations that require remote human approval, and situations that require physical human intervention.

A tool change is autonomous. A machine fault that requires physical repair triggers a remote alert to on-call maintenance staff with a predicted arrival window. A fire alarm triggers an immediate system shutdown and emergency response.

This taxonomy must be exhaustive, formalized in the ontology, and continuously updated as the system encounters new situations. The edge cases are where dark factories fail, and the ontology is where those edge cases get formalized into handled scenarios.

What Is Possible Today

Today, lights-out operation is production-ready for:

• Single-product, high-volume lines with fixed processes (electronics assembly, packaging, simple machining)
• Individual robotic cells within larger staffed facilities (CNC machine tending, welding, palletizing)
• Continuous process manufacturing with established automation (food processing, chemical production, pharmaceutical compounding)
• Lights-out shifts — running the third shift unmanned on stable jobs, with operators reviewing in the morning

The technology for these use cases is mature and the economics generally work out. High-volume, low-variability production lines are the natural starting point for lights-out operation.

What Is Five Years Out

Within five years, the following will be achievable for early adopters:

• Full-facility lights-out operation for medium-mix production (tens of product variants, not hundreds)
• Autonomous changeovers — the system reconfigures tooling, fixtures, and programs for a new product without human intervention
• Self-healing production — the system detects quality drifts and adjusts process parameters automatically, rather than stopping for human investigation
• Autonomous material logistics — AGVs and AMRs managing raw material delivery, WIP transport, and finished goods staging without human dispatching

What will not be achievable in five years: high-mix, high-complexity manufacturing with frequent engineering changes and novel materials. Aerospace job shops, prototype labs, custom fabrication — these still need human judgment in the loop. The creative, adaptive, judgment-heavy work remains human. The repetitive, rule-governed, coordination-heavy work moves to the machines.

The Knowledge Formalization Problem

The manufacturers who move toward lights-out operation first will not necessarily be the ones with the most robots or the biggest automation budgets. They will more likely be the ones who have done the hard work of formalizing their operational knowledge -- capturing the decisions that operators currently make intuitively, based on experience, and encoding them somewhere a machine can execute them.

A factory that still runs primarily on tribal knowledge -- the operator who "knows" that Machine 12 makes a funny sound before it fails, the scheduler who "just knows" which jobs to prioritize -- has a knowledge formalization gap to close before autonomy becomes practical. The robots and sensors are largely good enough for most single-cell applications. The coordination layer -- formal enough to replace the operator's intuition across a full facility -- is what most factories are still missing. Ontologies and digital twins are tools suited to that kind of formalization, but the work itself is slow and domain-specific regardless of tooling.