In control engineering, a loop is closed when the effect of the action is measured and that result feeds back into the next decision. This entry explains that definition, the three actuation patterns, the guardrails that make it acceptable and what the loop demands in data and in governance.
What it is in control engineering
In open loop an action is issued and the result is taken on trust. In closed loop the output is measured, compared with the target, and the difference corrects the next action. The operating criterion is blunt: the loop is closed only if the effect of the action is measured. If the effect is not measured there is no loop, there is a recommender. That is exactly the boundary between level 2 and level 3 of the Physical AI Ladder. A PID controller is the simplest case of the loop: one measured variable and a fixed law. The decision layer described here does not replace it, it sits above it and sets its setpoints.
The three actuation patterns
- Setpoint on an existing loop. The system writes a reference and the local controller follows it: target temperature, flow, contracted power, line speed. It is the lowest-risk pattern: the fast dynamics remain governed by the cabinet controller.
- Discrete action on the sequence. Starting, stopping, switching a machine, changing recipe or diverting product. Here the interlock rules: there are plant states in which the action never executes, and that condition is a deterministic rule.
- Order to a person or to another system. A maintenance work order, an adjustment notice to the operator, a task created in the ERP. The actuator is human or is management software, and the effect is measured just the same.
The guardrails
- Confirmation. Writing a setpoint is not acting. Acting is writing and verifying in the measured signal that the equipment accepted and executed it.
- Maximum data age. Every action declares how old the data justifying it may be. Past that limit it does not execute: it would be deciding on a state that no longer exists.
- Safe default behaviour. With missing data the answer is not to act, unless the opposite is justified in writing and tested.
- Independent envelope. The safety layer and the interlocks remain deterministic rules and do not depend on the model. With robots involved, the applicable framework is ISO 10218-1:2025 and ISO 10218-2:2025, published in February 2025, which also absorb the content of the former ISO/TS 15066, as reported by The Robot Report on 18 February 2025.
- Cost of a false positive. Asking for human confirmation on everything produces alarm fatigue and the operator stops looking. The cost of a false positive is that attention.
What it demands in data
Bounded latency, not just low latency: a loop with variable delay cannot be tuned, either it becomes slow in order to be stable, or it oscillates. The order of magnitude is set by periodic industrial communication, with common cycle times of 10 ms, 1 ms and 0.5 ms according to the 5G-ACIA and ZVEI paper on industrial Ethernet and 5G networks, published in November 2019. It also demands a common clock, which in measurement and control networks is solved with the precision time protocol IEEE 1588-2019, published on 16 June 2020. Without a common clock the history lets you count, not explain. And it demands a baseline: without a declared before and after reference nobody can show that the loop improves anything.
What it demands in governance
Traceability of every decision: input data with its timestamp, version of the rule or model applied, the person who supervised, the command written, the equipment confirmation and the effect measured afterwards. Declared authority: who approves the action, who can revert it and under which conditions it does not execute. Separation of read and write permissions, named access control and a rollback path that has been tested, not merely described. The industrial cybersecurity framework is the ISA/IEC 62443 series, whose part 2-1, in its August 2024 edition, sets the security programme requirements of the asset owner. Writing into the process crosses the responsibilities of production, maintenance and safety: if nobody owns the action, the project stays in pilot.
Related terms
The closed loop is level 3 of the ladder and the prerequisite for level 4 of physical AI, because it is the only one that produces pairs of action and effect. It rests on data acquisition and on the historian, it often runs from edge computing, and its organisational guardrail is human-in-the-loop. When a model decides, industrial AI agents and the monitoring of drift come into play. The architecture that supports it is described in the industrial data platform.