A factory can contain intelligent robots, digital twins, AI agents, safety systems and highly skilled people — and still fail as a system when each participant acts on a different view of reality.
Physical AI needs more than smarter machines.
It needs shared operational infrastructure.
A factory is increasingly becoming a system of independently intelligent participants sharing the same physical environment.
Robot perception can improve.
World models can improve.
Planning can improve.
Reasoning can improve.
And the factory can still fail as a system.
The problem appears when multiple autonomous systems, people and operational systems must understand the same situation, coordinate their intentions, respect changing authority and learn from what actually happened.
The bottleneck starts moving outside the individual machine.
This is where an operating layer for shared autonomous spaces becomes necessary.
Imagine a factory where Robot A intends to enter a shared workspace while Robot B has already reserved part of it.
The production system has just increased the priority of Robot A's task.
At the same time, a technician enters the area.
A safety system has issued a temporary restriction affecting part of the workspace.
Robot A's local perception still sees a navigable path.
Robot B knows about its reservation.
The production system knows that Robot A now has higher priority.
The safety system knows about the restriction.
A wearable or spatial sensor knows that the technician has entered the area.
And the technician needs to understand what the machines around them are going to do.
None of those systems necessarily has to be wrong.
But when they operate independently, without sharing a sufficiently aligned Operational Reality:
LOCAL INTELLIGENCE CAN BECOME GLOBAL CONFUSION.
How predictable is the physical environment in which the system must operate?
BOUNDED → DYNAMIC
Factories are engineered environments, but people, mobile robots, temporary restrictions, maintenance activity, production changes and failures continuously introduce uncertainty.
How strongly do the actions of one participant affect the actions and resources available to others?
HIGH
Robots, workers, production systems and infrastructure increasingly share corridors, workspaces, machines, chargers, lifts, loading areas and production priorities.
How widely is operational authority distributed across machines, systems, people and organisations?
MODERATE — AND INCREASING
A factory may still have a clear organisational owner, but operational authority is already distributed across production, maintenance, safety, robot fleets, software platforms and multiple vendors.
The factory is therefore not merely an automation problem.
It is increasingly a shared-operation problem.
To coordinate this temporary Operational Reality without replacing local autonomy, the operating layer needs six properties.
01 · PRESERVE LOCAL AUTONOMY
Each robot should still be responsible for its own motion and behaviour.
The safety PLC should remain responsible for the deterministic safety functions it owns.
The operating layer should act as a supervisory and coordination layer, not as a central brain.
Its role is to give participants enough shared operational context to make coherent local decisions.
Otherwise, we create latency, fragility and a new single point of failure.
And when connectivity degrades, the factory must not suddenly forget how to operate safely.
LOCAL AUTONOMY REMAINS LOCAL.
COORDINATION HAPPENS THROUGH SHARED CONTEXT, CONSTRAINTS AND INTENT.
02 · ESTABLISH SHARED OPERATIONAL CONTEXT
Each participant sees the situation differently.
Robot A sees a clear route.
Robot B sees a committed reservation.
The production system sees a priority change.
The safety system sees a restricted zone.
The human-tracking system sees the technician entering it.
Sharing raw data is not enough.
The participants need a sufficiently aligned understanding of what is happening now.
This is the role of the Operational Twin:
not another world model replacing the local models of robots, but a living shared operational referent reconciling the relevant state of the physical environment.
Shared operational context must carry more than position and geometry.
OTHERWISE THE FACTORY MAY COORDINATE PERFECTLY AROUND YESTERDAY'S REALITY.
EXPLORE THE OPERATIONAL TWIN03 · MAKE AUTHORITY EXPLICIT, SCOPED AND DYNAMIC
The production system can increase the priority of Robot A.
That does not mean Robot A is allowed to enter a space temporarily restricted because a technician is there.
The shared operating layer needs to represent more than what participants can do.
It must also represent what they are permitted to do now.
Authority needs operational context.
Authority is contextual, not a permanent property of the robot.
The same machine may be authorised to cross a workspace during normal operation, lose that authority while maintenance is taking place, and regain it once the condition has been cleared.
PRIORITY ≠ PERMISSION
04 · COORDINATE COMPETING INTENTIONS
Suppose the technician leaves and the safety restriction is removed.
Robot A and Robot B may now know exactly the same thing about the workspace.
They can still both decide to enter it.
Shared state and action coordination are different problems.
The operating layer therefore needs a way to expose intentions and commitments early enough to resolve conflict.
PERFECTLY RATIONAL LOCAL DECISIONS CAN STILL BE GLOBALLY INCOMPATIBLE.
05 · FEDERATE ACROSS BOUNDARIES
Now make the factory realistic.
Robot A comes from one vendor.
Robot B comes from another.
The safety infrastructure comes from somebody else.
The technician's wearable belongs to another system.
Production priorities come from MES.
Maintenance uses its own software.
Some systems exchange data through an industrial data platform.
Others use their own fleet managers, edge infrastructure or proprietary platforms.
The answer cannot be that one vendor owns the complete stack.
Participants need to federate around the operational scenario while keeping their local autonomy.
And that federation does not have to exist permanently.
Robot A, Robot B, the technician and the affected infrastructure may participate in the same Operational Reality only while their actions are mutually dependent.
Once that scenario ends, they return to their local context.
FEDERATION IS TEMPORARY.
LOCAL AUTONOMY PERSISTS.
EVERY NEW PARTICIPANT SHOULD NOT BECOME ANOTHER CUSTOM INTEGRATION PROJECT.
THAT IS NOT SCALING.
IT IS INTEGRATION DEBT.
06 · ACCUMULATE OPERATIONAL EVIDENCE
Imagine the situation is resolved.
The technician entered the area.
The safety restriction was activated.
Robot A stopped or rerouted.
Robot B released or changed its reservation.
A supervisor may have intervened.
Production eventually resumed.
If all the factory retains is:
“Robot A stopped at 10:42”
it has learned very little.
Operational evidence must preserve the context.
The next version of Robot A — or the next factory deploying the same system — should not need to rediscover the same exception from scratch.
A SMART FACTORY SENSES AND OPTIMISES.
A COGNITIVE FACTORY ALSO ACCUMULATES OPERATIONAL EXPERIENCE.
“Cognitive factory” is CompoundWorks terminology, not an external industry maturity category.
Humans cannot be treated as an exception outside the architecture.
They participate in the same Operational Reality as robots, machines and AI agents, but through different cognition, interfaces, responsibilities and authority.
The human participation layer must connect operational context with the person who needs to understand it and act on it.
The system may need to understand relevant human context such as:
The human needs access to:
A technician does not need the same operational window as a supervisor.
A supervisor does not need the same operational window as maintenance.
Safety personnel may need yet another view.
But they should all operate against the same sufficiently aligned Operational Reality.
ONE OPERATIONAL REALITY.
DIFFERENT WINDOWS.
Local hazard, affected equipment, nearby machine intent.
Production impact, exceptions, priorities and authority.
Asset condition and intervention scope.
Restricted areas, active constraints, authority and validity.
XR and spatial computing can therefore become more than visualisation.
They can become the human-facing spatial interface into shared Operational Reality.
No single technology solves the shared-operation problem.
The value emerges when multiple capabilities are composed into one operational architecture.
THESE TECHNOLOGIES BECOME MORE VALUABLE TOGETHER THAN SEPARATELY BECAUSE THE OPERATIONAL PROBLEM REQUIRES THEM TO COMPOUND.
THE COMPOUNDWORKS VISIONHuman intervention is part of the same evidence loop.
If a technician changes a route, a supervisor overrides a decision or an operator resolves an ambiguity, that intervention must not disappear into a black box called “human in the loop”.
It is operational evidence.
The goal is not simply to make the current deployment work.
Each deployment should make the next deployment easier to understand, validate and qualify.
The vendors of Robot A and Robot B could improve their perception, world models, reasoning and planning enormously and still not solve the shared-operation problem.
Because the bottleneck increasingly sits outside the individual machine.
Physical AI needs more than robot intelligence.
It needs operating infrastructure able to:
A SHARED OPERATIONAL ARCHITECTURE IN WHICH HUMANS, ROBOTS AND AI AGENTS PARTICIPATE THROUGH DIFFERENT INTERFACES, DIFFERENT COGNITION AND DIFFERENT AUTHORITY — BUT OPERATE AGAINST A SUFFICIENTLY ALIGNED OPERATIONAL REALITY.
COORDINATION UNDER COMPLEXITY
Bounded but increasingly dynamic environments.
High operational coupling.
Authority still relatively concentrated, but increasingly distributed across systems and vendors.
FEDERATION UNDER DISTRIBUTED OWNERSHIP
Open environments.
Very high operational coupling.
Multiple organisations, infrastructures, jurisdictions and untrained public participants.
TRUSTED OPERATION UNDER CONTESTED REALITY
Degraded and adversarial environments.
Extreme operational coupling.
Distributed and dynamically changing authority under incomplete or contested information.
THE FACTORY MAY BE ONE OF THE FIRST PLACES WHERE THIS OPERATING ARCHITECTURE CAN BE BUILT AND TESTED.
THE CITY AND THE BATTLEFIELD MAKE THE PROBLEM MUCH HARDER.
Scaling Physical AI is not only a model problem.
It is an architecture, integration, evidence and operational trust problem.