In the public imagination, the rise of physical AI is a story about perception and autonomy. Humanoid machines acquire better dexterity, buildings become responsive environments, road fleets turn into software-defined networks, and domestic devices quietly negotiate with one another in the background. Yet by mid-2026, the harder question is not whether these systems can act intelligently, but whether they can remain dependable once exposed to dust, weather, vibration, misuse, budget cuts and long service lives. The hidden bottleneck is not cognition but continuity.
That is an awkward conclusion for a sector drawn to model performance and pilot deployments. Maintenance does not photograph well. It is procedural, repetitive and expensive. But physical intelligence does not fail gracefully when maintenance is treated as optional. A misaligned camera in a warehouse robot, a stale certificate in a building controller, an unpatched telematics unit in a fleet vehicle or a degraded battery in an autonomous cleaning machine can each turn an apparently smart system into a hazard, a source of waste or simply a stranded asset.
From breakthrough narratives to service realities
Software culture still shapes much of the conversation around AI in the physical world. Bugs can be fixed later; features can be shipped iteratively; data can compensate for design limits. That logic weakens when systems have brakes, actuators, doors, grippers, medical functions or public-facing mobility roles. NIST’s work on cyber-physical systems has long stressed that computation, networking and physical processes are inseparable in such environments. Once a system touches the real world, downtime acquires safety consequences and routine servicing becomes part of governance, not merely operations.
In practice, the decisive variable is often not headline intelligence but the ability to preserve acceptable performance across months and years. Sensors drift. Mechanical tolerances change. Adhesives age. Connectivity contracts expire. Suppliers disappear. Sites are remodelled. Human operators develop workarounds that no design team anticipated. The resulting mismatch between laboratory capability and field reliability explains why many apparently promising deployments stall after the pilot phase.
The maintenance gap in smart spaces
Smart buildings offer the clearest example. Modern spaces are full of connected locks, cameras, occupancy sensors, heating controls, air-quality monitors and energy-management systems. The ambition is coherent orchestration. The reality is often a patchwork of protocols, installers and maintenance contracts layered onto legacy infrastructure. A sensor network is only as trustworthy as its servicing regime.
When a building system degrades, the consequences are usually interpreted as inconvenience rather than as intelligence failure. Rooms are overcooled, badge readers behave erratically, occupancy counts drift, lifts report phantom faults, emergency systems produce false positives. Yet the aggregate effect matters. If data quality decays, every optimisation layer above it becomes suspect. This is why the European Union’s recent cyber and AI frameworks matter less as abstract legal texts than as pressure towards lifecycle discipline: asset inventories, software update obligations, logging, incident handling and clearer accountability over who must keep connected products safe after installation.
A sensor network is only as trustworthy as its servicing regime.
Physical intelligence does not fail gracefully when maintenance is treated as optional.
Humanoid robots are maintenance-intensive by design
Humanoid and general-purpose service robots are frequently discussed as labour substitutes. A better framing is that they are maintenance-intensive co-workers. Their very promise rests on operating in environments designed for humans rather than machines: offices, hospitals, shops, homes and transport hubs. Those settings are dynamic, cluttered and socially dense. To function there, robots need high sensor fidelity, reliable actuation, robust battery performance, secure communications and regular recalibration.
Unlike industrial robots in caged environments, human-facing machines encounter dirt on floors, reflective surfaces, improvised obstacles, accidental impacts and changing lighting conditions. Their failures are therefore more ambiguous. A human might interpret a pause as hesitation, a navigation error as rudeness, a grasp failure as incompetence, or a software lock-up as danger. That reputational fragility raises the importance of preventive service. In health and care settings, WHO’s guidance on AI governance has underscored the need for human oversight, safety and accountability. Those requirements do not begin and end with model validation; they extend into the banal but essential question of who replaces worn components and verifies that updates have not altered behaviour in risky ways.
Vehicles show what lifecycle governance really means
Connected vehicles already reveal the shape of this future. The automotive sector has spent years moving from mechanical maintenance schedules to software-centred lifecycle management. Cybersecurity engineering standards such as ISO/SAE 21434 recognise that road vehicles are not static products but evolving systems exposed to persistent threats and update demands. The same is increasingly true for delivery robots, municipal machines, shuttles and construction equipment.
Vehicle networks make visible a broader truth: maintenance now includes cryptographic keys, anomaly detection, patch provenance and secure decommissioning. A brake actuator may function perfectly while the update channel around it is compromised. A fleet can be physically healthy but operationally vulnerable if supplier support ends prematurely or if diagnostics are opaque to independent repairers. This is one reason the politics of repair, interoperability and software support periods are becoming central to physical AI. The machine is no longer separable from the service stack that keeps it lawful and safe.
Why predictive maintenance is not enough
For years, predictive maintenance was presented as the elegant answer: instrument assets, model degradation, intervene just in time. There is value in that approach, and OECD work on industrial maintenance has highlighted the productivity gains from more data-driven practices. But prediction addresses only one layer of the problem. It assumes the organisation has spare parts, competent technicians, secure remote access, accurate configuration records and the authority to act on early warnings.
In many smart-space and robotics deployments, those preconditions are weak. Operators may lease devices they cannot inspect deeply. Integrators may control software without responsibility for hardware wear. Facilities teams may own uptime targets but lack access to vendor logs. Procurement may prioritise low upfront cost over supportability. Predictive tools in such settings become dashboards of deferred responsibility. They detect decay without resolving the institutional fragmentation that produced it.
The hidden bottleneck is not cognition but continuity.
The new politics of repairability
Repairability used to be discussed mainly in relation to consumer devices. In the phygital economy, it becomes a strategic issue for cities, hospitals, logistics operators and public infrastructure managers. If a robot, gateway or building subsystem cannot be repaired except by one remote vendor, resilience is weaker than the glossy system diagram suggests. If parts are proprietary, interfaces undocumented and software locks pervasive, maintenance becomes a rent rather than a capability.
WIPO’s work on innovation and diffusion is relevant here because physical AI is not merely an invention challenge; it is a capability-distribution challenge. Societies that can operate, adapt and repair complex systems will capture more value than those that only import finished equipment. The practical test is whether local technicians, accredited workshops and public operators can maintain deployed intelligence without breaching safety or cybersecurity requirements. This is not an argument for unfettered tinkering. It is an argument that repair rights, secure documentation and lifecycle transparency are now part of industrial policy.
Physical intelligence does not fail gracefully when maintenance is treated as optional.
Cybersecurity and maintenance have merged
One reason the maintenance debate has sharpened is that software support is no longer separable from safety support. The EU Cyber Resilience Act and related connected-device rules reflect a simple premise: products with digital elements must remain secure over their expected lifetimes. For physical AI, this folds cybersecurity into ordinary maintenance schedules. Patching windows must account for operational downtime. Authentication failures can disable equipment. Vulnerability disclosure becomes part of service management. End-of-support dates become operational risk markers.
This convergence changes job descriptions as much as regulation. The classic maintenance engineer who could diagnose wear through sound, vibration or heat now needs to understand logs, certificates and dependency chains. Conversely, software teams must grasp failure modes in motors, power electronics and embedded controllers. The professions are not merging entirely, but the wall between them is thinner than most organisations admit.
Labour, not just hardware, is the scarce input
There is a tendency to speak as if autonomy will reduce reliance on skilled labour. In reality, physical AI often shifts demand towards different kinds of skill: field technicians able to swap modules safely, cyber-physical engineers able to validate updates, facilities teams able to interpret noisy diagnostics, and supervisors able to distinguish user error from system degradation. The maintenance burden does not vanish; it becomes more specialised.
A sensor network is only as trustworthy as its servicing regime.
This has two implications. First, labour shortages in servicing professions may constrain deployment more than model quality does. Secondly, institutions that neglect training and certification will struggle to use complex systems responsibly. A hospital may procure mobile assistants or sensor-rich wards, but without a dependable servicing workforce it is merely accumulating points of failure. The same applies to municipalities trying to run connected transport assets or smart public buildings on thin budgets.
Insurance, liability and the economics of uptime
Maintenance is also where the economics become plain. Insurers and regulators care less about promotional claims than about incident frequency, audit trails and recoverability. If a connected machine causes harm, investigators will ask predictable questions: Was it serviced on schedule. Were updates tested. Were faults logged. Was a vulnerability known but unpatched. Could an operator override unsafe behaviour. These are maintenance questions in legal dress.
As the EU AI Act takes effect, risk management duties for certain systems may increase documentation burdens across supply chains. That will not end experimentation, but it will favour deployments that can evidence lifecycle control. Organisations able to show disciplined maintenance, versioning and human oversight will find it easier to defend uptime claims and, potentially, liability positions. The glamour lies in autonomy; the margin may lie in record-keeping.
What a mature phygital stack will look like
A more mature physical-AI environment is unlikely to be defined by the most humanlike robot or the most densely instrumented building. It will be defined by boring virtues. Components will expose health states in standardised ways. Update channels will be secure and auditable. Assets will have clear software-support horizons. Spare-part availability will be planned rather than improvised. Maintenance records will be portable across contractors. Decommissioning will include data sanitisation, not just disposal.
Just as importantly, intelligent gateways and supervisory agents will matter less as magic orchestrators than as disciplined intermediaries: verifying device identity, limiting privileges, preserving logs, mediating updates and helping users understand when a machine is drifting out of tolerance. In that sense, the secure personal or institutional gateway is valuable not because it centralises control for its own sake, but because it can make lifecycle accountability legible in environments where too many devices otherwise operate as black boxes.
The real test of physical intelligence
The next phase of the phygital economy will not be decided by demos alone. It will be decided in service corridors, loading bays, plant rooms, maintenance depots and municipal procurement offices. There, organisations will discover whether their robots can be recalibrated without vendor drama, whether their smart spaces can survive turnover in contractors, whether vehicle fleets can remain secure through years of updates, and whether logs are sufficient to reconstruct failures when something goes wrong.
That is less romantic than the usual future-of-automation script, but more consequential. A society that cannot maintain its physical intelligence will not, in any meaningful sense, possess it. It will merely rent episodes of functionality between outages. The strategic task in 2026 is therefore not only to build systems that can sense and act, but to build institutions, standards and labour markets capable of keeping those systems trustworthy over time.
