cs.AIMar 16, 2026
SaveResilience Meets Autonomy: Governing Embodied AI in Critical Infrastructure
Organizations: Department of Automation and Process Engineering (IAP), UiT The Arctic University of Norway, Tromsø, Norway · Department of Technology and Safety (ITS), UiT The Arctic University of Norway, Tromsø, Norway
Abstract
Critical infrastructure increasingly incorporates embodied AI for monitoring, predictive maintenance, and decision support. However, AI systems designed to handle statistically representable uncertainty struggle with cascading failures and crisis dynamics that exceed their training assumptions. This paper argues that Embodied AIs resilience depends on bounded autonomy within a hybrid governance architecture. We outline four oversight modes and map them to critical infrastructure sectors based on task complexity, risk level, and consequence severity. Drawing on the EU AI Act, ISO safety standards, and crisis management research, we argue that effective governance requires a structured allocation of machine capability and human judgement.
Figures & tables
| Oversight mode | Human role | Key characteristics | Main selection criteria |
|---|---|---|---|
| Fully AI-Automated (Human-out-of-the-Loop) | None (Autonomous) | High operational autonomy; rapid response and scalability; operates without routine human intervention, but within predefined safety constraints and fallback mechanisms. | Routine operation of infrastructure and instantaneous physical stabilisation. Used, e.g., for millisecond-level load balancing in smart grids or preventing cascading electrical failures where humans are physically too slow to act, provided that functional safety is assured by system design and safeguards. |
| Human-on-the-Loop (HOTL) | Supervisor (Passive) | The system operates autonomously under human supervision; the human can monitor, interrupt, override, or reclaim control if necessary. | Predictive maintenance and steady-state operations. Used when the system is stable, but where a human must be able to override the AI and assume control if it misinterprets a physical anomaly or system state. |
| Human-in-the-Loop (HITL) | Gatekeeper (Active) | Humans are a mandatory part of the decision chain; the system cannot proceed without approval in high-impact or safety-critical actions. | Service restoration and reconfiguration. Used for high-consequence decisions, where control transfer to the human must occur before execution because errors may cause major physical, social, or environmental harm. |
| Human-in-Command (HIC) | Policy-maker (Strategic) | Humans define goals, safety limits, rules of engagement, and escalation thresholds; AI operates only within these externally set constraints. | Crisis management, especially in unexpected events, and disaster recovery. Used during large-scale infrastructure failures or cyber-warfare scenarios where strategic trade-offs, exceptional uncertainty, or crisis escalation require political, legal, and ethical accountability beyond automated decision-making. |
TABLE I: AI autonomy and the oversight modes in critical infrastructure
| CI domain | Typical embodied AI application | Main resilience contribution | Dominant vulnerability | Most appropriate oversight modes |
|---|---|---|---|---|
| Energy infrastructure | UAVs, USVs, and mobile robots for inspection of transmission towers [ 31 ] , pipelines, offshore wind farms [ 33 ] , and subsea assets [ 22 ] | Continuous monitoring, early fault detection, reduced human exposure, faster maintenance response | Misinterpreted anomalies, cyber-physical manipulation, and cascading failure risk | HOTL / HITL; HIC in large-scale outages or crisis escalation. |
| Transport systems | Autonomous vehicles [ 6 ] , drones, and maritime autonomous surface ships [ 34 ] | Real-time navigation, operational continuity, and safety optimisation | Edge cases, unpredictable environments, accident response, and contested responsibility in emergencies | HOTL in routine supervision; HITL / HIC in emergencies and high-consequence decisions. |
| Water, wastewater, and digital infrastructure | Crawling or climbing robots, UAVs, and AUVs for inspection of pipes, tunnels, dams, towers, undersea cables, and offshore structures [ 29 ] | Inspection in hazardous or inaccessible environments, structural health monitoring, and continuous surveillance | Sensor degradation, poor visibility, communication limits, hidden defects, and sabotage or tampering | HOTL in routine inspection; HITL where intervention decisions affect safety or service continuity. |
| Health and hospital logistics | Delivery robots [ 15 ] , monitoring systems, perimeter security [ 30 ] , and robot-assisted surgery [ 18 ] | Efficiency, reduced routine burden, improved logistics, and potential clinical precision | Life-critical error consequences, explainability and trust problems, cognitive overload, and accountability | HITL as the default; HIC for strategic and ethically sensitive decisions |
| Food and agricultural systems | Autonomous agricultural robots for spraying, weeding, planting, harvesting, and field monitoring [ 38 ] | Precision, efficiency, resource optimisation, and continuity of food production | Unstructured environments, mixed vulnerabilities, and variable safety and reliability under changing field conditions | HOTL in routine operations; HITL where failure may affect safety, environment, or supply continuity |
| Space and other harsh environments | Orbital, planetary, and subsea robots for inspection, maintenance, retrieval, docking, and exploration [ 47 ] | Enables operation where direct human action is impossible or highly constrained | Extreme uncertainty, communication delays, environmental unpredictability, and limited recoverability after failure | Fully automated / HOTL in routine operation; HIC for mission-level strategic choices |
TABLE II: Embodied AI in critical infrastructure: operational promise, dominant vulnerabilities, and most appropriate oversight modes
More in this topic
AI Agent Governance
AI agent governance studies how to authorize, constrain, monitor, and hold autonomous and multi-agent systems accountable.
See all 442 papers