Skip to main content
Back

UiT Researchers Build a Governance Blueprint for Robots Running Critical Infrastructure

A new framework maps four human oversight modes onto embodied AI systems in critical infrastructure, tying autonomy limits to task risk and crisis dynamics.

Paper, Full Access
  • New framework argues embodied AI in critical infrastructure needs bounded autonomy, not full automation or simple overrides.
  • Defines four oversight modes: Fully Automated, Human-on-the-Loop, Human-in-the-Loop, and Human-in-Command.
  • Maps modes to sectors including energy, transport, water, health, food, and space robotics.
  • Grounded in the EU AI Act, ISO/IEC TR 5469:2024, and ISO/IEC TS 8200:2024 controllability standards.
  • Decomposes robot failure into exogenous, endogenous, and mixed vulnerabilities, including SLAM drift and adversarial data poisoning.
  • Position paper only: no code, dataset, or benchmark, and no empirical evaluation of the framework.

Autonomous drones inspect power lines. Crawling robots crawl through pipelines. Surgical robots assist in operating rooms. As embodied AI seeps into the physical systems that societies depend on, a quieter question keeps surfacing: when the model hits something it was never trained for, who is actually in charge? A new paper from researchers at UiT The Arctic University of Norway argues that the answer cannot be full automation or a simple human override button, and it lays out a governance architecture to prove the point.

The paper, Resilience Meets Autonomy: Governing Embodied AI in Critical Infrastructure, is a position and framework contribution rather than a model release. There is no repository, benchmark, or code artifact. What it offers instead is a structured taxonomy that maps four oversight modes onto critical infrastructure sectors, grounded in the EU AI Act, ISO safety standards, and crisis management research.

The blind spot in autonomous infrastructure

AI is typically designed to handle normal variation through training data and probabilistic models, yet this reliance on historical data can introduce bias and limit its ability to generalise to novel situations. The authors frame this as the problem of systemic surprise: cascading failures, unknown unknowns, and crisis dynamics that lie outside any training distribution. In those moments, autonomous systems tend to do one of two bad things.

  • Errors of omission - failing to act when action is required.
  • Errors of commission - acting inappropriately under operational pressure.

When confronted with genuinely novel conditions for which no adequate rules exist, behaviour may become unpredictable: systems may fail to act, revert to brittle routines, repeat earlier responses, or behave erratically. That fragility is amplified by tight coupling. Because infrastructures are interdependent, with electricity and ICT often acting as escalation hubs, an initial AI failure can cascade far beyond its point of origin.