A study from a team at Duke University — Zhehan Qu, Christian Fronk, Jaewoong Jeong, Pavel Manakhov and Maria Gorlatova — tested the central claim made for augmented reality in safety-critical work, and found it half true in a way that matters.
The claim: in tasks like search-and-rescue, where a person works alongside a robot, a head-mounted display with in-situ spatial visualisations keeps your eyes up and on the world, instead of down on a tablet. Better situational awareness follows.
The finding: AR improved awareness of the robot at the perception level only, with no gains in higher-level robot awareness, and no gains in environmental awareness at any level. Eye tracking explains why, and the explanation is the useful part.
The design
A between-subjects study with 30 participants compared a custom AR interface against a custom mobile interface presenting equivalent information — the comparison that makes the result interpretable, since a difference in information content would have explained any difference in awareness without telling you anything about the display.
Awareness was measured with the Situation Awareness Global Assessment Technique (SAGAT) across all three of its levels, for both the robot and the environment:
- Level 1, perception — noticing that an element is there
- Level 2, comprehension — understanding what it means right now
- Level 3, projection — anticipating what it will do next
Measuring all three separately is what the authors note is usually missing from AR human-robot collaboration work, where a single awareness score is reported and the levels are collapsed. Concurrent eye tracking ran throughout, to identify the attentional mechanism behind whatever differences appeared.
The result
Both interfaces achieved high usability — this is not a story about a bad prototype.
Relative to the mobile baseline, AR improved perception-level awareness of the robot. Participants were better at noticing where the robot was and what state it was in. That is a real benefit and it is what conformal visualisation — graphics registered to the physical robot and task — should deliver.
Everything else was flat. No improvement in comprehension or projection for the robot. No improvement in environmental awareness at any of the three levels.
The gaze analysis accounts for it: AR freed attention from the map, but that attention was re-invested in the conformal visuals rather than in the physical environment.
Why this is the finding and not a disappointment
The authors’ own summary is the sentence worth keeping: “Freeing the eyes from a screen is not the same as directing them to the world.”
The argument for AR in this setting has a hidden premise. It assumes that attention spent on a tablet is attention stolen from the environment, and that removing the tablet returns it. But a head-mounted overlay is also a display. If the information it presents is rich and task-relevant — and good AR design works hard to make it so — then attention stays on the display, which now happens to be in front of the world rather than in your hand.
Worse, the failure mode is harder to notice. A colleague staring at a tablet is visibly not looking at the room. A colleague staring through a headset at a dense overlay registered onto the room looks exactly like someone looking at the room.
The study’s phrasing of the design consequence is precise: this is “a distinction AR interfaces for safety-critical HRC must design around.” Not a reason to abandon AR — a specification for it. An interface that genuinely improves environmental awareness has to be designed to push attention off itself, which is close to the opposite of how display interfaces are normally optimised.
For non-safety-critical work
Most readers of this site are not building search-and-rescue interfaces. The transfer is still direct, because the mechanism is about attention rather than about robots.
If you are building a spatial-computing experience whose value depends on people attending to the physical space — a gallery piece with AR annotation, a site-specific tour, a performance with overlaid content — this study is evidence that putting compelling graphics in front of the room competes with the room, and that the competition is not resolved by the graphics being spatially registered. Registration makes the overlay relevant to the space. It does not make looking at the overlay the same as looking at the space.
The practical note is that this is measurable. The authors found it with eye tracking, which is now cheap enough to be standard in headsets. If your work depends on where people look, you can check rather than assume.
Caveats worth stating
Thirty participants, between-subjects, one task domain, one AR interface design. The result shows that this interface failed to redirect attention to the environment, not that no interface could. The authors frame it as a design constraint, and that is the right reading.
The paper is eleven pages with nine figures. Qu and Fronk are noted as contributing equally; this is version 2, revised on 13 October 2026.
Sources
- Better Situational Awareness in AR-HRC? A Comparative Study of Augmented Reality and Mobile Interfaces for Human-Robot Collaboration — arXiv:2609.01461
- DOI: 10.48550/arXiv.2609.01461
- Hart & Staveland, Development of NASA-TLX — the standard workload instrument alongside SAGAT
- Maria Gorlatova — Duke University