The level of aggregation should match the supervisory task's grain, not stay fixed

What it is

Which level of aggregation to show — individual robot, subgroup, or whole fleet — needs to match the grain the current supervisory task actually calls for, rather than staying fixed across every phase of a task. Switching levels this way is semantic zoom: it changes what kind of question the interface can answer, not just how large the graphics render.

Why it happens

This builds on the earlier aggregation principles and points to a need for dynamic adjustment: different supervisory decisions require different statistical units — resource balancing calls for fleet-wide distribution, a communication partition calls for network component structure, collision risk calls for a local neighborhood, and hands-on maintenance calls for a single robot's component state. The level of detail a task needs also shifts across its own phases — planning needs only an overall progress overview, execution monitoring needs to drill into a subgroup or individual once a region shows an anomaly, and a post-hoc review may need to zoom back out to look at trends. A fixed aggregation level cannot track these phase shifts. Drill-down solves the question of whether finer-grained information is reachable at all; this principle is about whether the interface's default aggregation level itself can move, rather than requiring the same number of drill-down steps every time just to reach the grain the current task needs. If a level switch also reshuffles object ordering and the timeline at the same time, it breaks the spatial memory and causal thread the supervisor had already built up — the cost most easily overlooked when designing this kind of switching.

Studying it

Testing dynamic aggregation-level switching involves giving supervisors a task sequence spanning distinct phases — planning, execution monitoring, anomaly response, and review — and comparing a fixed single level against a task-adaptive level on cross-scale question accuracy, diagnostic efficiency, navigation cost (the extra actions spent switching), and mode-error rate across the whole sequence, while controlling for whether the system's automatic level recommendation is actually correct — a wrong automatic switch can confuse more than no switching at all, while leaving zoom entirely to the user can leave them stuck at the wrong grain for too long.

Where it stops holding

A critical fleet-wide event needs to break through whatever level a supervisor is currently at and surface directly, rather than being buried because the supervisor happens to be looking at single-robot detail; some tasks also cross a tree-shaped hierarchy rather than following a clean nested global-region-robot structure. Semantic zoom cannot make up for missing underlying data — if a given level was never instrumented with enough granularity, switching the view produces no new information — and the aggregation rule at every level has to remain inspectable, since a supervisor who cannot tell how a number at some level was computed will trust dynamic switching less, not more. This principle matters most when the supervisory task genuinely spans phases with very different grain requirements; for a single, stable task with unchanging aggregation needs, adding dynamic switching may just be unnecessary complexity.

Applying it

Define a default aggregation level and its key metrics for each class of supervisory question — fleet distribution by default for resource-balancing problems, local neighborhood by default for collision-risk problems — and preserve object identity, the timeline, current selection, and color coding continuously across levels, so the same object does not look disconnected from itself when the level changes. Preview any automatic level change before it happens, and let the supervisor pin a level manually so an unexpected jump does not interrupt an ongoing judgment. Verify it with a sequence of scenarios — a fleet-wide resource crunch, a regional task partition, and a single-robot fault, occurring one after another — testing whether supervisors can move naturally across grains as the task develops and recover their judgment quickly, rather than having to reorient from scratch on every switch.

Related

• Same group: X7.07.1 Large numbers of robots need an aggregated overview, not a full list · X7.07.2 Aggregated views need a path down to single-robot detail · X7.07.3 Outlier robots must stand out in an aggregated view, not get averaged away
• Nearby: X7.04 Multi-robot supervision · X3.01 Visible Current State
• Search terms: semantic zoom · level of detail · human-swarm interaction