Supervising more robots at once thins the attention any single one receives
Aliases: attention limits in multi-robot supervision · human-swarm interaction
What it is
Supervising several robots at once means attention has to be time-sliced across them, and the sustained attention any single robot receives falls as fleet size grows. Supervisory control theory gives this a quantitative handle through the neglect tolerance / fan-out model: each robot has a neglect tolerance — the time it can go unattended before its performance drops below an acceptable level — and each human-robot exchange costs an interaction time. The theoretical fan-out, the largest fleet one operator can sustain, works out to roughly neglect tolerance divided by interaction time.
Why it happens
The value of the model is not the number it produces but the way it splits fan-out into two independently improvable quantities. Neglect tolerance tracks how reliable the robot's autonomy is — the less it needs bailing out, the longer it can go unattended. Interaction time tracks interface efficiency — the same status check or command can take seconds with a good display or tens of seconds with a poor one. That means fan-out can be raised two separate ways: investing in more capable autonomy, or shortening each interaction through better interface design, rather than simply asking the operator to concentrate harder or train longer. Both quantities also drift across a mission: a robot cruising in open terrain tolerates long neglect, but once it enters fine manipulation or unmapped territory it needs attention far more often, neglect tolerance shrinks, and the theoretical fan-out drops with it — it is not a constant that belongs in a fixed staffing table. Interaction time itself decomposes further into perceiving current state, forming a judgment, and issuing the command; an interface that lets an operator batch several commands into one interaction — queuing a sequence of waypoints instead of confirming each one on arrival — compresses interaction time on its own, without touching autonomy at all. That makes interface batching a third lever alongside autonomy investment, not a substitute for either.
Studying it
Typical studies have operators supervise varying fleet sizes under controlled tasks while autonomy level and interface design are manipulated, measuring overall mission completion, the performance drop each robot suffers relative to being supervised alone, and subjective workload. Observed performance breakpoints are then used to fit neglect tolerance and interaction time and check whether they reproduce the predicted fan-out. Some work also logs gaze-switching frequency and response latency across fleet sizes as a cross-check on the subjective workload scale.
Where it stops holding
Neglect tolerance and interaction time are not fixed parameters — the theoretical fan-out only holds within a window where mission phase, autonomy level, and interface design stay constant, and the same fleet can need a fresh estimate once it enters a new phase. Individual differences among operators — multitasking ability, familiarity with the platform, current fatigue — can also push the sustainable fleet size well away from the theoretical value, so that value is a planning ceiling, not a staffing mandate. Neglect tolerance also depends on how "unacceptable performance" is defined for the task: drifting off a nominal path is often tolerable in navigation but the same lapse can end a fine-manipulation task outright, so the same neglect-tolerance number carries different real risk across task types, and importing a value measured on one task type into another over- or under-estimates true capacity.
Applying it
Measure the robots' actual neglect tolerance and typical interaction time for the target task before setting fleet size, rather than picking an operator-to-robot ratio by convention. Validate by running representative operators through the intended fleet size in a real task or high-fidelity simulation, watching for performance drops or error-rate increases caused by delayed attention; if they appear, either shrink the fleet or prioritize the robots' autonomous recovery capability rather than expecting the operator to compensate through effort alone.