How often?
How many systems actually need a person to step in?
Remote WarriorResource PortalREMOTE WARRIOR / THE HUMAN SUPERVISION PROBLEM
How many balls can a human juggle?
Now imagine the balls can juggle themselves—until they can't.
That's the challenge of supervising autonomous machines: most can work independently, but what happens when several suddenly need a human at the same time?
Try it for yourselfTry the thought experiment
What happens when the balls stop catching themselves?
Each ball represents one autonomous system. Change the number in the air—then introduce a problem.
Systems in the air
What changes?
Autonomous systems10 in the air
Self-managing (10) Need the human (0)
One human supervisorNo catches needed right now
All systems are managing themselves. The person is still responsible for recognizing when that changes.
Visual analogy only. Five simultaneous alerts are not automatically five equal tasks; severity, available time, and automated contingency behavior determine the real burden.
THE POINT
It's not about how many balls are in the air.
It's about how many need catching, how quickly, and what happens when the person can't catch them.
Three things matter more than the number of machines.
How soon?
Does the person have enough time to understand and act?
What if they can't?
Does the system reach a safe, authorized condition on its own?
THE SHORT ANSWER
Can one person supervise 5, 10, or 50 systems?
Sometimes—but there is no universal safe number. A person might oversee fifty self-managing systems under ordinary conditions and struggle with three urgent, conflicting problems.
Monitoring something, making decisions about it, and recovering it from failure are different responsibilities. See what the research actually supports ↓
INTERACTIVE MATHHow much human intervention time would this require?Open the calculator and change the assumptions.+
Interactive educational model
Human intervention capacity
Change assumptions and observe how nominal human-required work grows. This is not a safety assessment.
1100
0.012.00
0.5 min30 min
Additional workload assumptions (optional)
These change the capacity accounting, not the calculated intervention demand. Monitoring work is an independently assumed average; adding operators assumes perfect work distribution.
Excluded: correlated failures, queueing, interruptions, fatigue, human reorientation, and intervention deadlines. More operators may add coordination overhead not represented here.
Selected scenario
16.7%Nominal intervention utilization for one operator
2.5Expected interventions / hr
10Expected minutes / hr
Nominal intervention demand10 / 60 min
Compare aircraft counts
5 aircraft1.25 interventions/hr · 5 min/hr · 8.3%
10 aircraft2.5 interventions/hr · 10 min/hr · 16.7%
50 aircraft12.5 interventions/hr · 50 min/hr · 83.3%
How this works
Aircraft × interventions per aircraft per hour × minutes per intervention = expected human intervention minutes per hour. Divide by 60 to express single-operator nominal utilization.
Important limitation: a result below 100% does not mean an operation is safe or viable. Simultaneous requests, finite deadlines, skill, situational awareness, fatigue, communications, and contingency design can make an intervention impossible even at a low average workload.
Educational use only. This simulator does not establish a safe operator-to-system ratio, certify human performance, or substitute for human-in-the-loop testing, safety assessment, or regulatory authorization.
A better interface shouldn't demand faster juggling.
It should keep routine tasks with the machines, explain when human judgment is needed, and provide enough context, time and authority to intervene. When that isn't possible, a verified contingency must protect the operation.
Less unnecessary catching. More meaningful human control.
THE EVIDENCEWhat does research actually show?Studies, original sources, limits and important qualifications.+
Evidence, not a magic aircraft-to-human ratio.
These studies address different tasks and environments. None establishes a universally safe number of aircraft for one supervisor.
SOURCE 01
Bogg & Birrell, 2025 — Monitoring automated vehicles ↗
In a study of 24 participants monitoring 3–9 simulated road vehicles, performance was generally strongest at 5–7. Participants delegated actual vehicle intervention to another operator.
Important limit: This examined monitoring road vehicles, not one person controlling 5–7 aircraft. It establishes no universal staffing ratio.
SOURCE 02
Cummings & Mitchell, 2008 — Supervisory capacity and delay ↗
Adding waiting, reorientation and loss-of-situation-awareness delays reduced predicted supervisory capacity by up to 67% in the paper's modeled conditions.
Important limit: A model-based result with specific assumptions—not a universal percentage reduction for all UAS operations.
SOURCE 03
NASA-STD-3001, Vol. 2 — Human–automation interfaces ↗
NASA's crew-interface requirements address system state, predicted state, health, configuration, responsibility and safe automation intervention.
Important limit: Human-spaceflight requirements are useful design references, not automatic UAS regulatory requirements.
SOURCE 04
FAA Part 107 — Multiple small UAS ↗
The FAA explains the waiver path for activities that depart from specified Part 107 operating rules, including applicable multiple-UAS operations.
Important limit: Technical feasibility and a research simulator do not grant regulatory authorization.
Why simultaneous failures change the answer
A calm hour may involve very little human work. Then shared communications, weather, or a common software issue can generate several overlapping requests. The human must prioritize, reorient, decide, and act within deadlines. Average workload alone cannot tell us whether those actions are possible.
As a fictional illustration, imagine four aircraft needing decisions within 25, 40, 60 and 90 seconds. Those figures are not real operating limits; they simply show how multiple requests can collide.
What happens when autonomy changes?
Remote Warrior's proposed evidence-continuity question asks whether prior conclusions about human supervision remain valid after changes in software, sensors, missions, operating limits, or responsibilities. A defensible assessment should preserve evidence of configuration, operating envelope, qualifications, change impact, reassessment, and decision authority.
This evidence-continuity framework is a research concept requiring testing—not a validated industry standard or new legal requirement.
What the calculator cannot prove
Average intervention time does not capture correlated emergencies, queuing, fatigue, cognitive reorientation, decision correctness or deadline failures. Even a workload below 100% does not establish safety. The calculator is an illustrative teaching tool, not a certification method.
For the underlying US civil rule, see 14 CFR §107.35 ↗.
GO DEEPERRead the discussion that led hereThe brain, a proper interface, and the 5/10/50 supervision question.+
The archive is an edited public reading edition rather than a word-for-word transcript. The complete verbatim export remains a separate review requirement.
Edited public reading edition. This preserves the three questions and their substantive arguments in an accessible condensed rendering, not a word-for-word transcript of all platform-generated content. A full verbatim export remains a publication acceptance item.
1. What's the human brain capable of?
The capabilities
The human brain learns, imagines, predicts, invents, adapts, experiences consciousness, and physically changes with experience. Neuroplasticity modifies neural connections and activity patterns. Mental simulation lets us rehearse possible futures, and expertise makes meaningful patterns noticeable that novices miss. Language, mathematics, collaboration and tools extend our individual capability.
Scale and limits
The brain has roughly 86 billion neurons and trillions of connections and draws a substantial share of the body's energy. Its conscious workspace is comparatively narrow. Working-memory experiments often find only a handful of simultaneously accessible chunks. The popular claim that people use only ten percent of the brain is false.
Outstanding questions
We have no complete explanation of consciousness, general flexible reasoning, memory consolidation or the limits of adult neuroplasticity. Brain–computer interfaces demonstrate increasingly consequential communication and control capabilities for some people with paralysis, but do not read arbitrary thoughts.
Human performance is not biological potential
The brain can struggle with sustained vigilance, task switching, false memories, biases and attention to multiple independent events. Capacity is what biology makes possible; performance is what someone reliably accomplishes under realistic conditions; potential also depends on learning and supporting environments. Changing the representation of information can improve performance without changing inherent intelligence.
The implication
Extending human capability may increasingly mean reducing the distance between minds, information and tools—without creating additional cognitive burdens.
2. What does a proper human–machine interface look like?
The purpose
A properly designed interface helps a person work with a system that presents what matters for the human's responsibilities, including its own limitations. Five autonomous aircraft should not automatically require five unrelated consoles. Routine automation and exceptional supervision should be distinguishable; essential raw evidence and controls must still be available.
Seven questions
What is happening now? What does the machine intend? What happens next? What information is missing or unreliable? What is authorized? What decisions can the human make and how much time is available? On what evidence and authority was the action based?
Normal, exceptional and emergency states
Normal operations need predictable low-clutter summaries that maintain operator awareness. Emerging problems need changed conditions, cause, time and alternatives. Emergencies require immediate recovery information and verified safe procedures, not multiple menus. Unpredictable display changes may themselves create confusion.
A human and a machine are different resources
Humans can interpret ambiguity, negotiate competing values and draw on experience. Machines can continuously calculate, record, compare and monitor. The correct allocation of functions depends on demonstrated capabilities and clearly specified authority, not treating the pair as equal decision-makers.
The continuity problem
After an autonomy software update, can the earlier human qualification still be defended? A complete supervisory system should make available technical configuration, approved envelope, the human's qualified tasks, change impact, required reassessment and decision authority. This goes beyond dashboard cosmetics and is a proposed evidence-continuity concept rather than a universal current HMI requirement.
How to validate the design
Human-in-the-loop tests should evaluate recognition, situation awareness, decision quality, response time, workload, recovery and sustained performance. Experienced and inexperienced personnel should face quiet periods, misleading automation, overlapping failures, handoffs and configuration changes. Visual polish does not establish safe performance.
The larger question
Can one person supervise five, ten or fifty autonomous systems without receiving more cognitive responsibility than they can reliably fulfill?
3. Can a person genuinely supervise five, ten, or fifty autonomous systems without the interface transferring more cognitive responsibility than they can reliably handle?
Conditional answer
Yes, under bounded conditions. Someone may watch fifty predictable, largely self-managing systems but struggle with three ambiguous simultaneous emergencies. The key variable is consequential decisions needing recognition, comprehension and resolution before deadlines—not the displayed count of aircraft.
Existing evidence
Researchers have investigated multi-vehicle supervisory capacity for many years. Cummings and Mitchell showed that reorientation and situation-awareness costs can lower capacity predictions. Bogg and Birrell examined simulated ground-vehicle monitoring, not control of multiple UAS, and found monitoring performance strongest generally at five to seven vehicles in their experimental conditions. These findings cannot be merged into a universal safe operating ratio.
A four-aircraft exception queue
One fictional scenario involves aircraft 07 facing a route conflict with 25 seconds remaining, 19 in a communications contingency with 40 seconds, 32 experiencing conflicting sensor inputs with 60 seconds and 44 approaching a mission boundary with 90 seconds. These are illustrative—not real observations. The human cannot necessarily investigate and resolve four such exceptions simultaneously. Prioritization does not manufacture attention or time.
First-order workload model
Expected intervention minutes per hour = number of aircraft × mean interventions per aircraft per hour × mean minutes per intervention. Nominal single-operator intervention utilization = those minutes ÷ 60 × 100%. At 0.25 interventions per hour of 4 minutes each: 5 aircraft imply 5 minutes/hour (8.3%); 10 imply 10 minutes/hour (16.7%); 50 imply 50 minutes/hour (83.3%). The interactive simulator above exposes these assumptions.
Why averages mislead
Rare common-cause failures (weather, shared communications, software defects) can cause clustered intervention requests. A low average workload does not guarantee enough time to meet a sudden deadline. The simple model does not account for task switching, queueing, fatigue, coordination or the quality of operator decisions.
Meaningful testing
A reasonable research protocol compares realistic missions at counts of 1, 3, 5, 10, 20 and 50. Evaluate event detection, comprehension, correct and timely action, safe recovery when intervention is impossible, and sustainable workload under prolonged operations. Include misleading automation, handoffs, simultaneous failures and changes in autonomy.
Qualitative distinctions
Five systems may be manageable for monitoring but not five simultaneous time-critical actions. Ten can be plausible with mature bounded automation and demonstrated intervention capacity. Fifty may allow mission-level oversight but needs particularly strong evidence when an individual is represented as the meaningful emergency decision-maker. These are interpretations, not certified limits.
Evidence after change
A supervisor's earlier qualification and operating authorization may be undermined by changed software, sensors, behavior or responsibility. The institutional question is who must recognize a changed assumption and act before continuing operations.
Core conclusion
If a system requires a person to exercise judgment but does not provide sufficient information, time, capability and authority to do so, it has assigned responsibility without ensuring the capacity to fulfill it.
Increasing the number of autonomous systems is not proof that human supervision has scaled with them.
Explore more Remote Warrior research →