Effects of Swarm Size Variability on Operator Workload

1. Problem Statement

Real-world deployments of human-swarm teams (HST) require operators to monitor and manage dynamic robot swarms. A critical challenge in Human-Swarm Interaction (HSI) is that swarm sizes are rarely static; robots may join or leave due to failures, redeployment, or changing mission requirements. These abrupt fluctuations create unpredictable shifts in operator cognitive demand.

While existing research often assumes fixed swarm sizes, real-world scenarios involve dynamic transitions. The core problem addressed is understanding how the magnitude (size of the change) and direction (increase vs. decrease) of swarm size changes affect operator workload and performance. Specifically, the authors investigate whether "workload history" (hysteresis) persists after a reduction in swarm size, whether gradual increases are easier to manage than abrupt jumps, and whether large changes induce a cognitive "reset" that overrides these history effects.

2. Methodology

The authors conducted two controlled studies ($N=34$) using a simulated drone monitoring task within the HARIS simulator.

• Task Design: Participants watched a 7-second video of $N$ drones flocking across a map using the boids algorithm. At a random interval (2–4 seconds), one drone disappeared. The participant's task was to identify the color of the vanished drone.
• Metrics:
  • Objective Performance: Accuracy in identifying the drone's color.
  • Subjective Workload: Self-reported mental workload on a 1–7 scale after each episode.
  • Post-Task Surveys: Included NASA-TLX, PANAS-X (for emotional response/surprise), and qualitative feedback.
• Experimental Structure:
  • Study 1 (Small Changes): Tested 9 swarm sizes ($N \in \{4, 6, ..., 20\}$) with small transitions ($\Delta = \pm 2$). This tested H1 (workload residue after decrease) and H2 (easing-in with gradual increase).
  • Study 2 (Large Changes): Tested 5 swarm sizes ($N \in \{8, 12, 16, 20, 24\}$) with large transitions ($\Delta = \pm 4, \pm 8, \pm 12, \pm 16$). This tested H3 (large jumps inducing a cognitive reset/shock).
  • Design: Episodes were grouped into "decks" using Eulerian cycles to ensure every transition occurred an equal number of times, minimizing positional bias.

3. Key Hypotheses

1. H1 (Workload Residue): Following a period of high demand (large swarm), reducing the swarm size will leave operator workload elevated compared to a baseline of the same size without prior overload.
2. H2 (Easing-In): Reaching a specific swarm size via gradual increases results in lower workload than reaching it via a single abrupt increase.
3. H3 (Cognitive Reset/Shock): Large magnitude changes in swarm size will override the effects of H1 and H2, causing a "shock" that resets the operator's cognitive state, leading to deviations from the baseline regardless of direction.

4. Key Results

• Objective Performance:
  • Small Changes ($\Delta = \pm 2, \pm 4$): Performance (accuracy) remained largely unaffected by the direction or magnitude of small changes.
  • Large Changes ($\Delta \ge 8$): A clear pattern emerged where large increases in swarm size caused a drop in accuracy, while large decreases improved accuracy relative to the baseline.
• Subjective Workload:
  • Small Decreases: Confirmed H1. When the swarm size decreased, workload did not return to baseline immediately; a "workload residue" persisted, leaving operators feeling more burdened than if they had started at that lower size.
  • Small Increases: Confirmed H2. Gradual increases were associated with equal or lower workload compared to the baseline, suggesting an "easing-in" effect.
  • Large Changes: Confirmed H3. As the magnitude of the change increased ($\Delta \ge 8$), the directional differences (residue vs. easing-in) collapsed. Large jumps, particularly increases, induced a "shock" or startle response (supported by PANAS-X surprise scores), effectively resetting the operator's cognitive state and invalidating the history effects seen in smaller changes.
• Qualitative Feedback: Participants reported feeling "relieved" but still "disrupted" after decreases (supporting residue) and "panicked" or "startled" by large increases.

5. Key Contributions

1. Empirical Evidence of Workload Hysteresis: The study provides quantitative proof that mental workload in HSI exhibits hysteresis; the history of the swarm size influences current workload perception, specifically creating a residue effect after reductions.
2. Characterization of Change Magnitude: The paper delineates a threshold where small, manageable transitions (allowing for easing-in or residue) give way to large, disruptive transitions that trigger a cognitive reset.
3. Actionable Design Guidelines: The findings offer specific strategies for interface designers and mission planners:
   • Avoid Abrupt Decreases: Do not assume that reducing the number of robots immediately reduces operator stress; the cognitive load may linger.
   • Utilize Gradual Increases: "Ease-in" operators to larger swarms through small, incremental steps to maintain lower workload.
   • Strategic Use of Large Jumps: While large jumps are disruptive, they can be used intentionally to "reset" an operator's cognitive state if they are stuck in a high-workload residue, though this comes at the cost of temporary performance degradation and stress.

6. Significance

This research moves beyond static swarm size assumptions to address the dynamic reality of robotic deployments. It highlights that workload management is not just about the current number of agents, but the trajectory of change.

The findings are critical for the design of robust HSI interfaces and adaptive autonomy systems. By understanding the "workload residue" and "shock" effects, system architects can better schedule robot redeployment, failure recovery, and mission scaling to prevent operator overload or under-engagement. The study suggests that managing the flow of swarm size changes is as important as the swarm size itself for maintaining human operator effectiveness in safety-critical environments.