Robust Geospatial Coordination of Multi-Agent Communications Networks Under Attrition

TL;DR

The paper tackles robust multi-agent geospatial networking under attrition (RTNUA) by formalizing the problem and introducing ΦIREMAN, a physics-informed, topological swarm algorithm that proactively builds redundant network geometries. ΦIREMAN uses a Task-Space Potential Field and Semi-Steiner tree concepts to drive drones into hexagonal-like configurations that sustain connectivity despite drone losses, without heavy learning or centralized control. Comprehensive simulations across 25 configurations show ΦIREMAN consistently outperforms the DCCRS baseline, achieving high task uptime even at large scales (100 tasks, 500 drones) and under substantial attrition. The work highlights how problem size and attrition rate constrain uptime, and demonstrates the value of geometry-aware, locally governed swarm dynamics for robust emergency communications networks.

Abstract

Fast, efficient, robust communication during wildfire and other emergency responses is critical. One way to achieve this is by coordinating swarms of autonomous aerial vehicles carrying communications equipment to form an ad-hoc network connecting emergency response personnel to both each other and central command. However, operating in such extreme environments may lead to individual networking agents being damaged or rendered inoperable, which could bring down the network and interrupt communications. To overcome this challenge and enable multi-agent UAV networking in difficult environments, this paper introduces and formalizes the problem of Robust Task Networking Under Attrition (RTNUA), which extends connectivity maintenance in multi-robot systems to explicitly address proactive redundancy and attrition recovery. We introduce Physics-Informed Robust Employment of Multi-Agent Networks ($Φ$IREMAN), a topological algorithm leveraging physics-inspired potential fields to solve this problem. Through simulation across 25 problem configurations, $Φ$IREMAN consistently outperforms the DCCRS baseline, and on large-scale problems with up to 100 tasks and 500 drones, maintains $>99.9\%$ task uptime despite substantial attrition, demonstrating both effectiveness and scalability.

Figures (5)

• Figure 1: Semi-Steiner task tree example with $c_b = 0.3$, balancing total edge length ($l_{\Sigma e} = 19.86$) against sum of distances to base station ($l_{\Sigma b} = 39.53$). Steiner nodes (circles) enable more efficient network topologies.
• Figure 2: Left; hypothetical task tree graph using only sum of distances, right; task tree graph including sum of distances to base station.
• Figure 3: Performance over time of $\Phi$IREMAN under the default configuration. Thick lines give per-timestep averages over 100 simulations. "Alive Agents" shows how many drone agents remain alive; "Alive Agents Active" shows the portion of alive agents which are networked; "Connected Tasks" shows how many tasks are connected; and "Connected Tasks (Ret)" shows how many of the previously connected tasks are currently connected. Once all tasks are connected for the first time, these last two overlap. The integral of the former is $TU1$, and the integral of the latter is $TU2$.
• Figure 4: Evolution of drone network under default configurations at $t = 0, 40, 80, 120, 160, 200$ timesteps. Alive drones (green triangles when active, red when disconnected) maintain connectivity between tasks (purple squares when networked, gray when disconnected) and base station (green square) despite ongoing attrition (gray triangles). The gray square is the field border. Green lines are network connections.
• Figure 5: Heatmap of differences between $\Phi$IREMAN and DCCRS $TU1$ and $TU2$ scores on problem configurations. Values shown are the result of subtracting the $TU1$ and $TU2$ scores of DCCRS from those of $\Phi$IREMAN for each configuration. Note that $\Phi$IREMAN generally outperforms DCCRS across all problems and has a maximum performance deficit of only 0.1 percentage points on $TU1$.