Source-linked AI summary
TriSAR: Task Coordination and Collision Avoidance for Aerial Robot Teams in Disaster Response
Aditya Anil Kapile, Pedro Machado, Isibor Kennedy Ihianle
TL;DR
Multi-UAV disaster-response coordination needs controlled evidence on the separate contributions of task allocation and local trajectory safety. TriSAR evaluates a hybrid GA–PSO system with a 2 × 2 ablation across allocation strategies and reactive repulsion. Repulsion reduced collision-threshold violations under both allocators, while GA improved efficiency only when repulsion was disabled.
Problem
The study addresses limited controlled evidence about how task-allocation and trajectory-control layers contribute individually and jointly to multi-UAV mission efficiency and safety.
Method
TriSAR combines GA task allocation, PSO trajectory control, and reactive repulsion, evaluated in a five-UAV, eight-target Gazebo scenario using a 2 × 2 factorial ablation.
Results
Reactive repulsion reduced collision-threshold violations under both allocators, while GA showed no efficiency advantage with repulsion enabled but significant advantages in steps, path length, and energy when disabled.
Takeaways & Limitations
The safety benefit of reactive repulsion is allocation-dependent in effect size, and GA’s efficiency benefit appears only without repulsion in the tested scenario.
Takeaways & Limitations
The ablation uses a single small scenario, so GA performance should not be generalized beyond that scale without larger-scale testing.
Abstract
from arXiv · showhide
Multi-Unmanned Aerial Vehicle (UAV) disaster-response systems require coordinated task assignment and local trajectory control, yet the individual and combined contributions of these coordination layers to mission efficiency and operational safety remain insufficiently characterised under controlled experimental conditions. TriSAR is evaluated as a five-UAV coordination system operating in a physics-based Gazebo simulation of an earthquake-damaged urban environment. A 2 x 2 factorial design compares two task-allocation strategies (Genetic Algorithm and greedy fitness-based allocation) with reactive collision avoidance enabled or disabled. Each of the four configurations was evaluated over 30 stochastic episodes in a common scenario of five UAVs and eight targets. Under greedy allocation, enabling repulsion eliminated recorded collision-threshold violations, confirmed by a Mann-Whitney test (U = 885, p = 4.03 x 10^-12, rank-biserial r = 0.97). Under GA allocation, the same protective effect was confirmed (U = 675, p = 1.26 x 10^-5, rank-biserial r = 0.50). For mission-efficiency metrics, GA-based allocation showed no statistically detectable advantage over greedy allocation when repulsion was enabled, but a significant advantage in steps, path length, and energy when repulsion was disabled (Welch's t-tests, |g| between 0.92 and 1.76). These results show that reactive repulsion provides a substantial, allocation-dependent safety benefit, while the additional computational complexity of GA-based task allocation yields a detectable mission-efficiency benefit only when repulsion is disabled.
I. INTRODUCTION
TriSAR addresses limited controlled evidence about how global task allocation and local collision avoidance independently contribute to multi-UAV disaster-response performance. It uses a factorial ablation to test efficiency, safety, and allocation-dependent effects.
- Coordinated disaster response requires assigning targets to UAVs while navigating safely around neighbouring agents and obstacles.
- The study isolates coordination-layer contributions instead of evaluating only aggregate performance from a fused architecture.
- TriSAR combines a GA task allocator, PSO trajectory controller, and reactive repulsion in a physics-based Gazebo environment.
- A controlled 2 × 2 factorial ablation compares allocation strategy with reactive collision avoidance enabled or disabled.
- GA provides no detectable efficiency advantage with repulsion enabled, but significant advantages on all three efficiency metrics when repulsion is disabled.
- Reactive repulsion improves safety under both allocators, with a substantially larger effect size under greedy allocation.
III. SYSTEM MODEL
The system model represents TriSAR in a simulated urban environment with five homogeneous quadcopters and eight mission targets. Agents are initialized in a circular formation to avoid positional clustering.
- Figure 1 depicts four layers: agent and kinematic state, coordination, middleware and simulation, and deployment and evaluation.
- The simulated metropolis spans 500 m × 500 m × 120 m and contains N = 5 homogeneous quadcopters.
- Three search-and-rescue agents address M = 8 targets, while two relay agents serve fixed rooftop relay stations.
- Figure 2 shows agents spawned in a circular ring with N = 5 and R = 4.0 m.
- The ring provides equal angular dispersion and radial spacing, avoiding initial positional clustering among UAVs.
IV. PROPOSED HYBRID GA-PSO COORDINATION
TriSAR implements a closed-loop coordination pipeline that registers tasks, allocates them with GA-based search, controls trajectories with PSO, and updates energy each simulation step.
- The coordination pipeline runs from shared-blackboard task registration through GA allocation and PSO trajectory control to per-step energy updating.
A. Task Allocation
Task allocation encodes assignment priorities as permutations and decodes them using a shared agent–task cost function. The GA searches these priorities while preserving the same cost model used by both allocation strategies.
- A chromosome is a permutation of task indices representing assignment priority order.
- Decoding assigns each task to the available agent with the lowest assignment cost c(a, t).
- The cost function combines distance, battery level, target urgency, and rooftop-type climb cost.
- With three search-and-rescue agents and eight targets, decoding assigns one task per available agent across successive rounds.
- Both allocation strategies use the same agent-task cost function, isolating the effect of allocation search strategy.
B. Trajectory Control
TriSAR updates per-agent candidate velocities with a local PSO variant, then applies reactive repulsion separately to the selected velocity. The pipeline closes the loop through task allocation, trajectory control, energy updates, and state feedback.
- Trajectory optimisation: At 10 Hz, each agent evaluates P = 10 candidate velocity vectors at projected positions and updates them for K = 5 inner iterations.Candidates are reinitialised around the current target-heading direction at each control step.
- Trajectory optimisation: Each agent tracks its own best candidate and population-wide best velocity, making the optimiser per-agent and velocity-space based.The formulation does not use one fleet-wide global-best velocity or search directly over position.
- Pipeline: The coordination pipeline runs from blackboard task registration through allocation, trajectory control, energy update, and per-step state feedback.Relay agents use fixed rooftop stations separately from the registered mission targets.
- Reactive repulsion: After optimisation selects the lowest-cost candidate velocity, repulsion separately adjusts it for nearby neighbours or obstacles within dsafe = 2.5 m.The adjustment scales unit vectors by fractional violation depth before summation.
- Reactive repulsion: Repulsion uses an artificial-potential-field structure and therefore inherits local-minima susceptibility.The repulsion gain is krep = 3.0 m/s.
C. Energy Model
TriSAR models battery drain with a cubic-drag formulation and uses the resulting energy measure as a normalised simulation proxy for relative condition comparisons.
- Energy model: Battery drain follows a cubic-drag model based on the parasitic-power term of Zeng et al.
- Energy model: The coefficients Phover = 0.28, Cdrag = 7 × 10^-5, and Cmass = 0.05 parameterise simulation-scale battery depletion.They are expressed as percentage points per second per corresponding unit.
- Interpretation: Energy is a normalised simulation proxy for relative comparisons, not a predictor of real UAV energy consumption.
V. EXPERIMENTAL SETUP
The experiment compares four TriSAR variants in a fixed five-drone, eight-target Gazebo scenario using independent stochastic episodes. Safety is measured by distance-threshold violation-steps, alongside Welch’s GA-versus-greedy efficiency tests.
- Scenario and variants: Each of four variants was evaluated in n = 30 independent episodes with five drones and eight targets in Gazebo Sim.The targets comprise four rooftop and four ground-level locations.
- Scenario and variants: The variants combine GA or greedy allocation with repulsion enabled or disabled: Full, No-GA, No-Repulsion, and Floor.Ground-level coordinates receive independent ±2 m axis jitter across episodes, while rooftop coordinates and environmental conditions remain fixed.
- Safety outcome: A collision-threshold violation is a realised pairwise separation below dsafe = 2.5 m, counted once per control step and combined across drone-drone and drone-obstacle events.No drone-obstacle violations were recorded in the reported batch.
- Statistical analysis: Welch’s t-tests with 95% confidence intervals and Hedges’ g compare GA against greedy allocation for steps, path length, and energy.
VI. RESULTS AND DISCUSSION
Reactive repulsion reduced collision-threshold violations under both allocators, while GA’s efficiency advantage appeared only without repulsion. The conditional pattern was observed in a common n = 30-per-condition ablation batch.
- Experimental basis: n = 30 episodes per condition provided the directly comparable ablation batch for all reported results.
- Mission efficiency: With repulsion enabled, GA showed no detectable efficiency advantage over greedy allocation; with repulsion disabled, GA significantly outperformed greedy on steps, path length, and energy.The disabled-repulsion comparisons had |g| between 0.92 and 1.76.
- Safety: U = 885, p = 4.03 × 10^-12, rank-biserial r = 0.97 confirmed repulsion’s protective effect under greedy allocation.The comparison was between Floor and No-GA.
- Safety: U = 675, p = 1.26 × 10^-5, rank-biserial r = 0.50 confirmed the same protective effect under GA allocation.The comparison was between Full and No-Repulsion.
- Safety: No-Repulsion recorded violations in 15/30 episodes, compared with 29/30 for Floor.
B. Ablation Study
The ablation compares efficiency and safety across four allocator–repulsion variants using per-episode distributions and representative trajectories. Repulsion consistently reduces violations, while GA’s efficiency advantage appears only without repulsion and may depend on scenario scale.
- Efficiency: Figure 4 shows right-skewed per-episode efficiency distributions, with means above medians and at least one high-value outlier in every variant.Welch’s test was used because it does not assume equal variance.
- Safety: Repulsion reduced collision-threshold violations under both allocation strategies, with stronger statistical effect under greedy allocation.Greedy comparisons yielded U = 885, p = 4.03 × 10^-12, r = 0.97; GA comparisons yielded U = 675, p = 1.26 × 10^-5, r = 0.50.
- Safety: 29 of 30 Floor episodes and 15 of 30 No-Repulsion episodes recorded at least one violation-step, versus zero in both repulsion-enabled conditions.Floor denotes greedy allocation without repulsion; No-Repulsion denotes GA allocation without repulsion.
- Safety: Figure 5 illustrates a No-Repulsion trajectory reaching 0.911 m separation at step 18, below the 2.5 m threshold, while the Full episode maintains separation.The figure is illustrative and was generated under an earlier system version rather than the reported batch.
- Efficiency: GA allocation improved steps, path length, and energy relative to greedy allocation only when repulsion was disabled.With repulsion enabled, no GA-versus-greedy efficiency difference reached significance; with repulsion disabled, all three comparisons were significant with large effects.
- Interpretation: The GA efficiency advantage is reported as conditional on repulsion and scenario scale, motivating tests at larger or more asymmetric problem sizes.The authors propose evaluating larger N and M and asymmetric target distributions.
C. Computational Cost
GA allocation required substantially more computation than greedy allocation, although the measured absolute overhead was small relative to episode flight time. Timing interpretation is limited because the measurements came from a separate verification sample rather than the canonical batch.
- Timing: GA allocation averaged 86.53 ± 13.83 ms, compared with 0.174 ± 0.039 ms for greedy allocation.GA times ranged from 67.6–109.3 ms, while greedy times ranged from 0.11–0.21 ms.
- Caveat: Timing results came from a separate n = 5 verification sample per allocator because allocation time was not logged for the main n = 30 batch.The canonical efficiency and safety results use the n = 30 batch, whereas the timing estimates do not.
- Timing: GA incurred an approximate 500× computational overhead relative to greedy allocation.Greedy allocation was also more consistent in timing.
- Interpretation: The absolute allocation overhead was small relative to multi-second per-episode flight time and did not materially affect reported mission-efficiency metrics.The authors nevertheless identify computational cost as relevant to future scaling of GA allocation.
- Timing: The GA stopped after 16–19 generations in every timing run, well below the 100-generation cap.This was reported as consistent behavior at the tested scenario size, while larger problems may behave differently.
VII. CONCLUSION AND FUTURE WORK
TriSAR’s controlled factorial ablation found that reactive collision avoidance reduced collision-threshold violations under both allocation strategies, while GA efficiency advantages depended on repulsion being disabled. The authors also identify reproducibility resources and several directions for extending validation.
- Conclusion: Reactive collision avoidance measurably reduced collision-threshold violations under both greedy and GA allocation in the tested scenario.The reported tests were Mann-Whitney comparisons across the allocation strategies.
- Conclusion: GA-based allocation had no detectable efficiency advantage over greedy allocation with repulsion enabled, but significantly outperformed it on all three efficiency metrics when repulsion was disabled.The comparison used Welch’s t-tests and covered steps, path length, and energy.
- Future Work: The study recommends larger asymmetric scenarios, formal interaction modelling, external baselines, heterogeneous fleets, and simulation-to-real validation.It also proposes benchmarking reciprocal velocity obstacles to address the potential-field-like repulsion term’s local-minima vulnerability.
- Reproducibility: Source code, simulation configurations, experiment scripts, and data-processing pipelines are publicly available for reproducing the reported results.The repository is identified as the TriSAR GitHub project.