This is the authors' abstract. We don't add key points for this paper.
Unmanned Aerial Vehicles (UAVs) are increasingly deployed for wildlife monitoring, anti-poaching patrols, and search-and-rescue operations. Deployment in remote environments remains constrained by three interdependent challenges: energy-efficient area coverage, real-time responsiveness to dynamic obstacles, and reliable safety intervention under limited onboard computational power. Existing solutions address these challenges in isolation, producing architectures that are inefficient, unreliable, or reliant on probabilistic safety mechanisms unsuitable for safety-critical missions. The objective of this study is to design, implement, and statistically validate a hierarchical hybrid navigation framework that simultaneously addresses all three challenges within a computationally efficient architecture suitable for edge-deployed wildlife monitoring UAVs. The framework separates autonomy into three coordinated layers: (1) a strategic layer employing a Genetic Algorithm (GA) enhanced with a 2-Opt topological mutation for globally efficient path planning; (2) a tactical layer driven by a Twin Delayed Deep Deterministic Policy Gradient (TD3) agent to generate smooth, energy-aware local trajectories; and (3) a reactive layer based on an Adaptive Fuzzy Logic (AFL) safety governor providing rule-based, low-latency, interrupt-driven collision avoidance. To mitigate deep reinforcement learning failure modes, the TD3 agent is trained via a three-stage curriculum learning strategy, validated in a 2D kinematic sandbox before transfer to a high-fidelity 3D simulator. The system was evaluated across two AirSim environments (Africa and LandscapeMountains) under dynamic obstacle conditions. Statistical validation across n = 20 independent planner trials and n = 15 ablation trials confirmed 100% mission completion, a 70% reduction in control jerk versus the PID baseline (41.97 versus 139.84 m/s 3 , p < 0.01), and zero collisions under AFL supervision across all five stress-test scenarios in both environments. Inference profiling confirmed an actor model size of 0.276 MB and a mean latency of 0.097 ms on CPU hardware, supporting compatibility with resource-constrained edge platforms. These results demonstrate that integrating evolutionary optimization, deep reinforcement learning, and fuzzy-logic safety supervision produces a statistically validated and computationally feasible autonomous navigation system for wildlife conservation.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Talib et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: