Liquid Neural Network-Based Multi-UAV PathPlanning in Dynamic, Obstacle-AwareEnvironments

Deploying multiple unmanned aerial vehicles (UAVs) for last-mile delivery in urban airspace requires real-time path planning under non-stationary conditions such as moving obstacles, intermittently active no-fly zones, variable wind, and tight energy budgets, while coordinating several vehicles simultaneously. Classical planners offer optimality guarantees but become brittle and computationally expensive when frequent re-planning is needed, whereas standard recurrent policies trained by behavioral cloning are prone to compounding errors under distribution shift. We propose a decentralized planning framework in which each UAV is governed by a compact Liquid Neural Network (LNN). The LNN’s continuous-time dynamics and learnable per-neuron time constants provide a structural inductive bias that filters transient disturbances while remaining responsive to persistent environmental changes. Policies are trained via masked behavioral cloning from a centralized, coordinated A* oracle, with legality masks enforcing collision and airspace constraints during both training and inference. Systematic evaluation on a reproducible multi-UAV delivery benchmark shows that the LNN policy consistently achieves the highest success rate, the shortest paths, and the lowest energy-failure rate among all non-oracle methods, substantially outperforming both a GRU sequence-model baseline and a strong greedy heuristic. An inference-time ablation confirms that the persistent LTC hidden state contributes measurably to closed-loop robustness beyond the feedforward mapping alone. The framework, simulator, and trained models are publicly available to support further research on energy-aware autonomous aerial logistics.