Intelligent methods for collision avoidance in UAV group flights using radar: a hybrid approach with trajectory prediction and federated adaptation

Abstract

The paper addresses one of the central challenges in modern multi-agent robotic systems: ensu-ring guaranteed flight safety for unmanned aerial vehicle (UAV) swarms under high density, dynamic obstacles, and partial observability. The fundamental limitations of traditional reactive methods are substantiated: Artificial Potential Fields (APF) suffer from local minima and oscillations, while geometric methods (such as ORCA) rely on constant-velocity assumptions, leading to excessive conservatism and deadlocks in maneuvering swarms. To overcome these challenges, a novel decentralized hybrid collision avoidance system, designated AI-RDR, is proposed.
The methodology integrates four core functional modules: (1) onboard FMCW radar perception robust to adverse atmospheric conditions; (2) nonlinear trajectory prediction of neighboring agents via an Attention-Bi-GRU recurrent neural network; (3) decentralized continuous 3D path planning using a modified Ant Colony Optimization (ACO) algorithm with a risk-aware heuristic driven by probabilistic occupancy predictions; and (4) continuous swarm-wide model adaptation based on the FedProx federated learning framework, which guarantees stable convergence under non-IID data distributions without centralizing raw sensor feeds.
Experimental validation was conducted within a high-fidelity simulation setup combining Ga-zebo 11, PX4 Autopilot in Software-in-the-Loop (SITL) mode, and ROS 2 nodes. Comparative evaluations demonstrate that the proposed approach achieves a collision-free success rate (SR) exceeding 99.8% in dense dynamic scenarios, significantly outperforming conventional APF and ORCA algorithms. The system maintains higher safety margins (AMD = 4.8 m) and path efficiency (PE = 0.92) with a decision-loop latency of 45 ms, confirming real-time feasibility for onboard deployment. The scientific novelty lies in establishing a synergistic framework where state forecasting directly feeds into risk-bounded planning, alongside the first application of FedProx for online adaptation of motion predictors in UAV swarms.

Keywords: UAV swarm, collision avoidance, FMCW radar, artificial intelligence, trajectory prediction, Attention-Bi-GRU, Ant Colony Optimization, federated learning, FedProx, proactive navigation, ROS 2, Gazebo/PX4.