Towards Intelligent UAV Swarms through Active Inference-Driven World Modeling

This research focuses on developing a self-learning, rational-based trajectory design framework for swarms of unmanned aerial vehicles (UAVs). The proposed framework enables autonomous, effective, safe, and energy-efficient trajectory planning, particularly in dynamic and uncertain environments where real-time adaptation, inter-UAV coordination, and collision avoidance are simultaneously required. The study begins with a comprehensive and critical review of existing approaches to UAV navigation and trajectory planning, comparing traditional, biologically inspired, and artificial intelligence–based algorithms. Through this comparative analysis, key factors such as computational efficiency, scalability, energy consumption, robustness under uncertainty, and the balance between centralized and decentralized control are systematically evaluated. The review also identifies significant research gaps, particularly in real-time adaptation and explainable decision-making, highlighting the need for more flexible, safety-critical, and generalizable strategies in future UAV swarm systems. Subsequently, the research introduces a novel GA–RF framework, which integrates genetic algorithms (GA) and repulsion forces (RF) to optimize the paths of multiple UAVs, minimizing collisions, overlaps, and interference. Coordination at the cluster level is enhanced through improved task selection and visiting order classification, particularly within the context of the Multi-Traveling Salesman Problem (MTSP). Simulation experiments demonstrate that this approach achieves shorter travel distances, better interference avoidance, and more efficient navigation compared to traditional heuristic and metaheuristic algorithms such as PSO, ACO, SA, and 2-OPT. The third part of the research presents a novel Active Inference–driven World Modeling for an adaptive UAV swarm Trajectory Design. The proposed framework enables UAVs to autonomously perform mission distribution, route ordering, and motion planning through probabilistic reasoning and self-learning. In the offline phase, expert trajectories are generated using a GA–RF optimizer and used to train a World Model that captures UAV swarm behavior across mission, route, and motion abstraction levels. During online operation, each UAV infers optimal actions by continuously minimizing divergence between current beliefs and reference states encoded in the world model, allowing the swarm to adapt to new targets and environmental changes in real time. The results demonstrate faster convergence, improved stability, and safer navigation compared to Q-Learning, establishing the proposed framework emerge as a scalable and knowledge-based solution for future intelligent UAV swarm networks. Moreover, the proposed model also performed effectively when tested on real-time simulated data, further strengthening its generalization ability and applicability to real-world scenarios. Overall, this research presents a scalable, secure, and domain knowledge–driven framework for UAV swarm control, offering an effective and promising direction for the advancement of AI-powered autonomous aerial networks.