Key points are not available for this paper at this time.
Autonomous vehicles tasked with area coverage in cluttered environments, such as for search and rescue or precision agriculture, require path planners that are not only optimal but also robust and adaptable. Traditional planning methods often fail to balance these demands, struggling with computational complexity in dynamic settings or a lack of generalization in unseen, complex environments. This article presents a systematic comparative study of two distinct algorithmic paradigms to address this challenge: a bio-inspired metaheuristic using Particle Swarm Optimization (PSO) and a machine learning-based approach using Neural Networks (NN). We develop a PSO-based planner integrated within a Moving Horizon Planning (MHP) framework for real-time, online trajectory optimization. In parallel, we design and train a deep NN on a diverse set of simulated environments, enabling it to learn and generalize efficient coverage paths. The performance of both methods is rigorously evaluated across various maze-like scenarios, with key metrics including path completeness, computational efficiency, and robustness, the latter assessed through extensive Monte Carlo simulations. Our findings demonstrate a clear trade-off: the PSO-MHP method excels in dynamic replanning and achieves efficient global optimization with fewer iterations, while the NN planner provides superior generalization and faster online inference in complex, unstructured environments. The primary contribution of this work is a detailed comparative synthesis that delineates the specific strengths and limitations of each approach. Furthermore, these insights directly pave the way for a novel hybrid NN-PSO framework, proposed as future work, which aims to synergistically combine the real-time adaptability of PSO with the generalization speed of NNs for robust autonomous navigation.
Ibrahim et al. (Wed,) studied this question.