Q-learning global path planning for UAV navigation with pondered priorities

Abstract

The process of path planning plays a crucial role in enabling self-directed movement, particularly for unmanned aerial vehicles. This involves accommodating diverse priorities, such as route length, safety, and energy efficiency. Traditional techniques, including geometric and dynamic programming, have historically been employed to address this challenge. However, recent years have testified to an increasing prevalence of artificial intelligence methodologies such as reinforcement learning. This study introduces a novel approach to offline path planning in static environments, utilizing Q-learning as its foundation. The method optimizes three pivotal factors: path length, safety, and energy consumption. By effectively balancing exploration and exploitation, this technique enables an autonomous agent to efficiently navigate from any initial point to a specified destination on the map. To evaluate the proposed strategy’s effectiveness, extensive simulations are conducted across diverse environments. A comparative analysis with three established strategies showcases the algorithm’s proficiency in generating feasible routes. The user can freely tailor the system’s priorities by modifying each of their weights prior to training. Additionally, scalability tests reveal the algorithm’s swift convergence, achieving stability within just 35 s for larger environments spanning up to 40 × 40 units. To further validate the proposed approach, both simulations and real-world experiments are employed, collectively demonstrating its performance and applicability.