Abstract

The rising global urban population has led to increased vehicles on the road and the demand for faster and more efficient transportation. As a result, the need for low-altitude spaces to support safe Unmanned Aerial Vehicles (UAV) operations has become increasingly urgent. To address this need, countries are constructing systems and implementing relevant regulations and techniques to ensure the safety of people and properties on the ground and in the air while enabling aerial vehicles to navigate and complete tasks autonomously amidst uncertainties.
In recent years, the rapid advancement in Artificial Intelligence (AI) and a corresponding exponential increase in computing power have unlocked new possibilities. This synergy enables UAVs equipped with AI abilities to improve over time and perform complex tasks more adeptly than traditional models. This innovative direction not only expands the range of tasks that UAVs can undertake but also enhances their safety, opening up a new dimension in various sectors of human activity by leveraging the evolving capabilities of intelligent systems.
This research advances the field of safe and intelligent UAV development through a systematic, four-stage process. In the initial stage, the exploitation and improvement of advanced sensors facilitated the integration of their outputs with control systems, leading to the creation of an emergency landing system that enhances reliability. The second stage involved the design of several UAV autonomous navigation and obstacle avoidance algorithms based on diverse Reinforcement Learning (RL) algorithms. This phase also explored the performance of multiple UAV units operating in adversarial and cooperative modes, laying the groundwork for subsequent studies.
Building on the aforementioned algorithms, the third stage saw the establishment of a hybrid intelligent SAC-FIS controller. This system combines an enhanced Soft Actor-Critic (SAC) method with a Fuzzy Inference System (FIS), integrating universal expert experience to streamline the learning process. It ensures real-time path planning in three-dimensional spaces, enabling UAVs to dodge obstacles and intercept multiple dynamic targets. It addresses two significant challenges in RL: dynamic-environment problems and multi-target dilemmas.
In the final stage, a multi-layered control framework was constructed to integrate all previously developed algorithms and functionalities. This structure enables decentralized swarm control among a custom number of UAVs for dynamic target interception and includes a distributed communication strategy for effective and dynamic target allocation. The system enhances robustness and maintains low risk by activating a failsafe mode for an emergency landing whenever an individual unit in the swarm fails, allowing other units to continue and complete the mission.
The effectiveness and superiority of all proposed algorithms and the control framework have been validated by simulations and experiments. Comparisons with previously published research work highlight the enhanced efficiency and higher success rate of these approaches.