American Institute of Aeronautics and Astronautics

This study introduces a simulation-based cooperative framework that integrates unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs) for autonomous road inspection and mapping. The proposed system leverages Multi-Agent Reinforcement Learning (MARL) to enable adaptive coordination, dynamic task allocation, and efficient mission planning across heterogeneous agents. Built within a high-fidelity CARLA–AirSim simulation environment, the framework allows extensive experimentation under controlled weather, lighting, and communication conditions, ensuring reproducibility and scalability. Quantitative evaluations demonstrate that MARL-based coordination significantly enhances overall mission performance compared to conventional rule-based strategies. Coverage improves from 89.3% to 96.1%, while redundant path overlap decreases by 44.3%. Mission completion time is reduced by 17.5%, and energy consumption drops by 14.7%, highlighting the efficiency gains achieved through learning-based cooperation. Moreover, integrating aerial and ground observations increases detection accuracy (IoU +9.9%, mAP +8.1%) and improves spatial map continuity to 98.7%, with a 55% reduction in geo-registration error. The results confirm that cooperative intelligence can substantially enhance the accuracy, efficiency, and reliability of autonomous infrastructure monitoring.

This work presents a simulation-based framework for cooperative road inspection using drones and mobile robots coordinated through multi-agent reinforcement learning (MARL). The framework integrates adaptive task allocation, multi-modal data fusion, and communication-aware modeling within a unified CARLA-AirSim environment, allowing systematic analysis of individual module contributions to overall mission performance. Results from extensive ablation experiments indicate that MARL-based coordination substantially improves coverage efficiency, detection accuracy, and energy utilization when compared with rule-based strategies. In addition, fusion of aerial and ground-level observations enhances defect characterization by up to 9.9% and reduces geo-registration error by more than 50%, while maintaining stable coordination under simulated communication delays of up to 100 ms. Scalability experiments further show that increasing the number of agents improves mission completeness and reduces overall energy consumption, although performance gains begin to saturate beyond a certain agent density.

ABSTRACT Small object detection, a critical technique for recognizing and localizing diminutive targets in visual data, plays a vital role in applications ranging from remote sensing and unmanned aerial vehicle (UAV) vision to autonomous driving. Current methodologies, however, face substantial challenges, including detection accuracy limitations, low‐resolution image processing difficulties, background noise interference, and target occlusion issues. To address these challenges, we propose FMF‐DETR, an innovative small object detection framework featuring frequency‐domain feature optimization through three key components: (1) High‐Low Frequency Fusion Model (HLFM), (2) Focused Diffusion Feature Pyramid Network (FDFPN), and (3) BiPathNet (BPNet). Specifically, the HLFM module enhances multiscale feature representation by emphasizing high‐frequency details while suppressing low‐frequency background noise. The FDFPN architecture improves detection performance in complex scenarios through multiscale feature fusion and saliency‐aware diffusion. BPNet introduces a dual‐path feature extraction mechanism that simultaneously enhances feature discriminability and reduces computational overhead. Through the synergistic integration of these components, the proposed framework enhances both detection accuracy and operational efficiency. Comprehensive evaluations on the VisDrone dataset demonstrate FMF‐DETR's superior performance, achieving a 2.2% accuracy improvement while reducing model parameters by 13.03M and computational complexity by 97.2G FLOPs compared to baseline methods. These results validate both the effectiveness and efficiency of our proposed framework.

Abstract With the continuous increase in turbine inlet temperature, traditional uncooled radial-inflow turbines are becoming inadequate for operation in higher-temperature environments. This study investigates both the overall layout of internal cooling passages and the characteristics of local cooling structures for a radial-inflow turbine. Using a conjugate heat transfer numerical approach, four cooling schemes are evaluated from the perspectives of cooling efficiency, as well as turbine stage aerodynamic performance. To enhance the thermal protection of the wheel, a novel sunken-type disk cooling scheme is first proposed. In this design, a portion of the coolant after being used for blade cooling is redirected toward the disk region, resulting in a reduction in both disk temperature and the temperature in high-stress root regions of the blade. To reduce the aerodynamic efficiency losses caused by the conventional full-cut trailing-edge slot design, this study proposed a novel pressure-side slot near the trailing edge. This approach preserves the structural integrity of the trailing edge and significantly improves the aerodynamic performance of the turbine stage. Turbine stage efficiency assessments reveal that the commonly used full-cut trailing-edge cooling design provides the least structural retention at the trailing edge, resulting in a 15.5% drop in aerodynamic turbine stage efficiency compared to the uncooled baseline. In contrast, the pressure-side trailing-edge slot cooling configuration offers a minimal aerodynamic efficiency reduction of 2.7% relative to the uncooled blade. The study also analyzes the flow and heat transfer characteristics associated with leading-edge, blade-tip, and trailing-edge cooling designs, summarizing the underlying fluid-thermal interaction mechanisms.

Abstract Legged robots operating in irregular environments are often subjected to compound disturbances such as tilts and vibrations, which can degrade attitude stability and motion reliability. This paper presents a real-time disturbance-adaptive control framework for a hexapod robot. The proposed system integrates quaternion-based attitude estimation using an extended Kalman filter (EKF), a double-threshold pose classifier, and a modular gait library and is implemented on an embedded controller with a 2 ms control-loop latency. Analytical verification and laboratory experiments demonstrate that the proposed control loop achieves uniform ultimate boundedness (UUB) under deterministic hybrid disturbances composed of controlled tilt and vibration, with a mean recovery time of 5.7 s. These results demonstrate that a lightweight rule-based controller can ensure reliable posture recovery within the experimentally validated laboratory scenarios, providing a foundation for future extensions to more complex environments. The main contributions of this work are (1) a disturbance-adaptive gait selection architecture for quasi-static stabilization, (2) a noise-robust EKF-based attitude estimation and double-threshold pose determination scheme, and (3) a concise Lyapunov-based stability analysis demonstrating UUB of the closed-loop system.