Legged robots face inherent challenges in energy efficiency and stability at high speeds due to the repetitive acceleration–deceleration cycles of swing-based locomotion. To address these limitations, this paper presents a motion strategy that uses rolling gait instead of swing gait to improve the energy efficiency and stability. First, a wheel-legged quadruped robot, R-Taichi, is developed, which is capable of switching to legged, wheeled, and RHex mobile modes. Second, the mechanical structure of the transformable two-degree-of-freedom leg is introduced, and the kinematics is analyzed. Finally, experiments are conducted to generate wheeled, legged, and RHex motion in both swing and rolling gaits, and the energy efficiency is further compared. The experimental results show that the rolling motion can ensure stable ground contact and mitigate cyclic collisions, reducing specific resistance by up to 30% compared with conventional swing gaits, achieving a top speed of 0.7 m/s with enhanced stability (root mean square error (RMSE) reduction of 22% over RHex mode). Furthermore, R-Taichi exhibits robust multi-terrain adaptability, successfully traversing gravel, grass, and obstacles up to 150 mm in height.

Deep learning-based object detection algorithms have recently made significant strides in analyzing natural scenes. However, they still face challenges in remote sensing imagery, particularly due to the large variation in object sizes, the prevalence of small objects, and the similarity between different classes. To address these issues, we introduce MLA-YOLO (Multi-Attention based YOLO), a novel network designed to enhance object detection accuracy in remote sensing images. Our approach incorporates a multi-attention mechanism into the YOLO framework, significantly improving the detection of small objects. Furthermore, we integrate a CA (Channel Attention) mechanism into the feature fusion network, bolstering its feature fusion capabilities. We evaluate MLA-YOLO using the DIOR and RSOD datasets, where it demonstrates superior performance with a mean Average Precision (mAP) of 86.9% on the DIOR dataset and 93.93% on the RSOD dataset. Notably, it achieves a mAP of 94.68% on the NWPU VHR-10 dataset, outperforming existing algorithms. MLA-YOLO shows particular effectiveness in detecting small and multi-scale objects in remote sensing images, marking a significant advancement in this challenging area.