This study introduces the hybrid throwable reconnaissance robot and explores foundational research on its visual odometry and object detection capabilities. CFH-based hybrid throwable reconnaissance robot combines the advantages of wheeled and spherical robot designs, featuring a CFH structure for improved shock absorption. Given the robot's role in collecting critical information in military operations to ensure troop safety and survivability, it is imperative to equip it with technologies for accurate self-positioning and object detection within its environment. Therefore, we developed an embedded system prototype of CFH-based throwable reconnaissance robot to implement on-board visual odometry and object detection. A monocular visual odometry algorithm was implemented on the embedded system prototype, achieving 13-15 frames per seconds (FPS). Using transfer learning, we trained an object detection algorithm on seven classes relevant to the battlefield, applied to the prototype with 15-20 FPS and an average mAP@50 of 0.868. Through these implementations, the study demonstrates the potential of CFH-based throwable reconnaissance robots in enhancing military operations by providing movement estimation and efficient object detection capabilities, thereby contributing to the safety and effectiveness of military personnel in complex operational environments.
