A Modular Vision System for Practical Object Detection on Resource-Constrained Humanoid Robots

Deploying modern deep learning-based vision systems on humanoid robots remains challenging due to limited onboard computational resources and legacy software constraints. This paper presents a modular vision system for practical object detection on resource-constrained humanoid platforms, based on the YOLOv9 framework. The proposed architecture adopts a dual-environment design, decoupling the perception pipeline from the robot control system to enable compatibility between modern deep learning libraries and a ROS-based platform. To support efficient deployment, task-specific lightweight models are trained and integrated into a modular pipeline optimized for CPU-only inference. The system is evaluated across multiple task scenarios derived from the FIRA RoboWorld Cup (Hurocup) competition, including Marathon, Basketball, and Archery. Performance is assessed in terms of detection accuracy and computational efficiency, demonstrating that reliable perception can be achieved at 4-8 FPS under constrained hardware conditions. The results show that the proposed approach improves robustness compared to traditional geometric vision methods, particularly in dynamic and visually complex environments, while maintaining practical responsive task-level perception for robotic decision-making. The work highlights the trade-offs between accuracy, computational cost, and system responsiveness and demonstrates the feasibility of deploying modern object detection models on embedded humanoid platforms.