Vision Perception Optimization and Adaptive Control for Resource-Constrained Platform: A Ping-Pong Ball Pickup & Place System

Abstract

This thesis presents the design, implementation, and experimental evaluation of an autonomous ping-pong ball collection robot system. The system integrates computer vision, object detection, visual servoing, and robotic manipulation to create a fully autonomous solution capable of identifying, collecting, and organizing ping-pong balls in dynamic environments. Using a mecanum-wheeled mobile platform equipped with a 6-DOF robotic arm and multiple cameras, a comprehensive system architecture was developed based on ROS (Robot Operating System) with a state machine design to orchestrate complex task sequences. The system demonstrates the ability to successfully detect, approach, and collect ping-pong balls while navigating in varied environments.

System performance was evaluated through comprehensive experiments, including YOLO v12 Nano model benchmarking, ablation studies of optimization techniques, spiral search strategy validation, and visual servoing performance analysis. The YOLO model evaluation demonstrates that a properly optimized Nano variant achieves 98.9% mAP@0.5 while maintaining sufficient inference speed on resource-constrained hardware. Ablation studies reveal that combining TensorRT with FP16 precision yields a 592.6% performance improvement with negligible accuracy loss. The spiral search strategy demonstrates effective target recovery capabilities when objects temporarily leave the field of view. Most notably, visual servoing experiments demonstrate that lower frame rates (15 fps) offer substantial advantages by enabling multi-camera operations and providing greater system functionality compared to higher frame rates (30 fps), despite conventional expectations.

This thesis presents a fully integrated robotic system tailored specifically for resource-constrained embedded platforms. The research introduces a refined YOLO-based vision pipeline, an adaptive spiral search method validated under diverse conditions, and a visual servo control strategy achieving millimeter-level accuracy. These contributions offer practical guidance for effectively deploying advanced robotic systems where computational resources are limited.