YOLOv5s-gse——基于YOLOv5s-gse的目标检测和机械臂吸附研究

YOLOv5s-gse—Study on Object Detection and Robotic Arm Suction Based on YOLOv5s-gse

摘要: 针对传统目标检测算法在杂乱工业环境中存在的识别精度不足和定位不稳定等问题，本文提出了一种面向机械臂抓取任务的轻量化改进模型YOLOv5s-gse，以提升机械臂在复杂场景下的目标感知与操作能力。该模型以YOLOv5s为基础，在主干网络中引入由轻量卷积GSConv、Shuffle通道重排以及ECA注意力机制构成的C3GSConv模块，以增强特征提取能力；同时降低模型参数量和计算开销。在检测完成后，通过外参标定实现相机与机械臂坐标的转换，并基于逆运动学求解各关节角度，构建Episode1机械臂抓取实验平台进行验证。实验结果表明，YOLOv5s-gse模型在保持较高前向推理速度的同时，其mAP较原YOLOv5s提升0.03%，漏检率与误检率均略有降低；结合机械臂的吸取实验成功率极高，充分验证了所提方法在机器人抓取场景中的有效性。该方案兼具检测精度与轻量化优势，成功将多种轻量化技术整合并应用于一个功能完备的机器人抓取系统。

Abstract: To address the limitations of traditional object detection algorithms in cluttered industrial environments—specifically insufficient recognition accuracy and unstable localization—this paper proposes a lightweight improved model, YOLOv5s-gse, designed for robotic arm grasping tasks. Building upon YOLOv5s, the model integrates a C3GSConv module into the backbone network, which is composed of lightweight GSConv, Shuffle-based channel reordering, and an ECA attention mechanism, thereby enhancing feature extraction while reducing model parameters and computational cost. After object detection, extrinsic calibration is performed to map coordinates between the camera and the robotic arm, and inverse kinematics is applied to compute joint angles. An Episode1 robotic grasping experimental platform is constructed for evaluation. Experimental results show that YOLOv5s-gse achieves higher mAP—an improvement of 0.03% over the original YOLOv5s—while maintaining high inference speed, with slight reductions in both missed detections and false detections. When combined with the robotic arm’s suction-based grasping mechanism, the system demonstrates a very high success rate, validating the effectiveness of the proposed method in robotic grasping scenarios. The approach offers both accuracy and lightweight advantages, successfully integrating multiple lightweight techniques into a fully functional robotic grasping system.