Fast and accurate identification and localization of fruits in complex natural environments is the key and difficult point of kiwifruit robotic arm picking.For kiwifruit real-time picking applications,deep learning-based kiwifruit classification and recognition algorithms,spatial localization model construction and robotic arm picking experimental research were carried out.The one-stage YOLOX algorithm model of Anchor_Free is preferred,and the feature fusion link of this model is improved by introducing an attention mechanism,adjusting the number of channels,increasing feature fusion,and stacking feature fusion,which improves the accuracy of classifying and recognizing kiwifruit fruits in unstructured environments.The experimental data show that the AP of the improved YOLOX algorithm for no occluded,leaf occluded,branch occluded and fruit occluded is improved to 99.38%,98.43%,93.82%and 92.98%, respectively,the mAP reaches 96.15%,its average F,value is improved to 94.24%,and its average detection time is 0.019 s/space at 640×640 resolution.Further,the three-dimensional spatial position of the recognized target is solved using the image depth information and verified by physical experiments,in which the fruit localization coordinate error is ≤3.3%.The algorithm achieves superior overall performance in terms of recognition rate,detection time and localization error to meet the real-time fruit recognition and localization requirements in kiwifruit robotic arm picking.