It is difficult to detect grasping position of irregular objects automatically which always leads to the failure of grasping task in the current arm robot application scenario. This paper proposes an adaptive grasping position detection method that uses the modified YOLO to recognize and detect contour of an image of the object which sampled by a camera. The detection contour results are then fed into a special regression network trained by prepared datasets to detect the grasping position of the object. The experimental results have shown that the grasping position detection method has good robustness and environmental adaptability and can meet the most needs of general applications in practical working condition.