Robot grasping based on object shape approximation and LightGBM

Object grasp planning is a challenging task. Recently, methods based on deep learning have made great progress in this area, but they are highly dependent on datasets, which means that they may encounter some difficulties when facing novel objects that are not available in the datasets. In this paper, a novel method is proposed to generate grasping candidate rectangles for object based on shape approximation, without datasets or shape priori of objects. Specifically, combining K-means and the minimum oriented bounding box algorithm for point sets, an adaptive K-means algorithm is applied for decomposing objects into multiple rectangles. The algorithm can independently select the number of K-means cores and automatically select the number of rectangles used to approximate the shape of the object. According to the parameters of each rectangle, a candidate grasping rectangle of the object is generated. In addition, using the Cornell grasping dataset, a LightGBM classifier is trained for the classification and evaluation of object candidate grasping rectangles. Experimental results show that our classification accuracy rate has reached 94.5% and the detection time is only 0.0003s. Among the candidate rectangles, the one with the highest score in the LightGBM model would be selected for real robot grasping. Finally, a multi-object grasping experiment conducted on a real robot platform shows that our algorithm can help the robot grasp new objects with an average success rate of 91.81%.