Graspable Object Classification with Multi-loss Hierarchical Representations

To allow robots to accomplish manipulation work effectively, one of the critical functions they need is to precisely and robustly recognize the robotic graspable object and the category of the graspable objects, especially in data limited condition. In this paper, we propose a novel multi-loss hierarchical representations learning framework that is capable of recognizing the category of graspable objects in a coarse-to-fine way. Our model consists of two main components, an efficient hierarchical feature learning component that combines kernel features with the deep learning features and a multi-loss function that optimizes the multi-task learning mechanism in a coarse-to-fine way. We demonstrate the power of our proposed system to data of graspable and ungraspable objects. The results show that our system has superior performance than many existing algorithms both in terms of classification accuracy and computation efficiency. Moreover, our system achieves a quite high accuracy (about 82%) in unstructured real-world condition.