SO3-CNN: Learning Rigid Displacement using Depth Images and Orthogonal Dual Tensors

Learning by demonstration, human-to-robot skills transfer or visual perception and control are tasks that can be tackled with techniques that combine machine vision and deep learning. In this paper, we take a step further from the classic deep learning based on the special Euclidean group (SE3) parameterization of displacement and motion. Considering an isomorphism between SE3 and the orthogonal dual tensor group, we propose an architecture that is set to improve the accuracy of computing the motion parameters. Thus, we introduce orthogonal dual tensors - convolutional neural network (${\widehat {{\mathbf{S}}\mathbb{O}}_3}$-CNN), a deep neural network set to learn from depth data how to extract semantic information regarding the layout of crobot’s segments while recovering the rigid displacement for each segment. To validate the design of ${\widehat {{\mathbf{S}}\mathbb{O}}_3}$-CNN, we employ a state-of-the-art database. The experimental results show that rigid displacements are estimated with higher accuracy by ${\widehat {{\mathbf{S}}\mathbb{O}}_3}$-CNN in comparison with a state-of-the-art SE3-CNN.