Equivariant Deep Dynamical Model for Motion Prediction

• DOI: 10.48550/arxiv.2111.01892
• Publication Date: 2021-01-01
• Authors: Bahar Azari; Deniz Erdoğmuş
• Keywords: Equivariant map; Representation; Generative model

Learning representations through deep generative modeling is a powerful approach for dynamical to discover the most simplified and compressed underlying description of data, then use it other tasks such as prediction. Most learning have intrinsic symmetries, i.e., input transformations leave output unchanged, or undergoes similar transformation. The process is, however, usually uninformed these symmetries. Therefore, learned individually transformed inputs may not be meaningfully related. In this paper, we propose an SO(3) equivariant model (EqDDM) motion prediction that learns structured representation space in sense embedding varies with symmetry transformations. EqDDM equipped networks parameterize state-space emission transition models. We demonstrate superior predictive performance proposed on various data.