We present Open3DSG the first approach for learning to predict open-vocabulary 3D scene graphs from 3D point clouds. The advantage of our method is that it can be queried and prompted for any instance ...