This paper combine several modeling ingredients (BNNs, ODEs, and VAEs) to produce a new family of models. It's not clear to my whether adding second-order dynamics in particular is advantageous over just adding extra latent dimensions to the state, which I think would be a generalization of the current approach. However, seeing a comparison against GPLVM-based models was nice, since these two approaches represent very different technical approaches to the same problem.