Improving Out-of-Distribution Generalization of Learned Dynamics by Learning Pseudometrics and Constraint Manifolds | IEEE Conference Publication | IEEE Xplore