Modelling rigid body dynamics with physics-informed neural networks

Functions describing physical systems of rigid bodies lie in a low-dimensional subspace of functions that neural networks can represent. If we could construct a NN architecture, that for any set of parameters yields dynamics of some rigid bodies’ system, then we could improve sample efficiency and generalization. We will review architectures that impose some physical plausibility including Neural ODEs, Lagrangian Neural Networks and the Koopman operator and discuss their pros and cons in modelling rigid bodies. We will also mention our effort to enable Lagrangian-based architectures to learn from positions and velocities only (i.e. without access to the accelerations) and how we can use these architectures in reinforcement learning.