Abstract
Learning physically structured representations of dynamical systems thatinclude contact between different objects is an important problem forlearning-based approaches in robotics. Black-box neural networks can learn toapproximately represent discontinuous dynamics, but they typically requirelarge quantities of data and often suffer from pathological behaviour whenforecasting for longer time horizons. In this work, we use connections betweendeep neural networks and differential equations to design a family of deepnetwork architectures for representing contact dynamics between objects. Weshow that these networks can learn discontinuous contact events in adata-efficient manner from noisy observations in settings that aretraditionally difficult for black-box approaches and recent physics inspiredneural networks. Our results indicate that an idealised form of touch feedback-- which is heavily relied upon by biological systems -- is a key component ofmaking this learning problem tractable. Together with the inductive biasesintroduced through the network architectures, our techniques enable accuratelearning of contact dynamics from observations.