Learning Contact Dynamics using Physically Structured Neural Networks

Learning physically structured representations of dynamical systems that include contact between different objects is an important problem for learning-based approaches in robotics. Black-box neural networks can learn to approximately represent discontinuous dynamics, but they typically require large quantities of data and often suffer from pathological behaviour when forecasting for longer time horizons. In this work, we use connections between deep neural networks and differential equations to design a family of deep network architectures for representing contact dynamics between objects. We show that these networks can learn discontinuous contact events in a data-efficient manner from noisy observations in settings that are traditionally difficult for black-box approaches and recent physics inspired neural networks. Our results indicate that an idealised form of touch feedback—which is heavily relied upon by biological systems—is a key component of making this learning problem tractable. Together with the inductive biases introduced through the network architectures, our techniques enable accurate learning of contact dynamics from observations. Proceedings of the 24 International Conference on Artificial Intelligence and Statistics (AISTATS) 2021, San Diego, California, USA. PMLR: Volume 130. Copyright 2021 by the author(s).

[1]  E Weinan,et al.  A Proposal on Machine Learning via Dynamical Systems , 2017, Communications in Mathematics and Statistics.

[2]  Anthony Gravouil,et al.  A new heterogeneous asynchronous explicit–implicit time integrator for nonsmooth dynamics , 2017 .

[3]  Jan Peters,et al.  Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning , 2019, ICLR.

[4]  Jerrold E. Marsden,et al.  Nonsmooth Lagrangian Mechanics and Variational Collision Integrators , 2003, SIAM J. Appl. Dyn. Syst..

[5]  David Duvenaud,et al.  Neural Ordinary Differential Equations , 2018, NeurIPS.

[6]  J. Moreau,et al.  Unilateral Contact and Dry Friction in Finite Freedom Dynamics , 1988 .

[7]  P. Panagiotopoulos,et al.  New developments in contact problems , 1999 .

[8]  J. Marsden,et al.  Variational integrators for constrained dynamical systems , 2008 .

[9]  Joshua B. Tenenbaum,et al.  End-to-End Differentiable Physics for Learning and Control , 2018, NeurIPS.

[10]  Carl E. Rasmussen,et al.  Gaussian Processes for Data-Efficient Learning in Robotics and Control , 2015, IEEE Transactions on Pattern Analysis and Machine Intelligence.

[11]  Jason Yosinski,et al.  Hamiltonian Neural Networks , 2019, NeurIPS.

[12]  Matthew West,et al.  Decomposition contact response (DCR) for explicit finite element dynamics , 2005, International Journal for Numerical Methods in Engineering.

[13]  J. Marsden,et al.  Discrete mechanics and variational integrators , 2001, Acta Numerica.

[14]  J. Nathan Kutz,et al.  Deep learning in fluid dynamics , 2017, Journal of Fluid Mechanics.

[15]  Victor M. Martinez Alvarez,et al.  DyNODE: Neural Ordinary Differential Equations for Dynamics Modeling in Continuous Control , 2020, ArXiv.

[16]  Miles Cranmer,et al.  Lagrangian Neural Networks , 2020, ICLR 2020.

[17]  Alexander Herzog,et al.  Learning a Structured Neural Network Policy for a Hopping Task , 2017, IEEE Robotics and Automation Letters.

[18]  Michel Saint Jean,et al.  The non-smooth contact dynamics method , 1999 .

[19]  Eldad Haber,et al.  Stable architectures for deep neural networks , 2017, ArXiv.

[20]  Hsiu-Chin Lin,et al.  Kinematics-based estimation of contact constraints using only proprioception , 2016, 2016 IEEE-RAS 16th International Conference on Humanoid Robots (Humanoids).

[21]  Nicholas Rotella,et al.  Unsupervised Contact Learning for Humanoid Estimation and Control , 2018, 2018 IEEE International Conference on Robotics and Automation (ICRA).

[22]  Marc Peter Deisenroth,et al.  Variational Integrator Networks for Physically Structured Embeddings , 2020, AISTATS.

[23]  J. M. Sanz-Serna,et al.  Symplectic integrators for Hamiltonian problems: an overview , 1992, Acta Numerica.

[24]  Maziar Raissi,et al.  Deep Hidden Physics Models: Deep Learning of Nonlinear Partial Differential Equations , 2018, J. Mach. Learn. Res..

[25]  David Dureisseix,et al.  Benchmark cases for robust explicit time integrators in non-smooth transient dynamics , 2019, Advanced Modeling and Simulation in Engineering Sciences.