Physics-penalised Regularisation for Learning Dynamics Models with Contact

Gabriella Pizzuto, Michael Mistry

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Robotic systems, such as legged robots and manipulators, often handle states which involve ground impact or interaction with objects present in their surroundings; both of which are physically driven by contact. Dynamics model learning tends to focus on continuous motion, yielding poor results when deployed on real systems exposed to non-smooth frictional discontinuities. Inspired by a recent promising direction in machine learning, in this work we present a novel method for learning dynamics models undergoing contact by augmenting data-driven deep models with physics-penalised regularisation. Precisely, this paper conceptually formalises a novel framework for using an impenetrability component in the physics-based loss function directly within the learning objective of neural networks. Our results demonstrate that our method shows superior performance to using normal deep models for learning non-smooth dynamics models of robotic manipulators, strengthening their potential for deployment in contact-rich environments.
Original languageEnglish
Title of host publicationProceedings of the 3rd Conference on Learning for Dynamics and Control
EditorsAli Jadbabaie, John Lygeros, George J. Pappas, Pablo A. Parrilo, Benjamin Recht, Claire J. Tomlin, Melanie N. Zeilinger
PublisherPMLR
Pages611-622
Number of pages12
Publication statusPublished - 7 Jun 2021
Event3rd Annual Learning for Dynamics & Control Conference - Online
Duration: 7 Jun 20218 Jun 2021
https://l4dc.ethz.ch/

Publication series

NameProceedings of Machine Learning Research
PublisherPMLR
Volume144
ISSN (Print)2640-3498

Conference

Conference3rd Annual Learning for Dynamics & Control Conference
Abbreviated titleL4DC 2021
Period7/06/218/06/21
Internet address

Keywords

  • Dynamics Model Learning
  • Physics-guided Learning

Fingerprint

Dive into the research topics of 'Physics-penalised Regularisation for Learning Dynamics Models with Contact'. Together they form a unique fingerprint.

Cite this