We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: NewtonianVAE: Proportional Control and Goal Identification from Pixels via Physical Latent Spaces

Abstract: Learning low-dimensional latent state space dynamics models has been a powerful paradigm for enabling vision-based planning and learning for control. We introduce a latent dynamics learning framework that is uniquely designed to induce proportional controlability in the latent space, thus enabling the use of much simpler controllers than prior work. We show that our learned dynamics model enables proportional control from pixels, dramatically simplifies and accelerates behavioural cloning of vision-based controllers, and provides interpretable goal discovery when applied to imitation learning of switching controllers from demonstration.
Subjects: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (stat.ML)
Cite as: arXiv:2006.01959 [cs.LG]
  (or arXiv:2006.01959v2 [cs.LG] for this version)

Submission history

From: Miguel Jaques [view email]
[v1] Tue, 2 Jun 2020 21:41:38 GMT (5819kb,D)
[v2] Mon, 26 Apr 2021 21:49:30 GMT (6206kb,D)

Link back to: arXiv, form interface, contact.