eduzhai > Applied Sciences > Engineering >

Heteroscedastic Uncertainty for Robust Generative Latent Dynamics

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: Learning or identifying dynamics from a sequence of high-dimensionalobservations is a difficult challenge in many domains, including reinforcementlearning and control. The problem has recently been studied from a generativeperspective through latent dynamics: high-dimensional observations are embeddedinto a lower-dimensional space in which the dynamics can be learned. Despitesome successes, latent dynamics models have not yet been applied to real-worldrobotic systems where learned representations must be robust to a variety ofperceptual confounds and noise sources not seen during training. In this paper,we present a method to jointly learn a latent state representation and theassociated dynamics that is amenable for long-term planning and closed-loopcontrol under perceptually difficult conditions. As our main contribution, wedescribe how our representation is able to capture a notion of heteroscedasticor input-specific uncertainty at test time by detecting novel orout-of-distribution (OOD) inputs. We present results from prediction andcontrol experiments on two image-based tasks: a simulated pendulum balancingtask and a real-world robotic manipulator reaching task. We demonstrate thatour model produces significantly more accurate predictions and exhibitsimproved control performance, compared to a model that assumes homoscedasticuncertainty only, in the presence of varying degrees of input degradation.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×