eduzhai > Applied Sciences > Engineering >

Generalization Guarantees for Imitation Learning

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 18 pages

Abstract: Control policies from imitation learning can often fail to generalize tonovel environments due to imperfect demonstrations or the inability ofimitation learning algorithms to accurately infer the expert s policies. Inthis paper, we present rigorous generalization guarantees for imitationlearning by leveraging the Probably Approximately Correct (PAC)-Bayes frameworkto provide upper bounds on the expected cost of policies in novel environments.We propose a two-stage training method where a latent policy distribution isfirst embedded with multi-modal expert behavior using a conditional variationalautoencoder, and then "fine-tuned " in new training environments to explicitlyoptimize the generalization bound. We demonstrate strong generalization boundsand their tightness relative to empirical performance in simulation for (i)grasping diverse mugs, (ii) planar pushing with visual feedback, and (iii)vision-based indoor navigation, as well as through hardware experiments for thetwo manipulation tasks.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...