eduzhai > Applied Sciences > Engineering >

Learning to Play by Imitating Humans

  • Save

... pages left unread,continue reading

Document pages: 7 pages

Abstract: Acquiring multiple skills has commonly involved collecting a large number ofexpert demonstrations per task or engineering custom reward functions. Recentlyit has been shown that it is possible to acquire a diverse set of skills byself-supervising control on top of human teleoperated play data. Play is richin state space coverage and a policy trained on this data can generalize tospecific tasks at test time outperforming policies trained on individual experttask demonstrations. In this work, we explore the question of whether robotscan learn to play to autonomously generate play data that can ultimatelyenhance performance. By training a behavioral cloning policy on a relativelysmall quantity of human play, we autonomously generate a large quantity ofcloned play data that can be used as additional training. We demonstrate that ageneral purpose goal-conditioned policy trained on this augmented datasetsubstantially outperforms one trained only with the original human data on 18difficult user-specified manipulation tasks in a simulated robotic tabletopenvironment. A video example of a robot imitating human play can be seen here:this https URL

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×