Tweeted By @svlevine
Deep RL, from images, conditioned on user-specified goals, trains in a few hours on real robots:https://t.co/317cmDyYph
— Sergey Levine (@svlevine) July 20, 2018
Key idea: let the robot "practice" goals that it proposes to itself to learn more about the world.
w/ @Vitchyr Ashvin Nair, M. Dalal, S. Bahl, S. Lin