We propose a model-free deep reinforcement learning method that leverages a
small amount of demonstration data to assist a reinforcement learning agent. We
apply this approach to robotic manipulation tasks and train end-to-end
visuomotor policies that map directly from RGB camera inputs to joint
velocities. We demonstrate that our approach can solve a wide variety of
visuomotor tasks, for which engineering a scripted controller would be
laborious. Our experiments indicate that our reinforcement and imitation agent
achieves significantly better performances than agents trained with
reinforcement learning or imitation learning alone. We also illustrate that
these policies, trained with large visual and dynamics variations, can achieve
preliminary successes in zero-shot sim2real transfer. A brief visual
description of this work can be viewed in https://youtu.be/EDl8SQUNjj0
Description
[1802.09564] Reinforcement and Imitation Learning for Diverse Visuomotor Skills
%0 Generic
%1 zhu2018reinforcement
%A Zhu, Yuke
%A Wang, Ziyu
%A Merel, Josh
%A Rusu, Andrei
%A Erez, Tom
%A Cabi, Serkan
%A Tunyasuvunakool, Saran
%A Kramár, János
%A Hadsell, Raia
%A de Freitas, Nando
%A Heess, Nicolas
%D 2018
%K 2018 arxiv deepmind imitation-learning reinforcement-learning robotics stanford
%T Reinforcement and Imitation Learning for Diverse Visuomotor Skills
%U http://arxiv.org/abs/1802.09564
%X We propose a model-free deep reinforcement learning method that leverages a
small amount of demonstration data to assist a reinforcement learning agent. We
apply this approach to robotic manipulation tasks and train end-to-end
visuomotor policies that map directly from RGB camera inputs to joint
velocities. We demonstrate that our approach can solve a wide variety of
visuomotor tasks, for which engineering a scripted controller would be
laborious. Our experiments indicate that our reinforcement and imitation agent
achieves significantly better performances than agents trained with
reinforcement learning or imitation learning alone. We also illustrate that
these policies, trained with large visual and dynamics variations, can achieve
preliminary successes in zero-shot sim2real transfer. A brief visual
description of this work can be viewed in https://youtu.be/EDl8SQUNjj0
@misc{zhu2018reinforcement,
abstract = {We propose a model-free deep reinforcement learning method that leverages a
small amount of demonstration data to assist a reinforcement learning agent. We
apply this approach to robotic manipulation tasks and train end-to-end
visuomotor policies that map directly from RGB camera inputs to joint
velocities. We demonstrate that our approach can solve a wide variety of
visuomotor tasks, for which engineering a scripted controller would be
laborious. Our experiments indicate that our reinforcement and imitation agent
achieves significantly better performances than agents trained with
reinforcement learning or imitation learning alone. We also illustrate that
these policies, trained with large visual and dynamics variations, can achieve
preliminary successes in zero-shot sim2real transfer. A brief visual
description of this work can be viewed in https://youtu.be/EDl8SQUNjj0},
added-at = {2018-04-23T16:43:56.000+0200},
author = {Zhu, Yuke and Wang, Ziyu and Merel, Josh and Rusu, Andrei and Erez, Tom and Cabi, Serkan and Tunyasuvunakool, Saran and Kramár, János and Hadsell, Raia and de Freitas, Nando and Heess, Nicolas},
biburl = {https://www.bibsonomy.org/bibtex/280f6de4f5a59599665dd707f674e42c1/achakraborty},
description = {[1802.09564] Reinforcement and Imitation Learning for Diverse Visuomotor Skills},
interhash = {0e749758bd091da3766d33e6799c22fc},
intrahash = {80f6de4f5a59599665dd707f674e42c1},
keywords = {2018 arxiv deepmind imitation-learning reinforcement-learning robotics stanford},
note = {cite arxiv:1802.09564Comment: 13 pages, 6 figures},
timestamp = {2018-04-23T16:43:56.000+0200},
title = {Reinforcement and Imitation Learning for Diverse Visuomotor Skills},
url = {http://arxiv.org/abs/1802.09564},
year = 2018
}