Training RL agents to solve robotics tasks still remains challenging due to the difficulty of exploration in purelycontinuous action spaces . We find that our simple change to the interface of the RL algorithm substantially improves both the learning efficiency and task performance irrespective of the underlying RL algorithm . We perform a thoroughempirical study across challenging tasks in three distinct domains with image input and a sparse terminal reward. We find it significantly outperforms prior methods which learn skills from offline expert data, which can be learned from offline data . Code and videos at https://mihdalal.io/raps/ and at http://www.g.com/rps.org/rp/mihdalal. The code is available to download and watch the videos at www.gtr.com.org.com

Author(s) : Murtaza Dalal, Deepak Pathak, Ruslan Salakhutdinov

Links : PDF - Abstract

Code :
Coursera

Keywords : rl - mihdalal - learning - algorithm - videos -

Leave a Reply

Your email address will not be published. Required fields are marked *