MoveIt! - motion planning with Reinforcement Learning

asked 2020-02-06 08:31:20 -0500

MPeregrini gravatar image

Hi everyone,

I'm quite new to ROS and Gazebo. I'm currently working on a long-term project which plans on training robots with reinforcement learning in Gazebo ; the goal of these robots is to learn how to intervene on complex industrial plants in case of crisis thanks to simulation inspired from real incidents and accidents. The next step is to provide to the real robots this knowledge coming from simulation. These robots consist in a mobile base with a 5 DOF arm mounted on top ; various tools and cameras are to be equipped as end effectors. For the moment, I'm focusing on the management of only one robot at a time, but I will deal with fleet management sooner or later.

In this context, I wanted gather opinions about the relevance of coupling MoveIt! (or OMPL with MoveIt!, or any motion planning tool) with Reinforcement Learning, or if it was worth training simulated robots without any motion planning tool. In the case the answer is positive, are there any open source works available on the Internet ?

Thanks a lot!

Configuration : Ubuntu 18.04, Gazebo 9, ROS Melodic

edit retag flag offensive close merge delete


Since this is a discussion and not a technical question, I think this would live better on

fvd gravatar image fvd  ( 2020-02-26 20:16:19 -0500 )edit