I worked on a personal project over the last two weeks to teach myself reinforcement learning. Here’s a write-up on my work - https://robosamir.github.io/DDPG-on-a-Real-Robot/
Very interesting work!
I did not really go into the code, so these questions might be trivial:
- What part of the MoveIt! framework do you still use? You write that you bypass it by interacting with joint states directly. Could your algorithm be an alternative for the RRT algorithm, or one of the other planning algorithms?
- Would it be easy to extend this to obstacle avoidance as well?
In the published iteration, I don’t use MoveIt! at all. Details of why I phased it out are in the blog.
It’s not my algorithm (i wish it was !); but yes it could be an alternative to RRT.
Obstacle avoidance will be somewhat involved because the state has to incorporate information about occupied voxels and the reward function has to be modified to include a cost for distance to collision objects.
Nice!! I shall try and implement it myself,… somehow.
Was a bit of a surprise to see an old ROSAnswers post when I read through the article ;). Never did get the Fanuc running, much to the benefit of health and safety.
Nice work! Something similar your stuff is also implemented by OpenAI for connecting OpenAI Gym and ROS. You can find ye package here: https://github.com/openai/rosbridge