Setting up an environment for reinforced learning

edit

Hi, I am trying to set up an environment that will used to teach a gripper how to grasp different objects, I've been trying to do it using erlerobots gym-gazebo package, but I am failing at one thing- how to prepare the environment. I want to use DDPG algorithm with continuous action space for joint control. I guess that most of the environment would be similar to how gym-gazebo turtlebot example(i will also use cameras).

I can see in the gym documentation that for that one uses Box space, but:

1. What would be the shape?
2. I can't figure out how to send messages to that action space, since I think it can't be similar to Discrete ones that are used in tutrlebot example.

With my model I am publishing to JointTrajectory interface, so I can send it the velocities, efforts and positions, but that is probably not how the messages should be sent in continuous action space. Any tips how to proceed further with this?