Almost no movement using ROS Control and KUKA LWR 4+ sim in Gazebo

[Jack White]

Hello,

Hopefully there are still people listening out there!

I have written a ROS Control controller derived from RobotPlugin (first into a SingleArmPlugin, removing all the passive arm stuff, then a plugin for the KUKA arm, derived from the hardware interface. It took absolute ages to work it all out, but I'm at a point where I can send position commands and that works reasonably well.

I have copied and modified the PR2 agent to work with the single arm; in fact, I've done it twice, to be sure.

The trouble is with the trial commands. I have the Gazebo sim running. I have an experiment set up, based on pr2_example. It works sort of. I see the mean cost diminishing, but the robot moves barely a centimetre and the end effector just moves backwards and forwards over this distance. It's not obvious that it's actually moving in more than one joint. The samples are definitely very, very wrong, if the PR2 is anything to go by.

I am not really good enough with the underlying maths to fully comprehend GPS, or the experiment hyperparameters. This is really my first project in robotics.

I have played with the initial trajectory distribution stiffness and the LQG cost terms and been able to change the shape of the mean cost graph, but nothing has made the arm move in any manner apart from this gentle centimetre-scale swinging back and forth.

Has anyone else encountered this problem? Have you fixed it? Can anyone help, please?

Many thanks,

Jack White