0
$\begingroup$

Rosanswers logo

Hey,

I am right now trying to implement some different reinforcement learning algorithms to my Turtlebot for navigational tasks. In preparation for transferring those after training to the real turtlebot, I am right now trying to make the localization "more real". Due to friction and slip effects, the robots simulated odometry data is not suitable to deliver "realistic" cooardinates for planning the path. Therefore I am looking for ways to make the simulation more realistic. I already have some thoughts about it:

  1. I found out about the libgazebo_ros_diff_drive.so plugin which is used in simulation as a controller. There I can switch

    `<odometrySource>world</odometrySource>`
    

to <odometrySource>encoder</odometrySource>

and I will get data from the encoder integration instead of from the gazebo world (and no perfect coordinates). Thus I hope I can get effects like slipping better into the simulation

  1. Right now I am also thinking about using tf as well. If I have a look in RVIZ, I can see that the locations of the cooardinate systems of the Map and Odom start to differ with time. Here it might be useful to have a look at the rqt_tf_tree which has the followed structure (with encoder setting):

map => odom => base_footprint => base_link => wheel_left_link, wheel_right_link, lmu_link, base_scan, center_back_link

Would it here be sufficient to connect the both map and base_link directly in order to get a higher reality factor?

Would be great if you may also have some ideas of how to make the localization of the robot more realistic.

Kind regards Marcel


Originally posted by Tima1995 on ROS Answers with karma: 17 on 2019-08-01

Post score: 0

$\endgroup$

1 Answer 1

0
$\begingroup$

Rosanswers logo

I might recommend modifying the plugin for some better noise/slip modelling. That publishes the real position of the robot, maybe add some noise in in accordance with the noise you see on your encoders. Modelling accurate weight/inertial/acceleration/coef. of friction of materials of the robot and environment will be critical for gazebo to do its job and give you reasonable slipping. With those 2 things you can probably get something pseudo-realistic, but I sort of doubt the slipping will be consistent since I don't believe anything in gazebo models multi-point-of-contact tires effectively.

http://docs.ros.org/jade/api/gazebo_plugins/html/gazebo__ros__diff__drive_8cpp_source.html


Originally posted by stevemacenski with karma: 8272 on 2019-08-01

This answer was ACCEPTED on the original site

Post score: 2


Original comments

Comment by Tima1995 on 2019-08-02:
Thanks for your answer!. Yes that is definetly something I will try to make better. But still I have a problem when getting the localization data. If I get them from Gazebo, they are perfect and not to compare with those which are collected by the odometry of the real Turtlebot. Thats why I am thinking about switching localization from Odometry to something like the AMCL package to get data from the scan data.

Comment by stevemacenski on 2019-08-02:
Yes, you can definitely do that. Localization is one of those fields I tend to not test in simulation because as you mention, its hard to model it well, but robots (at least in my experience) are plentiful for testing data. But obviously doesn't work for your online training.

Best of luck!

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.