GMapping Application

GMapping is a map generating tool that uses the OpenSlam software library. The application allows you to create maps to use in other applications.

The GMapping application uses the LIDAR capabilities of the Carter reference robot.

Note

Mapping is a computation-intensive and storage-intensive activity that may require fine tuning to generate an appropriate building map. For best results, record your mapping logs and tune the GMapping parameters offline.

Leverage odometry or the pose tree for the robot pose, depending on the capabilities of your robot. Use an inertial measurement unit (IMU) to improve results.

While you can run the GMapping application on a desktop, for actual mapping it must be deployed and run on a Carter robot.

  1. Deploy //apps/carter/gmapping:gmapping-pkg to the robot as explained in Application Console Options.

  2. Run the GMapping application while selecting the target robot profile:

    Copy
    Copied!
                

    bob@carter1:~/gmapping-pkg$ ./apps/carter/gmapping/gmapping --more "apps/carter/robots/carter_1.json"

Output

The sample application writes the maps into the /tmp/map.img_<N>.png folder throughout the run. Specify the output path with the following configuration parameter for the Gmapping codelet:

Copy
Copied!
            

"config": { "gmapping.gmapping": { "GMapping": { "file_path": "/tmp" } } }


The sequence of images below illustrates the mapping process over time, from the first LIDAR data capture to a map of a building.

map1.png

map2.png

map3.png

map4.png

map5.png

map6.png

map7.png

The speed at which the robot moves during mapping has an impact on results. The slower the speed, the higher number of LIDAR samples, resulting in increased accuracy. Avoid sharp turns. Configure the robot to limit the maximum linear and angular speeds.

Match and close path loops regularly to correct drifts and errors in odometry and inertial measurements during mapping. The matching depth is finite. Where possible, navigate in circles around blocks in the building, such as cubicle corrals and large architectural elements. There is no need to drive through an already mapped area; it increases noise.

Maintain enough anchor points from frame to frame. This is especially important when exiting or entering new areas or turning into a hallway. Avoid driving too close to walls. Select a quantity of (matching) particles that is as high as your building topology allows you to without losing scan matching or reverting to odometry only, which results in poor mapping results.

Use a long enough range to help maintain anchor points but use a small update range to draw a sharp map image. Ideally, your update range should be no larger than half the length of the largest area to map.

Record your scan and odometry channels to replay and create new maps with different parameters after the real world capture is completed with the log mapping applications. Tune the configuration parameters to experiment for your mapping use case.

When working with maps generated by GMapping or logamppings, trim the gray edges of the map image. This reduces the size of the file and improves the performance of the algorithms using the map image. After modification, save these maps as grayscale, compressed, in PNG format, for significant reduction in file size.

© Copyright 2018-2020, NVIDIA Corporation. Last updated on Oct 30, 2023.