SLAM

[diemea]

Hi!

So far I have been able to complete the setup of PyCuVSLAM using Docker. Inside the container I can run the EuRoC example (monocular) with success. The viewer windows appear and there is "a line" drawed at the world view.

However, my goal is to create a map via SLAM (like here). Would it be possible to get some assistance in adjusting the track_*.py to run Visual SLAM for a monocular camera (not inertial) or could someone provide a sample .py for monocular SLAM with a given dataset?

Also, I am very interested in using PyCuVSLAM monocular SLAM together with ROS2. How would I create a .py tracker for grapping the camera input from a ROS2 topic (e. g. /camera/image_raw or /camera/image_raw/compressed) and run SLAM with that saving the map afterwards?

Many thanks in advance for your help!

PS: One additional question: When I am running the track_euroc example, I get the following warning:

Running on a GPU/graphics driver with very limited abilitites. Consider updating your driver.

I am using a RTX 5060 with this stats:

NVIDIA-SMI 590.48.01 Driver Version: 590.48.01 CUDA Version: 13.1

Is this because of the weak GPU or is something wrong with the driver/cuda version?

[zwdoescode]

@diemea What kind of map do you want to create? Have you checked out this tutorial of creating map with isaac_mapping_ros? It creates three different kinds of map: cuVSLAM Map, Visual Global Localization Map and Occupancy Grid Map.

The "line" you observed in the viewer windows is the trajectory of camera poses.

Please check the system requirements for PyCuVSLAM https://github.com/NVlabs/PyCuVSLAM?tab=readme-ov-file#system-requirements-and-setup. It only supports CUDA 12.6 for this released version. But you have CUDA version 13.1.

[diemea]

@zwdoescode Thank you for your response. The maps created by isaac_mapping_ros seem to be what I'm looking for. I have two questions regarding this:

1. For the map creation pre-recorded rosbags are used. Do you know, if it's also possible to use live cam images instead?
2. Does this make use of PyCuVSLAM in any way or is it a separate thing?

For the CUDA version, I could downgrade to 12.6 if this would be helpful.

[zwdoescode]

You will probably find better support about this in its github page or NVIDIA Isaac ROS developer forum.

Since I haven't tried isaac_mapping_ros myself, I will try my best to answer you:

1. AFAIK, the algorithm for creating Visual Global Localization Map cannot be run online. Theoretically Occupancy Grid Map should be able to create online. But you probably need to write your own script.
2. Isaac ROS is a separate thing. Isaac ROS provides ROS2 wrapper for various NVIDIA libraries. PyCuVSLAM is a python wrapper for cuVSLAM.