I was testing nvidia isaac vslam on zed2i with stereo + imu.
vSLAM algorithms tend to struggle in featureless plane environments and I was testing nvidia’s vslam in a similar environment. I noticed the global loop closure isn’t working or does not exist. I moved the camera for a quite a while in a known region but it still could not converge the global slam path.
I am curious, does this algorithm not support non temporal constraints because I do see you all use graph slam?
Hi @vishrut1
Thank you for your post. We just released the new Isaac ROS 3.0, and I suggest you try the new features.
The new Isaac ROS vslam now works with Jetpack 6.0
The Isaac ROS vslam has a global loop closure on the cvslam algorithm.
For details about the algorithm, there is a webinar made last year: Pinpoint, 250 fps, ROS 2 Localization with vSLAM on Jetson
Hi @Raffaello
I just updated the packages to the latest 3.0 release and the tracking seems to be deteriorated a bit especially in turns and loop closure still does not seem to be working. I tried it on the same bag and here are the results.
Hi @vishrut1
Your issue may be related to the ZED camera, which is not fully supported on Jetpack 6.0 at this time.
We are investigating with our engineers your issue and I keep you posted when there are further updates
Awesome, thanks a lot @Raffaello Please let me know if I can share any data to help.
Also curious, since zed2i is providing rectified stereo data + imu data + params(camera+imu), why would the algorithm behave differently in case of support/unsupported cameras.
I forward your issue to the engineering.
It would be helpful for us if you could share a rosbag file with us, as it will help us better understand the issue if your data is not confidential.
Quick update. The Isaac ROS 3.1 is released and restored the demo with the ZED camera
https://nvidia-isaac-ros.github.io/concepts/scene_reconstruction/nvblox/tutorials/tutorial_zed.html
