search

NVIDIA-ISAAC-ROS / isaac_ros_visual_slam

Visual SLAM/odometry package based on NVIDIA-accelerated cuVSLAM
https://developer.nvidia.com/isaac-ros-gems
Apache License 2.0
819 stars 127 forks source link

Is this sufficiently robust for long-term and/or large scale use? #58

Open SteveMacenski opened 1 year ago

SteveMacenski commented 1 year ago

If so, I'd be happy to have some documentation about working with Nav2 + VSLAM in our tutorials: https://navigation.ros.org/tutorials/index.html. However, its important that its suitable for Nav2 users for practical applications!

hemalshahNV commented 1 year ago

We're racking up mileage on Isaac ROS Visual SLAM but if there is feedback on its performance in large-scale use for long-terms, it would be greatly appreciated. It should be suitable for Nav2 users for practical applications but I'd rather have Nav2 users confirm this of course.

SteveMacenski commented 1 year ago

I don't have any in particular to share, but would love to hear as well as comparison of this to other methods. I'd be happy to have it as a tutorial or integration into ROS 2 as an alternative option for SLAM/localization using vision if its sufficiently good for practical use. Having a vision option to compliment the 2D-laser scanner methods could be a good synergy of collaboration between ROS-Nvidia.

Does this have a localization mode from a previous map and/or continue a previous session?

hemalshahNV commented 1 year ago

Yes, Isaac ROS Visual SLAM does have multi-session mapping (save a map, restore, continue mapping). It would be great to work out a series of benchmarks together to qualify both navigation behavior and subsystems including localization, floor plan mapping, trajectory generation, etc. if there isn't already one you would recommend (continuing the discussion from a ROS Discourse thread).

SteveMacenski commented 1 year ago

How about a pure localization mode? Or does the continued mapping mode cull out old features so it doesn't scale infinitely?

Its also good to make sure benchmarks aren't just creating the "11th competing standard". I'd think some work that previously exists in literature may very well be suitable.

hemalshahNV commented 1 year ago

Yes, it does cull features associated with landmarks visible in the frustum of the camera in recent frames but failed to be matched in the current frame.

Elbrus is state-of-the-art quality on the KITTI benchmark (here) and compares well feature-wise and engineering-wise to the other solutions you reviewed in your paper (here).

We're looking to integrate Isaac ROS Visual SLAM as more than just an odometry source into Nav2 at present. Any guidance or advice is appreciated as always.