Skip to content

isaac_ros_visual_slam (cuVSLAM)

GPU-accelerated visual SLAM using stereo cameras. Estimates the robot's pose and builds a map of visual landmarks.

Uses a custom launch script at ~/workspaces/isaac_ros-dev/launch_cuvslam.sh:

docker exec -it isaac_ros_dev_container bash /workspaces/isaac_ros-dev/launch_cuvslam.sh
Launch script launch_cuvslam.sh (custom — NVIDIA's default launch file doesn't expose TF params)
Subscribes to /front_stereo_camera/left/image_rect_color, /right/image_rect_color, and their camera_info topics
Publishes /visual_slam/tracking/odometry, /visual_slam/vis/landmarks_cloud, /visual_slam/tracking/slam_path
Foxglove topics Add landmarks_cloud (decay=0), slam_path, and observations_cloud to the 3D panel

TF ownership and the coordinate system

The ROS TF chain is map → odom → base_link. On a real robot, SLAM owns this entire chain — there's no simulator giving ground-truth odometry.

In simulation, Isaac Sim and cuVSLAM both want to publish odom → base_link, but they compute different poses (Isaac Sim uses perfect physics, cuVSLAM uses visual feature tracking). If both publish, the lidar point cloud flickers between the two competing poses in Foxglove.

The fix has two parts:

  1. Isaac Sim side — the headless launch script (headless-sample-scene.sh) disables Isaac Sim's odom → base_link TF publisher via the OmniGraph API:

    import omni.graph.core as og
graph = og.get_graph_by_path("/World/Nova_Carter_ROS/transform_tree_odometry")
node = graph.get_node(".../ros2_publish_raw_transform_tree")
node.set_disabled(True)

  2. cuVSLAM side — the custom launch script sets publish_odom_to_base_tf: True and publish_map_to_odom_tf: True so cuVSLAM is the sole authority.

This matches real-robot behavior: cuVSLAM owns the full TF chain, and the lidar and SLAM landmarks render in the same coordinate system.

cuVSLAM TF params must be set at launch time

cuVSLAM caches publish_odom_to_base_tf and publish_map_to_odom_tf as C++ member variables at startup. Using ros2 param set at runtime has no effect — the node never re-reads them. Always set these in the launch file's parameters dict.

Stopping on-demand nodes

Killing a tmux session running docker exec does not kill the process inside the container — it only kills the exec client on the host. The node keeps running and will conflict if you launch another instance. Always stop container processes from inside:

docker exec isaac_ros_dev_container bash -c "pkill -f visual_slam"