Validating Isaac ROS Visual SLAM on a Robot

Overview

This document outlines a procedure that can be used to validate the successful installation and configuration of Isaac ROS Visual SLAM on real hardware.

Prerequisites

  • [ ] NVIDIA Jetson

  • [ ] Standalone computer with ROS data visualization tool (RViz, Foxglove, etc.)

  • [ ] Compatible stereo camera

  • [ ] Enough space to navigate (~40m loop)

  • [ ] Tape measure

Prepare Physical Environment

  1. Examine the designated testing area and identify a single, non-intersecting loop whose path length is approximately 40 meters. This is the path along which the camera is moved and tracked using Isaac ROS Visual SLAM.

  2. Define a starting pose along the identified loop. Label this pose with masking tape or a different marker.

    One possible approach is to mark an L-shape on the floor, aligning the camera’s position and orientation to the corner of the shape.

  3. Using a measuring tool, measure out a secondary pose along the loop that is exactly 1 meter along the forward axis from the starting pose. Label this pose in the same manner.

https://media.githubusercontent.com/media/NVIDIA-ISAAC-ROS/.github/main/resources/isaac_ros_docs/concepts/visual_slam/cuvslam/validation_tape.jpg/

Download and Configure Packages

  1. On each of the Jetson and standalone computers, set up your development environment by following the dev setup instructions.

    Note

    ${ISAAC_ROS_WS} is defined to point to /ssd/workspaces/isaac_ros-dev/ or ~/workspaces/isaac_ros-dev/.

  2. Set up your stereo camera and connect it to the Jetson, depending on your camera type:

  3. On both devices, clone isaac_ros_common and this repository under ${ISAAC_ROS_WS}/src.

    cd ${ISAAC_ROS_WS}/src

    git clone https://github.com/NVIDIA-ISAAC-ROS/isaac_ros_common.git

    git clone https://github.com/NVIDIA-ISAAC-ROS/isaac_ros_visual_slam.git

Build ROS Workspace

  1. On the Jetson, launch the Docker container in a new terminal:

    cd ${ISAAC_ROS_WS}/src/isaac_ros_common && ./scripts/run_dev.sh ${ISAAC_ROS_WS}

  2. Inside the container, install this package’s dependencies.

    sudo apt-get install -y ros-humble-isaac-ros-visual-slam

Select and Start ROS Visualization Tool

  1. Ensure that the standalone computer and Jetson device are on the same network.

  2. On the standalone computer, start RViz from the provided configuration file:

    Start the Docker container for RViz on the standalone computer:

    cd ${ISAAC_ROS_WS}/src/isaac_ros_common && ./scripts/run_dev.sh ${ISAAC_ROS_WS}

  3. (Optional) Configure the ROS_DOMAIN_ID environment variable on both devices to eliminate any cross-talk from other ROS nodes on the same network.

    In each terminal on each device:

    export ROS_DOMAIN_ID=5 # Any number between 1-101; use the same number for each terminal

    Note

    This environment variable must be set in each terminal, on both the Jetson and the standalone computer.

    From this terminal, run RViz2 to display the output:

    rviz2 -d src/isaac_ros_visual_slam/isaac_ros_visual_slam/rviz/default.cfg.rviz

  4. On the standalone computer, start logging the estimated transform:

    Attach another terminal to the running container for tf2_echo:

    cd ${ISAAC_ROS_WS}/src/isaac_ros_common && \
./scripts/run_dev.sh ${ISAAC_ROS_WS}

    From this terminal, run tf2_echo to display the transform between the /map frame and the appropriate camera frame:

    • HAWK: /camera

    • RealSense: /camera_link

    ros2 run tf2_ros tf2_echo map camera_link

Run Isaac ROS Visual SLAM and Monitor Results

Visual Odometry Test

The purpose of this test is to validate that the system is able to provide accurate measurements of the robot’s local motion.

  1. Initialize the camera at the marked starting pose.

  2. On the Jetson, run the Isaac ROS Visual SLAM launch file for your camera type:

    • HAWK:

      ros2 launch isaac_ros_visual_slam isaac_ros_visual_slam_hawk.launch.py

    • RealSense:

      ros2 launch isaac_ros_visual_slam isaac_ros_visual_slam_realsense.launch.py

    The /map, /odom, and camera frames must all be initialized to the starting pose.

  3. Carefully move the camera straight forward to the marked secondary pose.

  4. On the standalone computer and within the tf2_echo terminal, note the latest transform reported between /map and the appropriate camera frame.

    https://media.githubusercontent.com/media/NVIDIA-ISAAC-ROS/.github/main/resources/isaac_ros_docs/concepts/visual_slam/cuvslam/validation_test1_tf2.png/

    1. Validate orientation:

      Validate that the orientation error between the measured rotation unit quaternion and the expected rotation unit quaternion ([0 0 0 1]) is less than 15 degrees. This error can be calculated using the process explained here:

      \[q_{measured} = w_m + x_m i + y_m j + z_m k\]
      \[q_{expected} = w_e + x_e i + y_e j + z_e k\]
      \[2 \arccos(w_m w_e + x_m x_e + y_m y_e + z_m z_e) \leq 15 \degree\]

    2. Validate translation:

      Verify that the magnitude of the measured translation is within 5 centimeters of the expected translation (1 meter):

      \[0.95 \text{m} \leq \|t_{measured}\| \leq 1.05 \text{m}\]

If all real-world measurements correspond to the reported estimates within the expected tolerance, the Visual Odometry test is a success.

Loop Closure Test

The purpose of this test is to validate that the system is able to appropriately realign the estimated pose path to previously-seen features.

  1. Initialize the camera at the marked starting pose.

  2. On the Jetson, run the Isaac ROS Visual SLAM launch file for your camera type:

    • HAWK:

      ros2 launch isaac_ros_visual_slam isaac_ros_visual_slam_hawk.launch.py

    • RealSense:

      ros2 launch isaac_ros_visual_slam isaac_ros_visual_slam_realsense.launch.py

    Validate that the /map, /odom, and camera frames are all initialized to the starting pose.

  3. Carefully move the camera along the predetermined 40m path.

    Note

    To maximize the accuracy of this test, ensure that dynamic obstacles are removed from the robot’s field of view. If a nontrivial dynamic obstacle is introduced while the robot is performing the loop, you must restart this test.

  4. As the camera nears the completion of its first lap, move the camera in a small circle with the radius of ~0.5m around the origin of the starting pose. This step helps ensure that the previously-seen visual features are recognized and matched.

  5. Place the camera at rest, taking care to precisely align it with the floor marker indicating the starting pose.

  6. On the standalone computer and within the RViz window, ensure that a nontrivial number of visual features now appear red. Additionally, ensure that the camera frame appears nearly coincident with the /map frame.

    Note

    It is acceptable for the /odom frame to jump around significantly with respect to the /map frame. For the purposes of this test, only the transform between the /map and camera frames is relevant. For additional information about the design intent behind these frames, see REP-0105.

  7. On the standalone computer and within the tf2_echo terminal, note the latest transform reported between /map and the appropriate camera frame.

    1. Validate orientation:

      Validate that the orientation error between the measured rotation unit quaternion and the expected rotation unit quaternion ([0 0 0 1], because the path is a closed loop) is less than 15 degrees. This error can be calculated using the process explained here:

      \[q_{measured} = w_m + x_m i + y_m j + z_m k\]
      \[q_{expected} = w_e + x_e i + y_e j + z_e k\]
      \[2 \arccos (w_m w_e + x_m x_e + y_m y_e + z_m z_e) \leq 15 \degree\]

    2. Validate translation:

      Validate that the magnitude of the measured translation is within 2 meters (5% of the 40m path length) of the expected translation (0 meters, because the path is a closed loop):

      \[0 \text{m} \leq \|t_{measured}\| \leq 2 \text{m}\]

If all real-world measurements correspond to the reported estimates within the expected tolerance, the Loop Closure test is a success.

Next Steps

Troubleshooting

If any step of the tests did not succeed, then there might be a configuration error with your system. Follow these troubleshooting suggestions.

If the prepared troubleshooting steps are unable to resolve the test failure, contact your NVIDIA support representative and file an issue report with all relevant details.

Integrating Isaac ROS Visual SLAM

While the Isaac ROS Visual SLAM node exposes a multitude of interfaces, the core way to consume the node’s output is through the reported transform between the /map and camera frames.

Based on your robot’s precise specifications, you can set up a static transform publisher or URDF-based publisher that connects the camera frame to a relevant planning frame on the robot (typically named /base_link). Then, your navigation stack can use the transform between /base_link and /map to plan movements within the local space.

Consider further augmenting your robot’s navigation stack with other Isaac ROS packages, such as Isaac ROS Nvblox for 3D reconstruction.