Extrinsic Calibration for Mobile Robots#

You can use these instructions to calibrate the positions and orientations of sensors on a mobile robot. This process is known as Extrinsic Calibration and it enables robotics functions that combine information from multiple sensor modalities.

The extrinsic calibration results are encoded in a URDF calibration file, which contains the relative position of different sensors on the robot.

Instructions#

We suggest evaluating existing calibration solutions from companies such as Main Street Autonomy (MSA) or Tangram Vision for your particular needs.

Note

Make sure that the calibration solution supports the Extrinsic Calibration Requirements and that the resulting URDF calibration file follows the Isaac Requirements for URDF files.

The usual process for most calibration solutions include:

Define a nominals URDF file and publish its transformations to the /tf_static topic, so that they can be recorded in the rosbag in the next step.
Recording a rosbag with your robot. This step may include driving the robot in a particular motion or placing a calibration target around the robot.
Processing the rosbag with a calibration tool or sharing it with the calibration provider. The calibration tool or service will generate a URDF calibration file.
If the chosen calibration solution does not support camera - ground calibration, you may need to use additional calibration tools to complement the extrinsics calibration URDF.
Placement of the URDF calibration file in the robot.

See the sections below for additional details.

URDF nominals file:

This step may be optional depending on the calibration solution and sensors that you are calibrating. For example, camera-camera calibration with overlapping Field of View (FoV) may not require nominals. On the other hand, IMU, 2D LIDAR, or robot odometry may not be fully observable or reliable during calibration. In those cases, nominals allow to constrain the calibrated URDF file to the best known measurements from a CAD model. Make sure to follow the requirements and understand the limitations of the chosen calibration solution.

Unless nominals are provided for the robot that you are calibrating (see the Using Nominals section), in the general case it is strongly recommended that you define a nominals URDF file for your robot and publish its transformations to the /tf_static topic, so that they can be recorded in the rosbag that will be used for calibration.

Rosbag recordings used for calibration:

Most calibration solutions expect a rosbag as input data. You may need to use appropriate rosbag conversion tools to uncompress the messages in the rosbag to the types expected by most calibration solutions:

Camera streams: Uncompressed sensor_msgs/msg/Image messages.

3D LIDAR data: sensor_msgs/msg/PointCloud2 messages (with fields x, y, z, intensity, ring, timestamp).

Extrinsic Calibration Requirements#

Note

Recommended prior reading: Isaac Requirements for URDF files

In order to be compatible with Isaac applications, extrinsic calibration tools must:

Generate a URDF calibration file complying with the Isaac Requirements for URDF files.
Perform extrinsic calibration for all sensors without optimizing intrinsic calibration or stereo calibration of any of the sensors (that is, by fixing the calibration from the sensor drivers).
- In the event that it is necessary to re-calibrate intrinsics or stereo calibration, make sure that all the following are true:
  The re-calibrated sensors have a native a mechanism to overwrite their intrinsic calibration and/or stereo calibration (e.g., the calibration stored in the EEPROM).
  
  The overwrite process is performed before using the extrinsic URDF calibration file with Isaac applications.
  
  The resulting URDF still complies with the Isaac Requirements for URDF files (stereo calibration is not included in the robot URDF).
- Calibration transformations for all sensors must include 6 Degrees of Freedom (DoF), unless certain DoF are not observable during the data recording (see following points).
Perform 6-DoF sensor-to-robot calibration and encode the result in the transformation to base_link, unless certain DoF are not observable during the data recording (see following points).
Fix nominal transformations (from /tf_static) for the DoF that are not observable (or unreliable) during the data recording.
- Unobservable DoF depend on the calibration process, robot motion (if any), and calibration target design and motion (if any), among others.
- Some specific examples for wheeled robots include :
  IMU: Axes that are not excited during data recording.
  
  base_link: X and Y-translation and yaw rotation for robots with unreliable odometry.
Perform sensor-to-ground calibration. See clarifications below:
- Ground calibration refers to the 3-DoF transformation between a sensor and the ground plane: Z-translation, and pitch and yaw rotations.
- The ground plane should be computed from sensor data.
- The transformations between base_link and the robot sensors must be modified so that base_link is on the ground, with its X-Y axes representing the ground plane.
- Ground calibration must not modify the transformation between individual sensors. It must only modify the transformation between all sensors (as a rigid body) and base_link, such that ground points perceived by sensors lie on the X-Y plane of the base_link frame.

Using Nominals#

It is recommended that you always use the isaac_calibration.urdf file generated in the calibration instructions above, because it accurately reflects the true extrinsics of your physical robot. However, as a backup option, nominal xacro files may also be available for some platforms.

By default, these nominal values are used by Isaac applications, if a calibrated URDF file is not present in the system.

Note

If using nominal values with sensors mounted on a robot, it is still necessary to provide a URDF encoding their position on the robot.