An example procedure for adding and evaluating a new node#

Overview#

This page provides a guide for evaluating Autoware when a new node is implemented, especially about developing a novel localization node.

The workflow involves initial testing and rosbag recording using a real vehicle or AWSIM, implementing the new node, subsequent testing using the recorded rosbag, and finally evaluating with a real vehicle or AWSIM.

It is assumed that the method intended for addition has already been verified well with public datasets and so on.

1. Running Autoware in its standard configuration#

First of all, it is important to be able to run the standard Autoware to establish a basis for performance and behavior comparison.

Autoware constantly incorporates new features. It is crucial to initially confirm that it operates as expected with the current version, which helps in problem troubleshooting.

In this context, AWSIM is presumed. Therefore, AWSIM simulator can be useful. If you are using actual hardware, please refer to the How-to guides.

2. Recording a rosbag using Autoware#

Before developing a new node, it is recommended to record a rosbag in order to evaluate. If you need a new sensor, you should add it to your vehicle or AWSIM.

In this case, it is recommended to save all topics regardless of whether they are necessary or not. For example, in Localization, since the initial position estimation service is triggered by the input to rviz and the GNSS topic, the initial position estimation does not start when playing back data unless those topics are saved.

Consider the use of the mcap format if data capacity becomes a concern.

It is worth noting that using ros2 bag record increases computational load and might affect performance. After data recording, verifying the smooth flow of sensor data and unchanged time series is advised. This verification can be accomplished, for example, by inspecting the image data with rqt_image_view during ros2 bag play.

3. Developing the new node#

When developing a new node, it could be beneficial to reference a package that is similar to the one you intend to create.

It is advisable to thoroughly read the Design page, contemplate the addition or replacement of nodes in Autoware, and then implement your solution.

For example, a node doing NDT, a LiDAR-based localization method, is ndt_scan_matcher. If you want to replace this with a different approach, implement a node which produces the same topics and provides the same services.

ndt_scan_matcher is launched as pose_estimator, so it is necessary to replace the launch file as well.

4. Evaluating by a rosbag-based simulator#

Once the new node is implemented, it is time to evaluate it. logging_simulator is a tool of how to evaluate the new node using the rosbag captured in step 2.

When you run the logging_simulator, you can set planning:=false or control:=false to disable the launch of specific component nodes.

ros2 launch autoware_launch logging_simulator.launch.xml ... planning:=false control:=false

After launching logging_simulator, the rosbag file obtained in step 2 should be replayed using ros2 bag play <rosbag_file>.

If you remap the topics related to the localization that you want to verify this time, Autoware will use the data it is calculating this time instead of the data it recorded. Also, using the --topics option of ros2 bag play, you can publish only specific topics in rosbag.

There is ros2bag_extensions available to filter the rosbag file and create a new rosbag file that contains only the topics you need.

5. Evaluating in a realtime environment#

Once you have sufficiently verified the behavior in the logging_simulator, let's run it as Autoware with new nodes added in the realtime environment.

To debug Autoware, the method described at debug-autoware is useful.

For reproducibility, you may want to fix the GoalPose. In such cases, consider using the tier4_automatic_goal_rviz_plugin.

If your implementation works successfully, please consider a pull request to Autoware.

It is also a good idea to start by presenting your ideas in Discussion at Show and tell.

For localization, YabLoc's Proposal may provide valuable insights.