Development Ideas

The AI Challenge leverages open-source software. By utilizing the code and web platform provided by the organizers, you can skip the initial development phase and immediately start developing in line with the competition theme. This approach has the significant advantage of avoiding "reinventing the wheel." Additionally, it allows anyone to easily participate in the competition and enables the competition to be run with consistent evaluation criteria.

For first-time participants, you will start from a state where most of the functions necessary for autonomous driving are already in place, standing on the foundation built by your predecessors. From here, you have the opportunity to deepen your unique development in the competition field through the community's "publication of efforts." Furthermore, to deepen your understanding of autonomous driving, we recommend using the " Autoware Practice " prepared by the organizers and the learning programs provided by the ROS 2 community, such as " ROS 2 ".

For those who have already participated in the challenge, we encourage you to share your experiences, contribute to the community, and help the competition evolve. Your active participation will contribute to making the competition even more fulfilling.

Participants will proceed with development by customizing this code and parameters. However, if you are unfamiliar with Autoware, we recommend going through the introductory course first.

• For those who want to know the specifications, such as those developing independently without using the repository's code, refer to the interface specifications and simulator specifications pages.

Try switching the control mode

You can switch the control method by changing the control_method argument in reference.launch.xml. If you are unsure what to do after setting up the environment, try the various control methods to experience how the vehicle behavior changes.

• mpc (default): MPC-based control
• pure_pursuit: Pure Pursuit-based control
• tiny_lidar_net: TinyLiDARNet End-to-End control (outputs acceleration and steering angle directly from 1080 LiDAR points)
• pilot_net: PilotNet End-to-End control
• joycon: Manual teleoperation

For example, to try End-to-End control, change the default value of control_method in reference.launch.xml to pilot_net and run it.

Try changing the control parameters

Next, let's see how control parameters affect behavior. We'll change the parameters of the simple_pure_pursuit control module.

First, change the default value of control_method in reference.launch.xml to pure_pursuit.

Then adjust the value entries in $HOME/aichallenge-racingkart/aichallenge/workspace/src/aichallenge_submit/aichallenge_submit_launch/launch/control/pure_pursuit.launch.xml:

<node pkg="simple_pure_pursuit" exec="simple_pure_pursuit" name="simple_pure_pursuit_node" output="screen">
    <param name="use_external_target_vel" value="false"/>
    <param name="external_target_vel" value="10.0"/>
    <param name="lookahead_gain" value="0.5"/>
    <param name="lookahead_min_distance" value="3.5"/>
    <param name="speed_proportional_gain" value="1.0"/>

After making adjustments, try running again. You should be able to confirm that the behavior has changed.

For example, setting lookahead_gain to 0.1 and lookahead_min_distance to 0.5 makes the lookahead distance extremely short, causing the vehicle to zigzag and struggle to drive straight. This is a good way to feel the impact parameters have on driving behavior.

Try creating a new package

Try creating a new custom package. First, copy open-source packages or the autoware practice. It would be best if you proceed as follows.

1. Copy the original package and change the following
   • Package name
   • Folder name
   • Code
   • package.xml
   • CMakeLists.txt
2. Place it in aichallenge_submit

3. Change the launch file (reference.launch.xml) in aichallenge_submit_launch

4. Please ensure that you do not violate the license of the copied package.

[Optional] Try editing the map

For the 2025 AI Challenge, we recommend editing maps such as point cloud maps and lanelet2 maps using tools like the Trajectory Editor created by the Gifu University team.

See the Readme for step-by-step instructions.

Store the created lanelet2 map in aichallenge/workspace/src/aichallenge_submit/simple_trajectory_generator/data.

Check the evaluation results

When you run make eval and the required 6 laps are completed, the evaluation results are automatically saved under output/<timestamp>/d<domain_id>/. You can also access them via symlink at output/latest/d<domain_id>/. Try running it with the default code first to understand your current baseline.

Try submitting

After customizing the workspace, refer to this to submit.

Read reference articles by voluntary participants

The efforts of voluntary participants are summarized in the Advent Calendar, so please refer to them.

If you are unsure where to start, we recommend starting with this article written by Mr. Arata Tanaka, who won the Community Contribution Award in 2023.