These steps are to setup the repo to work on your own PC. We utilize docker to enable ease of reproducibility and deployability.
Why docker? It's so that you don't need to download any coding libraries on your bare metal pc, saving headache :3
- This repo is supported on Linux Ubuntu >= 22.04, Windows (WSL), and MacOS. This is standard practice that roboticists can't get around. To setup, you can either setup an Ubuntu Virtual Machine, setting up WSL, or setting up your computer to dual boot. You can find online resources for all three approaches.
- Once inside Linux, Download Docker Engine using the
aptrepository - You're all set! You can visit the WATO Wiki for more documentation on the WATO infrastructure.
demo_nov30.mp4
We use Ignition Gazebo for physics simulation. The main world file is robot_env.sdf, which defines the entire simulation scene.
A basic model of the rover is defined directly in the SDF with a differential drive base. Key components:
- Chassis: 2.0m x 1.0m x 0.5m box with inertial properties
- Wheels: Four cylindrical wheels with revolute joints
- IMU: Mounted on chassis, publishes to
/imu - RGBD Camera: Simulated RealSense D435 mounted on the front
The depth camera is configured as an rgbd_camera sensor with realistic FOV, intrinsics, etc.
It publishes:
/camera/image(Image) - RGB image (used by YOLO object detector)/camera/points(PointCloud2) - 3D point cloud (used by costmap)
The environment is loaded from mars_env.urdf, which includes meshes for the mars terrain, mallet, and water bottle.
The terrain has collision geometry so the rover can't drive through obstacles.
The ros_gz_bridge translates Ignition messages to ROS2 topics, bridging /cmd_vel, /imu, camera streams, and TF transforms. This way, our ROS2 autonomy stack interface seamlessly with the Gazebo sim.
- gps_sim: Simulates GPS measurements by reading ground truth transforms from the simulator and injecting realistic noise and dropouts. Publishes NavSatFix messages.
- imu_sim: Adds sensor noise and bias to clean IMU data to mimic real-world sensors. Takes ground truth IMU and outputs noisy orientation and angular velocity.
- localization: Runs an Extended Kalman Filter (EKF) to fuse (simulated) noisy GPS and IMU data into a smooth, filtered odometry estimate. Outputs robot pose for downstream modules.
- costmap: Converts RGBD camera point clouds into a local 2D occupancy grid centered on the robot. Marks obstacles and inflates them for safety margins.
- map_memory: Stitches together local costmaps into a persistent global map as the rover explores. Only updates when the robot moves beyond a threshold distance to reduce computational load.
- yolo_inference: Runs YOLOv8 object detection (ONNX runtime) on camera images to detect objects like bottles and mallets. Publishes annotated images with bounding boxes.
- planner: Implements A* search on the global occupancy grid. Takes the current robot pose and a goal point, outputs a collision-free path. Replans when new goals arrive.
- control: Pure pursuit controller that tracks the planned path. Uses a PID loop (configurable Kp, Ki, Kd gains) to compute steering corrections based on cross-track error. Outputs angular velocity to steer toward the target waypoint while maintaining constant forward velocity, publishing Twist commands to
/cmd_vel.
The pipeline runs as follows:
- Sensor sims generate noisy data
- Localization fuses it into an estimate of the robot's position (odometry)
- Costmap builds local obstacles
- Map memory accumulates into global map
- Planner finds path to goal
- Control executes path
- Motor commands actuate the rover