Team project · UQ Space · 2025–26
Theseus V rover autonomy
1st of 17
The ROS 2 autonomy stack I wrote on my own for UQ Space's Theseus V rover. It ran the autonomy task at the 2026 Australian Rover Challenge, where the team placed 1st of 17.
- My role
- Lead avionics engineer. Sole developer of the autonomy stack (82 of 83 commits) and of Theseus GCS, the ground control station (87 of 87 commits)
- Team
- Avionics, one of the 6 UQ Space sub-teams that developed Theseus V
- When
- Oct 2025 – Apr 2026
- Tools
- ROS 2, Nav2, FAST-LIO2, C++, Python, Gazebo, YOLO, React
- Code
- Private: UQ Space team repositories
- Press
- Interviewed for the Australian Space Agency's Gen Space feature (opens in a new tab) (28 Apr 2026) on the team's win
The full write-up
- The problem
- What I did
- Result
- What went wrong
- What's next
- Credits
The problem
In the Australian Rover Challenge's mapping and autonomy task, the rover is dropped into an unknown arena. It has to explore, build a map and report what it finds with nobody driving.
What I did
Autonomy stack
- The stack is 9 ROS 2 packages on two computers. An Orange Pi 5 Plus runs Nav2, FAST-LIO2 LiDAR-inertial odometry from a Livox MID-360 LiDAR, and KISS-SLAM, which turns FAST-LIO2's output into the 2D occupancy map that exploration works from. A Jetson Xavier NX runs YOLO and ArUco detection on the cameras.
- A frontier detector picks the next goal, and a 10-state supervisor runs the mission.
- For safety I wrote 4 costmap layers in C++ (the maps Nav2 plans around) and a 4-layer geofence that slows, then stops, the rover near the arena boundary.
- I built a Gazebo arena with placards and randomly placed cubes, so I could test the whole stack in simulation before the hardware was ready.
- The rover doesn't spin in place. Spinning a skid-steer rover can dig its wheels into sand and draw high current, and KISS-ICP, the LiDAR-only odometry I started with, lost track during pure rotation. So the 9-stage recovery sequence I wrote for Nav2 clears the costmap, waits, nudges forward and backs up instead, and the path follower turns in arcs.
Theseus GCS
The ground control station is a React app. Its autonomy tab sends the arena layout, launches the autonomy stack and switches the rover between teleop (remote driving) and auto.
Result
At the competition Theseus GCS ran the autonomy and mapping task; the driving tasks used a separate camera feed with less delay.
What went wrong
- IMU vibration: the MID-360's built-in IMU sits next to the LiDAR's spinning prism motor, and the motor's vibration coupled straight into it. Every LiDAR-inertial odometry we tried on the rover (FAST-LIO2, Point-LIO and DLIO) folds the IMU into its filter, so the extra noise showed up as persistent drift in the odometry, while LiDAR-only KISS-ICP did better on the same data. Before the competition I added an IMU filter node that calibrates the gyro bias at start-up, updates it whenever the rover is stopped and clamps the gyro after a stop, but it couldn't remove vibration while the rover was driving. For Theseus VI we moved to an external IMU on the rigid sensor plate and kept the MID-360's IMU for logging only.
- Placard detections: shortly before the competition I switched the ArUco detectors to best-effort QoS, but my topic mux still subscribed to them as reliable. ROS 2 never connects a reliable subscriber to a best-effort publisher, so live placard detections couldn't reach the mission supervisor during the run. I fixed the mux's QoS the day after.
What's next
For Theseus VI I've designed a sensor head that indexes the LiDAR to 10, 15 or 20 degrees, and an avionics bay layout with checked clearances. Nothing is built or tested yet.
Credits
Theseus V is a UQ Space team rover. The page covers the software I wrote for it; the rover itself is the work of UQ Space's aerostructures, avionics, payload, propulsion, recovery and systems engineering sub-teams.