Understand the measured challenge examples
The Leaderboard's Example results section shows a task's reference score, Download code and View five runs together. Select a task to see its example. The reference is excluded from participant ranks.
The six examples below use WWW-POSIM challenge scoring rules and Python SDK 0.4.2. Each score comes from five completed official runs. The commented file and the exact scored original are both available. The publication checks their Python ASTs for equality: comments explain the program without changing its execution logic. The record preserves its original submission number, run IDs and layout seeds.
| Task | Code | Five measured scores | Aggregate | Record |
|---|---|---|---|---|
surface_station |
Commented · Scored original | 70.03, 16.03, 6.16, 0.74, 1.18 | 15.2108 | ↗ |
surface_gates |
Commented · Scored original | 20.35, 21.27, 20.08, 15.11, 20.17 | 18.5392 | ↗ |
surface_slalom |
Commented · Scored original | 16.00, 16.00, 16.00, 16.00, 1.00 | 10.6000 | ↗ |
underwater_gate |
Commented · Scored original | 13.33, 13.33, 13.33, 13.33, 13.34 | 13.3338 | ↗ |
underwater_path |
Commented · Scored original | 15.00, 15.00, 15.00, 15.00, 10.00 | 13.2002 | ↗ |
underwater_return |
Commented · Scored original | 13.33, 10.00, 6.67, 10.00, 10.01 | 9.3347 | ↗ |
Run one example
- Download the matching commented code and save it as
main.py. Open it in VS Code; its comments in the selected language describe the task and control blocks. - In the installed app, prepare the Practice arena, choose the same task and condition, and keep ROS 2 direct control selected. Start the generated world.
- Press Start scoring on the task card, then run
python main.pyin the project's prepared terminal. Installed practice gives the complete setup. - Stop with Ctrl+C or Stop input. The SDK releases control. Inspect the recorded path, then restart the world for another trial.
- For an official result, follow Submit your challenge program. The server prepares its own Python/SDK environment. Keep credentials out of the code.
What each task's controller does
| Task | Goal | Active parts of the code |
|---|---|---|
surface_station |
Reach and continuously hold the target position and heading | Approach/hold hysteresis, heading error, bounded forward and yaw commands |
surface_gates |
Pass gates in order, return through the corridor and hold at home | Target progression, retraced return points and steady-current compensation |
surface_slalom |
Clear the buoy course and reach the final hold | Visibility-graph detours, a heading limit before forward motion and current compensation |
underwater_gate |
Dive, cross the gates and maintain the final depth/heading | Gate-plane crossing, entry-side recovery and six-thruster depth/attitude control |
underwater_path |
Follow inspection targets and their depth/hold requirements | Inspection transit gains and bounded drift-compensating thrust |
underwater_return |
Traverse obstacles, return and reach the final ascent/hold | Small-hull detours, velocity-error integration and thruster allocation |
The route files share helper functions. mission['task'] chooses the active branches; a helper for another task can remain unused in the selected trial. The controller's target index is its own navigation state. The server measures the vehicle independently and awards the score.
Read the program in sections
- Connection and mission:
Session.local(control=True)discovers the prepared session.sim.profile['mission']supplies this trial's targets and conditions; positions are not fixed in the script. - Sensor input: the code selects
nav_msgs/msg/Odometryby message type.sim.latest(..., max_age=0.5)requires a sample received within 0.5 wall-clock seconds. - Pose and target error: quaternion orientation becomes yaw/roll. Positions and mission targets share map/world axes in metres; positive z is up, so underwater targets use negative z. Odometry twist is in body axes. Yaw/roll use radians; mission current angles and swing use degrees.
- Navigation state: the code checks target proximity, gate-plane passage and continuous holding. Losing the hold conditions resets the local timer. The gate helper assumes the published X-aligned gate planes.
- Control: WAM-V uses forward speed in m/s and yaw rate in rad/s. BlueROV2 uses six forces in native joint order in newtons. The allocation uses the bundled 45-degree horizontal thrusters and lever arms of 0.167 m and 0.109 m.
- Timing: ROS stamps, hold timers and integrator
dtuse simulation seconds. SDK commandduration=0.25uses wall-clock seconds, refreshes commands at 10 Hz and releases the command lease before returning. A slow simulation therefore advances less simulation time per call. - Stopping and errors:
control_lease_endedends the loop normally. Other errors remain visible. Leaving theSessioncontext closes the session and releases control.
Tune with the five records
Read the task-specific comments before changing a gain. Adjust one parameter at a time: distance-to-speed gain, heading gain, speed/force limit, gate clearance or drift-compensation gain. Compare the path and error values across all five conditions. The underwater loop bounds its integral estimate and uses a fixed vertical trim; these values belong to the bundled model.
The displayed scores are the linked measured results. A new run can differ with timing and environment. These learning controllers have modest measured scores and do not guarantee complete autonomous navigation. Comments preserve the scored logic; changing executable statements creates a new controller that needs a new evaluation.