Mert Güler published a set of 42 Gazebo simulation worlds on October 18, aimed at testing navigation, exploration and SLAM. The repository is Apache 2.0. The worlds are not built in Gazebo's editor but in Blender, driven by Python scripts.
Key takeaways
- 42 worlds in the mertgulerx/world-collection repository, Apache 2.0 licensed
- Generated through Python and Blender, exported to .sdf files
- The set covers mazes, offices, warehouses, forests and moving-obstacle scenes
- For now these are two-dimensional, floor-plan based environments
- The author plans 3D textured worlds for perception and mapping work
The problem: Gazebo is awkward to author in
The reason for the collection is practical. Creating and developing worlds directly in Gazebo tends to be painful, so the author moved the whole process into Blender and wrapped it in Python scripts.
The output lands in .sdf files. That format gives compatibility with other simulation stacks, makes geometric calculations easier and lets scenes be changed without clicking through an editor.
git clone https://github.com/mertgulerx/world-collection.gitWhat is in the set
| Category | What it is for |
|---|---|
| Mazes | exploration planner tests in tight corridors |
| Offices and warehouses | typical interiors for service and logistics robots |
| Forests | irregular obstacles, no straight walls |
| Recreations of real places | scenarios closer to deployment |
| Moving obstacles | response to a changing environment |
| Challenge-oriented scenes | hard cases picked deliberately |
That range targets navigation and exploration testing, LiDAR work and robot model setup.
What is still missing
There is no 3D geometry and no texturing. The author himself notes these are needed for perception tasks, 3D mapping, object detection and robot training, and says he intends to add them.
Where the collection came from
The worlds were not built as a product in their own right. Since April, Güler has been developing frontier_exploration_ros2, a C++ frontier exploration package for autonomous mobile robots released under Apache 2.0. It runs with ROS 2 Jazzy, but the core is a C++ library for custom integrations.
Why it matters
Navigation and exploration benchmarks suffer from a lack of shared scenes. Every team builds its own, which makes results hard to compare. An open set of 42 environments cuts that cost, and procedural generation means producing a scene variant instead of clicking one together from scratch. The two-dimensional limit does close off perception testing.
The two-dimensional limit is not accidental. Obstacle geometry is enough to score a path planner — textures and a third dimension do not change the result. Only when a perception module is under test does the lack of visual appearance become a blocker, and that is exactly what the author flags as the next step.
What's next
- The author plans to add three-dimensional, textured worlds needed for perception, 3D mapping and object detection
- The repository was created on October 18 and is being actively updated
