Identity subt exploration behaviour on urban environments algorithm gaps.

[okyspace]

Issue by: katcujic-cmu Created at: 2021-05-25 05:48:52

Assignees: katcujic-cmu

---

No description provided.

[okyspace]

Commented by: katcujic-cmu 2021-06-07 07:02:58

---

Step 1:

Do more simulations with:

• openstreetmap
• google earth import?
• lidar -> mesh ?
• Airlab's AirSim

See how the uav drone behaves in such instances on longer runs. See how multi uav behave in such instances on longer runs. See how multi uav/ugv behave in such instances on longer runs.

[okyspace]

Commented by: katcujic-cmu 2021-06-09 06:25:32

---

Some initial findings:

OpenStreetMap

Limitations:

• Cannot load a large map easily -> gazebo takes long to load and eventually UAV sim fails
• Better textures need to be added
• No outside/inside transition in buildings
• No trees other objects from OSM download
• Download, format conversion is not automated.
• No dynamic objects (such as people).
• Robot initial start point must be in an available 'opening'

Google Earth Import:

Limitations:

• Not yet tried out, however it requires a very very ugly, not able to be automated hack to download the datasets...
• Similar to OSM. However it looks like the Google Earth textures are better when downloaded.

Lidar2Mesh:

Limitations:

• Conversion is not automated, requires manual user steps.
• 2Mesh is not the most accurate representation: jagged, holes, not smooth, etc.
• Gazebo robot & 2Mesh scale might be incorrect.
• Robot initial start point must be in an available 'opening'
• There is some py/c++ code available online, but might not similarly be as accurate. Most similar was proprietary software that is too expensive to buy.
• Will not easily work with UGVs. Since the ground places are not nicely flat, the UGV will get stuck very often.

AirSim:

• Have not yet set it up.
• Might require a change in setup to work with the SubT drones?
• There looks to be a OSM import available for Unreal Engine? Perhaps OSM + UE would be a good combination for larger maps?

From all of these available, similar issues will need to be addressed:

• How to run multiple simulations, headless, with everything automated for various world types.
• How to run large maps.
• Downloaded maps fail to have inside/outside, will require some automation to convert the downloaded objects to have inside/outside (such as doors, windows).
• Dynamic objects
• Different lighting conditions.

Perhaps a combination of these world types, such as:

• Case 1: Gazebo, OSM open area, no dynamic obstacles,
• Case 2: Gazebo, OSM open area, indoor/outdoor, trees, more 'non building' objects.
• Case 3: Gazebo, Cloud2Mesh world, more underground cases, city scans too, indoor/outdoor.
• Case 4: AirSim: OSM open area
• Case 5: AirSim: OSM open area, indoor/outdoor, trees, glass windows
• Case 6: AirSim: OSM open area, indoor/outdoor, trees, dynamic obstacles
• Case 7: AirSim: OSM open area, indoor/outdoor, trees, dynamic obstacles, lighting changes

Then we would need to define a few 'scenarios' for each world type:

• Multi robot simulation, where all the robots are using the in same world type
• Where the simulation would start a 'cluster' of computers to have the multi robot simulation.
• Have multiple scenarios with different robot starting points, other cases, etc.
• Results in an evaluation of various algorithm 'accuracy metrics'