[FollowLine] using the Monaco 3D circuit increases too much the computational cost of simulation

[jmplaza]

The Real Time Factor (RTF) drops to unacceptable low values and so the simulations is too slow. It does not happen with other circuits, where the RTF keeps at high values.

[javizqh]

Linked with JdeRobot/RoboticsInfrastructure#632

[aryan-py]

Good Day @jmplaza @javizqh, I looked into this issue and it seems to be a physics engine problem due to the number of triangles used in the model of the track specifically due to Bounding Volume Hierarchy in which because on flat surface the force is normal the triangles are grouped together but on curved surface it is not possible, so the rtf fall to prevent tunneling as the collision with triangles are done separately .

[aryan-py]

I tried decimating the .dae model to reduce the triangles in the mesh but it does not work,
I think something like using:
levels
heightmap
convex decomposition
may work what are your thought on this ??

[Sarvesh-Mishra1981]

Hello @jmplaza
After looking into the performance issue in the FollowLine exercise with the Monaco circuit, it seems that the slowdown may be coming from how Gazebo handles collisions with the track. The Monaco circuit is stored as a detailed .dae mesh with thousands of tiny triangles to capture all the curves and surface details. If this same complex mesh is also used for collision detection, the physics engine has to constantly perform calculations against a very large number of triangles. This can put significant load on the CPU and cause the Real Time Factor (RTF) to drop,

One possible idea to improve this can be to separate the visual model of the track from the collision model. The detailed mesh could still be kept for rendering so the circuit looks exactly the same, but a simpler representation could be used for collisions. For instance the elevation of the track surface could be extracted from the mesh and converted into a 2D heightmap. Gazebo could then use this heightmap to determine the ground height instead of calculating collisions against thousands of triangles. Since heightmaps are much lighter computationally, this approach could potentially reduce the load on the physics engine what do you think?

cheers!!

[tony19053000]

Hello @jmplaza , I tested this locally, and I think a possible reason for the lower RTF in Monaco is that monaco_track.dae (Monaco track file) is being used directly for collision, and that file includes both the road and the side walls together, which may be heavier. In the simpler circuits, the road and walls are handled separately, which looks lighter.

In Monaco, the track file is being used directly for collision:

In simple_circuit, the road collision is a plane, and the wall collision is separate:

I also tried a simpler collision version locally just to check the behavior, but in my setup, the Monaco RTF dropped even more after that test.

So at least from what I checked, I think this looks more likely related to how the Monaco circuit collision is set up, rather than a Follow Line controller issue.