Refactor of robot repositories

[alaurenzi]

@liesrock @DavideAntonucci @EdoardoRomiti
Hi all,
I am doing some work to improve our workflow concerning robot URDF/SRDF/capsules generation, and simulation, too.

Why

The weak points of our previous infrastructure, in my opinion, are the following:

1. one single big bash script doing everything and every time it is invoked
   • generate urdf and srdf for all existing config files (from urdf/config/*.urdf.xacro)
   • generate capsule models
2. does not report errors (or exit on error)
3. does the generation even though no file has actually changed
4. generates intermediate files inside the source tree (thus polluting it)
5. unclear which files are generated (byproducts of generation) and which are actual source files
6. no unit tests and continuous integration
7. no debian generation
8. issues with gazebo11 due to unsupported floating joint

How

I implemented some solutions inside a refactor branch. Key points are:

1. basic structure does not change (all xacros stay where they are)
2. main xacro must support an additional argument config with a default value pointing to a default configuration file
3. cmake-based xacro generation with single targets for every config file; benefits are:
   • generation skipped if source files haven't changed
   • files are generated inside the build tree and published to the source tree with an explicit command
   • errors causes early exit
4. dropped support for drag-and-drop simulation in gazebo
5. perception simulation with additional arguments to be provided to the launch file (realsense:=true velodyne:=true)
6. a basic integration test demonstrating that simulation is spawned successfully and xbot2 can start correctly
7. Travis CI
8. cpack based deb generation (example here)
9. conditional inclusion of a floating joint (just for control)

Usage

1. cloning the repository yields a ready-to-use package, generation is not required
2. make install support in case the source folder is not part of ROS_PACKAGE_PATH
3. from a developer's perspective, generation must be performed with a make <target-name> from the build folder, such as
   • make generate_urdf_<config_name>
   • make generate_urdf_<config_name>_capsule
   • make publish_urdf_<config_name> (copy to source tree)
4. make package produces a deb package

Feedback

I'd like to hear some feedback from you before going further ahead with this project.
Do you see a better way to accomplish some of the goals?
Do you see further shortcomings of the current architecture that can be solved?

@EdoardoRomiti I am not sure how and if this affects you work (generation not from xacros but from discovered topology)

[EdoardoRomiti]

Personally I like many of the ideas here. With modular we follow a slightly different workflow for urdf/srdf generation than with Centauro, but I would say we have the same weakness in the Why section (apart from 1,4 and 5):

• In our case 4 and 5 are generally not an issue because we don't actually build anything, but only write files. We even removed the CMakeLists.txt from our sources, which is instead only generated for the deployed packages in order to make it a ROS package. We keep the "build space" in the /tmp folder and files are only "installed" to the desired directory (usually part of ROS_PACKAGE_PATH) only when calling a deploy script. Therefore the source directory can be isolated.

• Issue 1 is instead solved by having a python API that allows to call methods such as: read_from_ethercat(), add_module(), write_urdf(), write_srdf(), deploy_robot(), etc., allowing then to choose in the python script or console which operation to perform and which robot to operate on. I would like to keep this possibility (in modular) to write and deploy a robot programmatically as I think is quite useful.

• All the others are issues also for us, or anyway what you propose could improve what we normally do. So i would be happy to integrate what you have done also in modular. For the automatic discovery of the topology this will not be immediate and will require to change some things, but I agree would be better in the long run.

[DavideAntonucci]

Hi @alaurenzi,

Thanks a lot that you've involved me in this discussion. Personally, I really like the approach described before to create and maintain the URDF (Usage points 3 and 4).
Another good point is Debian packaging creation that I've already used and it works.

The HOW points (4,5 and 6) are very useful for me when I want to play immediately with the robot.

I don't see weakness points using this approach right now.

[alaurenzi]

@nkashiri give this deb package a try
Also check if the issues you encountered are being addressed here

[nkashiri]

following the installation of the deb file, and sourcing the bashrc, it is needed to
~$ rospack profile
so that the robot can be simulated via
~$ roslaunch centauro_gazebo centauro_world.launch

[nkashiri]

There are these warnings:

[INFO] [1637584501.179652, 0.105500]: Calling service /gazebo/spawn_urdf_model
Warning [parser_urdf.cc:1115] multiple inconsistent exists due to fixed joint reduction overwriting previous value [Gazebo/Residential] with [Gazebo/BlackTransparent].
Warning [parser_urdf.cc:1115] multiple inconsistent exists due to fixed joint reduction overwriting previous value [Gazebo/Residential] with [Gazebo/BlackTransparent].
Warning [parser_urdf.cc:1115] multiple inconsistent exists due to fixed joint reduction overwriting previous value [Gazebo/Residential] with [Gazebo/BlackTransparent].
Warning [parser_urdf.cc:1115] multiple inconsistent exists due to fixed joint reduction overwriting previous value [Gazebo/Residential] with [Gazebo/BlackTransparent].

Solution: when removing the wheels material line from centauro.gazebo.xacro, the warning is removed

[nkashiri]

Bonus point! it would be nice if the robot would fall on the ground from a shorter height.

[liesrock]

Grande @alaurenzi, I try to answer step by step:

Why

The weak points of our previous infrastructure, in my opinion, are the following:

1. one single big bash script doing everything and every time it is invoked

   • generate urdf and srdf for all existing config files (from urdf/config/*.urdf.xacro)
   • generate capsule models

2. does not report errors (or exit on error)

3. does the generation even though no file has actually changed

4. generates intermediate files inside the source tree (thus polluting it)

5. unclear which files are generated (byproducts of generation) and which are actual source files

6. no unit tests and continuous integration

7. no debian generation

8. issues with gazebo11 due to unsupported floating joint

The motivation are clear and the issues in all the reported points should be tackled.

How

I implemented some solutions inside a refactor branch. Key points are:

1. basic structure does not change (all xacros stay where they are)

2. main xacro must support an additional argument config with a default value pointing to a default configuration file

3. cmake-based xacro generation with single targets for every config file; benefits are:

   • generation skipped if source files haven't changed
   • files are generated inside the build tree and published to the source tree with an explicit command
   • errors causes early exit

4. dropped support for drag-and-drop simulation in gazebo

5. perception simulation with additional arguments to be provided to the launch file (realsense:=true velodyne:=true)

6. a basic integration test demonstrating that simulation is spawned successfully and xbot2 can start correctly

7. Travis CI

8. cpack based deb generation (example here)

9. conditional inclusion of a floating joint (just for control)

Point 3 is great! Actually I would keep the support for drag-and-drop simulation in gazebo (point 4): why this choice? This can be useful (ROS-free) and easier to debug in my experience.

For the point 5 I would like to have also @aled96 feedback.

6,7 are super, as well as the debian generation (point 8) which we already used successfully on CENTAURO control pc.

[aled96]

Nice work !

5. perception simulation with additional arguments to be provided to the launch file (realsense:=true velodyne:=true)

For the point 5, I think it can be very useful to enable the cameras like this in order to speed up the customization of the robot sensors without the need to create the urdf again.
It is also useful because we have to consider that the realsense d435i slows down the simulation and so it can be better to remove it if not needed.

[alaurenzi]

Point 3 is great! Actually I would keep the support for drag-and-drop simulation in gazebo (point 4): why this choice? This can be useful (ROS-free) and easier to debug in my experience.

The rationale for this choice is twofold

1. running gazebo and then dropping one of our robots is not enough for a correct simulation, since some WorldPlugins would be missing (namely the one for syncing to simulated clock); we would need a gazebo wrapper a la gazebo_ros
2. oftentimes (and for sure in the case of centauro), dropping it into a default world results in a very unstable simulation due to the default value of integration step size (1 ms)

I agree it is useful to have, but requires a bit of additional work that we could tackle at a second time

[liesrock]

Ok I see, it was a plus in the past, but probably no one is using it in a drag-and-drop fashion in these days: the points you mentioned are clear.

[aled96]

For what concerns the velodyne and the realsenses, setting to true the corresponding arguments in the launch file allows to see in gazebo the robot with the sensors (since it uses the xacro), but when we launch xbot the sensors frames are not published since xbot takes as input the urdf (which currently does not contain the cameras).

I copy here also the links of the sensors xacros, if needed:

• https://github.com/m-tartari/realsense_gazebo_description
• https://github.com/m-tartari/realsense_gazebo_plugin
• https://bitbucket.org/DataspeedInc/velodyne_simulator/src/master/

[alaurenzi]

To address this I will generate all urdfs with perception flags set to true. However, this currently means gazebo-specific plugins are also included inside the urdf files (in addition to the sensor-related frames).

This is a bit dirty in my opinion. The only solution I see is to modify the macros _t265, _d435i, and VLP-16 so to expose a parameter that avoids the inclusion of plugins.

Would this work @aled96 ? Do you see any drawback ?

[aled96]

I think that it can be done without much problems. For the realsenses we have the xacro file with the links/joints and then the inclusion of the gazebo.xacro, so with an additional parameter we can simply avoid to include the plugin.

For the velodyne we have a single xacro file but, of course, we can still add a xacro:if clause to avoid considering the gazebo part.

I tried removing the gazebo inclusion from the D435i xacro file and there are no errors, we can see the model of the camera in the robot but no topics are published.