SDF Loader

[johnwason]

I have started investigating loading SDF robot definition files and have found that the standard sdformat library that comes with Xenial and Bionic is very primitive and does not provide functionality beyond @Levi-Armstrong 's URDF parser. My suggestion is that we modify the URDF parser to be able to parse both URDF and SDF files since there is a lot of overlap. I will talk to RPI about assigning a student to this task. @Levi-Armstrong do you agree with this approach?

[johnwason]

Looping in @wlawler45 and @controlguy168

[Levi-Armstrong]

I think this is a good approach to support both for now. We did have a discussion at ROSCon about the future of URDF and should we migrate to SDF or and standard format like STEP, etc.. Based on the discussion I believe it maybe leaning toward SDF, but there is still more discussions to be hade, but for the sake of Tesseract I am on board with supporting both.

[johnwason]

After doing more research on SDF, it looks like there are some compatibility issues.

The first issue is the "world" and "model" concept used by SDF. SDF breaks up the scene into hierarchical units, with namespaces. For instance, the ljoint of a joint would be "my_robot::joint1" instead of a global name as in URDF. Models can contain models, allowing for nested designs. This is a cleaner design, but this model hierarchy is separate from the link/joint graph. Joints can jump across models, for instance attaching a gripper to a robot: http://gazebosim.org/tutorials/?tut=attach_gripper . In this example, the following joint jumps between models:

<joint name="hokuyo_joint" type="fixed">
  <child>hokuyo::link</child>
  <parent>simple_gripper::palm</parent>
</joint>

This format breaks the current scene graph design, and it doesn't look like subgraphs will help. (There is no way to attach properties to subgraphs, or query which subgraphs a vertex belongs to.)

The second issue is the frame semantics as described here: http://sdformat.org/tutorials?tut=pose_frame_semantics . (Note that the "frame" attribute and "frame" tag are not currently supported.) Frames are currently always defined in relation to the parent "model", not to the parent link/joint. This makes loading into tesseract more difficult.

Is SDF still the frontrunner for replacing URDF?

[gavanderhoorn]

The second issue is the frame semantics as described here: http://sdformat.org/tutorials?tut=pose_frame_semantics . (Note that the "frame" attribute and "frame" tag are not currently supported.) Frames are currently always defined in relation to the parent "model", not to the parent link/joint. This makes loading into tesseract more difficult.

I'm not entirely sure, but I believe (at least some of) this is being looked at: Pose frame semantics: proposed behavior (Discourse post here).

[Levi-Armstrong]

I looked into boost subgraph and it is just another adjacency list so is support its own subgraph properties. For example the subgraph name property could be the namespace. I think between the main SceneGraph and the addition of the SubSceneGraph we could accomplish the support for SDF. @johnwason If you kinda outline how you think a user need to interact with the main SceneGraph, example getSubGraphLinks, getSubGraphLinkName, ect. I will try and mock up another version of the SceneGraph that leverages boost subgraph.

[johnwason]

Because of the way the models and frames work, there are 2-3 overlapping graphs that describe the joint/link topology, model topology, and frame topology. I am not sure how we can capture that using subgraphs.

[Levi-Armstrong]

Do you know if there is a good diagram showing the structure of the different components for SDF?

[johnwason]

No, not that I have found. The functionality of SDF seems to be a work in progress. I am planning on writing a simple importer using Python for our use, so that may proved some insight. It will manage the models and frames outside of tesseract.

[johnwason]

Another possibility is to use boost::variant or have nullable properties attached to the vertex/edges in the graph, then filter for the scenario we care about. For instance, if we just want the joint/link graph we would filter out any vertex/edge that has null link/joint.

[Levi-Armstrong]

Are you planning to write your own parser or use the sdf library?

[johnwason]

The SDF library would be preferable, but we would need at least version 1.8. Version 1.6 which comes with Ubuntu doesn't really do much more than a typical XML library. Version 1.8 and higher fully parses the scene but will need to be built from source or use the OSRF APT repo.

If I write a parser in Python it will probably be custom.

[johnwason]

I have been looking into this again recently. @Levi-Armstrong do you mind if we add a newer version of libsdformat to tesseract_ext? Version 6 (included with ROS) is very limited, it doesn't parse out all the various elements. The current Version 9 parses the entire scene and would be relatively easy to load into tesseract.

https://bitbucket.org/osrf/sdformat/src/sdformat9_9.1.0/

[Levi-Armstrong]

I do not have an issue.