mbp: Support programmatic composition of "sub-plants"?

[EricCousineau-TRI]

Composition

It would be nice to be able to assemble different plants together, i.e. throw them both into the same world, and then throw in some welds.
This would most likely require #12703 (the topology solution proposed in the issue).

Use Cases

• Compose MultibodyPlants when parsing, without having to do weird gymnastics to remember hierarchy
  • As part of composition/proposal: Expand on interface-type API gazebosim/sdf_tutorials#3, I wrote a WIP integration in Drake; here are lines of me complaining about this ;) - https://github.com/EricCousineau-TRI/drake/blob/ee010353b3a44dc5e7bd1e18efad8479e1d4b528/multibody/parsing/detail_sdf_parser.cc#L432-L437

[sherm1]

@EricCousineau-TRI a few thoughts:

• I'm imagining that you would need also to transfer part of the larger plant's state to the sub-plant's state. Is that part of the decomposition task?
• These seem like two unrelated tasks. Reading in multiple models and welding them together before doing anything with them seems very easy. Extracting sub-pieces of a plant, determining that they form a coherent connected set, and possibly transferring state seems MUCH harder. If I've understood that correctly please consider filing separate issues so we can prioritize and scope separately.

[ggould-tri]

This is a really nice idea -- it would go a long way toward cutting the knot on #12849 . However as @sherm1 notes it would also incur all of the complexity that motivates the 12849 case -- the correct shuffling and interpretation of state and the implementation of state readback into the parent -- and it would lay a serious landmine under any logic that uses fortran style "offset into flat vector" logic to address mbp states (ie, nearly all of it 😭 😡 ).

[EricCousineau-TRI]

Sorry for the lag! I will try to break this apart as requested by EOD today.

[EricCousineau-TRI]

Will update this issue to handle the harder task (decomposition). Will make a new issue for the easier task (composition).
Will edit this post once I'm done.

EDIT: Done #13177

[EricCousineau-TRI]

MATLAB version: https://www.mathworks.com/help/robotics/ref/rigidbodytree.addsubtree.html

[sherm1]

Lowering priority for now since this was a lesser issue for users per the Mutating Multibody Models meeting.

[EricCousineau-TRI]

Per convo, also should be achievable by current public API, without having to a method on MultibodyPlant:
Drake Dev Meeting Agenda, 2020/05/11

[EricCousineau-TRI]

I am working on a prototype of this in Python, as part of #13133

Similar to what Robert ran into, I'm seeing parsing taking up at least 50% of my team for clutter gen. I want to see if I can cut that down by some sort of subgraph decomposition and assembly and stuff.

[EricCousineau-TRI]

I think I have a prototype in Anzu that covers the workflows I want. See here for a copy-pasta version:
MultibodyPlantSubgraph (permalink)

This can handle composition via sub-plants.
Currently, the Anzu PR is in place, and will let it bake for a few weeks there per Jeremy's suggestion.

[EricCousineau-TRI]

See user pain point that could be partially addressed by the "subgraph" approach: https://stackoverflow.com/questions/65583969/visualizing-an-autodiff-multibodyplant-in-pydrake#comment115962920_65584221