Add comprehensive testing strategy for real-world deployment on iCub and ErgoCub

[flferretti]

The current testing framework for jaxsim lacks validation against the actual robots intended for deployment. This issue aims to address this gap by establishing a comprehensive testing strategy that ensures the reliability and performance of the simulator in real-world scenarios.

Perhaps we can test against also the iCub/ergoCub models, as most probably those are the one we are going to use the feature against?

Kind of unrelated to the PR, but just because I realized it: testing real world models is ok, but in my experience you may have some blinds spots. In iDynTree I historically found many problems with programmatically generated random models, generated by getRandomModel function.

It could make sense to do something similar here, or also directly use it here as we can export the randomly generated model to URDF, something like

import idyntree.bindings as idt
nrOfJoints = 20
randomModel = idt.getRandomModel(nrOfJoints)
mdlExport = idt.ModelExporter()
mdlExport.init(randomModel)
urdfString = "dummy"
urdfFilePath = "/dummy/dummy"
mdlExport.exportModelToString(urdfString)
# or
mdlExport.exportModelToFile(urdfFilePath)

Originally posted by @traversaro in #75 (comment)

[diegoferigo]

I agree about testing on iCub/ErgoCub models, but note that the JIT performance we currently have would make the tests take really a lot of time, and they are already very slow. This was the original reason to test smaller fixed-base and floating-base models.

[traversaro]

I agree about testing on iCub/ErgoCub models, but note that the JIT performance we currently have would make the tests take really a lot of time, and they are already very slow. This was the original reason to test smaller fixed-base and floating-base models.

Ah, good point. I was not considering that.

[diegoferigo]

In #106, I'm testing -among other models- a reduced version of the ErgoCub robot. It could be a good tradeoff between CI time and comprehensiveness of the testing strategy.

[diegoferigo]

Furthermore, random models could be created with rod similarly to the existing box and sphere models.

[diegoferigo]

The testing suite targeting the new functional APIs test against a reduced version of ErgoCub (only the DoFs of the legs are kept). It seems a acceptable trade-off between the comprehensiveness of the tests and the time necessary to run them (~20 mins at the time of writing).

I guess we can close this issue now.