Python mjSpec recompilation #1822
Description
Hello Mujoco team,
Loading and compiling a model and then adding a lone geom directly to the world body and then recompiling does not work with the new mjSpec in Python.
Following some examples from the spec_tests.py file the following works:
spec = mj.MjSpec()
spec.compile()
geom = spec.worldbody.add_geom()
geom.name = 'my_geom'
geom.size[0] = 1
geom.pos = [1, 1, 0]
spec.compile()
but when first loading a spec from a xml or a string things do not work
model_string = """
<mujoco>
<worldbody>
<body>
<freejoint/>
<geom size=".15" mass="1" type="sphere"/>
</body>
</worldbody>
</mujoco>
"""
spec = mj.MjSpec()
spec.from_string(model_string)
spec.compile()
geom = spec.worldbody.add_geom()
geom.name = 'my_geom'
geom.size[0] = 1
geom.pos = [1, 1, 0]
spec.compile()
Gives me the following error:
ValueError: Error: incompatible id in geom array, position 1
Element name '', id 0, line 6
Note that spec.recompile does not work either, in this case I get a mj_setPtrData: mjData buffer size mismatch error.
Is this intended? Looking the mjSpec Python tests it seems like none of the test cases preloads a xml and then adds a geom.
Adding a body first and then a geom to that body does work, for example the following as no issues:
model_string = """
<mujoco>
<worldbody>
<body>
<freejoint/>
<geom size=".15" mass="1" type="sphere"/>
</body>
</worldbody>
</mujoco>
"""
spec = mj.MjSpec()
spec.from_string(model_string)
spec.compile()
body = spec.worldbody.add_body()
body.pos = [1, 2, 3]
body.quat = [0, 1, 0, 0]
geom = body.add_geom()
geom.name = 'my_geom'
geom.size[0] = 1
geom.pos = [1, 1, 0]
spec.compile()
On a separate note it seems like mjSpec recompilation in Python does not happen in place? As in the test_recompile test in the spec_tests.py it seems like recompile has to be used like
model_new, data_new = spec.recompile(model, data)
as opposed to just
spec.recompile(model, data)
and the model and data being updated accordingly. This is what I thought would happen based off of the mj_recompile documentation, but it seems like this is not that case.
Is this intended and just a product of having to deal with Python jank or can the recompilation be made in place in the future?
Thank you,
Jeremy