Fixing qpos order in the xml of the basketball task.

[devinluo27]

I found that the order to initialize mujoco bodies in the xml of the basketball v2 task is slightly different with others. This results in unexpected initial robot poses. An example is shown below:

In this function, the object is assumed to follow the robot qpos, however, in the xml for the basketball file, the order is inconsistent, which leads to unexpected initial pose of the robot.

Metaworld/metaworld/envs/mujoco/sawyer_xyz/sawyer_xyz_env.py

Lines 248 to 254 in 87ac948

	def _set_obj_xyz(self, pos):
	qpos = self.data.qpos.flat.copy()
	qvel = self.data.qvel.flat.copy()
	qpos[9:12] = pos.copy()
	qvel[9:15] = 0
	self.set_state(qpos, qvel)

In this pull request, I put the basketball behind the robot in the xml, and this issue seems to be resolved.

[reginald-mclean]

@devinluo27 what version of mujoco are you using?

[devinluo27]

@devinluo27 what version of mujoco are you using?

I installed mujoco via pip, and pip list returns mujoco 2.3.5

[reginald-mclean]

@devinluo27 can you try 2.3.7?

[devinluo27]

@devinluo27 can you try 2.3.7?

I upgraded to 2.3.7 just now, and the results turned out to be identical to mujoco 2.3.5.

[reginald-mclean]

Interesting. I'll look into it tomorrow, thanks!

[pseudo-rnd-thoughts]

Does this problem affect any more environments?

[devinluo27]

Does this problem affect any more environments?

I haven't looked into every task closely, but so far I only encountered this problem in this basketball environment.

[reginald-mclean]

I manually checked the XMLs of each task, the only task where the xyz_base.xml file is also loaded in a different order is for the coffee environments. Changing the order screws up the environment completely. As far as I can tell, basketball is the only affected environment