Mujoco-py: Meaning of the observation list

[theobdt]

Hi,
Is there a way of knowing the "meaning" of the elements from the observation list ? There is good documentation for the basic environments :
https://github.com/openai/gym/wiki/CartPole-v0
But I can't find something similar for the mujocos envs. Thank you very much.

[BrunoBSM]

I agree, the Mujoco envs don't have such good doccumentations. If you've played around with mujoco-py you may recognize some names within the _get_obs() function (like in the ant environment https://github.com/openai/gym/blob/master/gym/envs/mujoco/ant.py ), like qpos and qvel which correspond to cartesian positions, rotation quartenions and velocities for joints. On other environments there are some names I do not recognize like cfrc. In order to find out what they are I recommend looking at the MuJoCo doccumentation http://www.mujoco.org/book/ and playing around with mujoco-py's sim.data and sim.model with a known simmulation you know how many and where are its joints. By using variables = [i for i in dir(sim.model)] and then printing variables you can see how many variables and functions are inside each one for you to read or write. Hope this helps, and I hope someone comes with a much simpler way.

[machinaut]

@theobdt this package is the python bindings to the mujoco library, so really just focused on physics simulation.

Reinforcement Learning environments (and their observations, etc) mostly live in different packages.

If this question is about an environment, it'd be better to ask that on the repo for that environment.

Similarly, MuJoCo is documented in its own book here: http://www.mujoco.org/book/index.html and if you have questions about the mujoco library, there is a forum here http://www.mujoco.org/forum/index.php