Understanding Gym Mujoco observation space

Discussion in 'Modeling' started by Alexis Duburcq, Aug 26, 2020.

  1. Hello,

    I'm trying to understand the meaning of the various quantities gathered in the observation space of some advanced environment like Humanoid, namely:
    - data.qpos.flat
    - data.qvel
    - data.cinert
    - data.cvel
    - data.qfrc_actuator
    - data.cfrc_ext

    Most of them are pretty clear to me, but not 'cinert' nor 'cvel'. According to the documentation, 'cinert' is "com-based body inertia and mass (nbody x 10)" and Todorov said on another post that it corresponds to "the mass and inertia of a single rigid body relative to the com of the subtree to which the body belongs". I get this more or less but more information about this quantity would be helpful. Regarding 'cvel', the only description I could find is "com-based velocity [3D rot; 3D tran] (nbody x 6)". I don't understand the meaning of the rotation part, since for me the center of mass is a position only. Besides, for both quantities the frame in which they are expressed is not mentioned, which is problematic.

    I know that those quantities "are intermediate results that were not meant to be documented and exposed to the user", as Todorov said, but since it is a key feature of extremely important benchmark in the field of reinforcement learning, it is very critical to make sure everybody can find extensive information about them.