mujoco-maze icon indicating copy to clipboard operation
mujoco-maze copied to clipboard

Published 20 hours ago • verified publisher icon kngwyu

Maze Ant observation shape

Open pavlosSkev opened this issue 4 years ago • 2 comments

Hello. I am just curious why the observation space of Maze Ant has a shape 30? The original Ant has 111. Is something excluded? If yes, how could I use the full observation space?

Thanks in advance.

pavlosSkev avatar Dec 12 '20 16:12 pavlosSkev

  • The first 29 elements are Mujoco model's qpos and qvel. I'm sorry but this is inherited from the implementation of HIRO paper (https://github.com/tensorflow/models/blob/master/research/efficient-hrl/environments/ant.py#L81) and I don't know the design choice precisely. Maybe for including the coordinates?
  • The last one is timestep. I'm going to make this optional in the future release.

kngwyu avatar Dec 13 '20 13:12 kngwyu

  • The first 29 elements are Mujoco model's qpos and qvel. I'm sorry but this is inherited from the implementation of HIRO paper (https://github.com/tensorflow/models/blob/master/research/efficient-hrl/environments/ant.py#L81) and I don't know the design choice precisely. Maybe for including the coordinates?
  • The last one is timestep. I'm going to make this optional in the future release.

Thank you for your answer. It's difficult to understand the observation space as everything leads back to the MujocoEnv class (at least from my attempt). Looking forward to your future releases. Cheers!

pavlosSkev avatar Dec 13 '20 18:12 pavlosSkev