The observation in the env
Closed this issue · 2 comments
Hello,
I have a question of fetchreach_env.
In fetchreach_env,i don't konw the gripper_vel and gripper_state really mean,because the gripper doesn't need to open or close.Mabye those obs are useful in pickandplace_env.
And I try change obs in fetchreach_env.
Like this :obs = grip_pos.
Just the coordinates of the grip,and exclude the gripper_vel and gripper_state, then I use HER to train a model ,the success rate after 10 epochs is 1.
So, in fetchreach_env do I really need the gripper_vel or gripper_state.
Thank you.
I exclude the grip_velp as well,the success rate are the same.
So,how the velocity actually work in the env.
Hello 😃
First of all, since your question is about the physical property of the environment I strongly recommend you to ask it from the mujoco developers here:
https://github.com/openai/mujoco-py
Secondly, take a look at this issue. It's been discussed here:
openai/gym#1503
But at last, if you want my opinion:
I believe since these approaches (such as DDPG) have been innovated in Model-free RL field, so any non-applicable input feature would be ignored by weights of its related neural network.