[question] Why are observation and achieved_goal the same in bit flipping env?

In the observation dictionary returned from the bit flipping env, observation and achieved_goal are the same:

https://github.com/hill-a/stable-baselines/blob/3069a0e859afd23879def492b8c57896a6c45faa/stable_baselines/common/bit_flipping_env.py#L70-L80

Aren’t all of the dictionary elements {observation, achieved_goal, desired_goal} fed to the neural network as input? Since observation and achieved_goal are the same, isn’t the neural network being fed the same input twice? Wouldn’t it be better for observation to be an empty list then?

Or do the neural network inputs consist only of observation and desired_goal?

[araffin]

I’m closing this issue as the initial question was answered.

Exactly, so why is a) only correct when the object is always at the same place?

for the reason I mentioned before. It is more for correctness (e.g. if you use another implementation of HER). In practice, with the current version of SB, it does not matter.

[siferati]

Anyway, as mentioned in #745 , the current implementation also gives achieved_goal to the agent.

Exactly, so why is a) only correct when the object is always at the same place? Since achieved_goal is fed to the agent, it doesn’t matter if the object is always at the same place or not, right?