Purpose of split_and_pad_trajectories() and unpad_trajectories() for recurrent neural network

[dxyy1]

Hi there,

Apologies if this question seems obvious but I wonder if and why the two functions split_and_pad_trajectories() and unpad_trajectories() are necessary, if rollout_storage already stores the Memory class' hidden states?

My understanding is that because the RNNs process the observation sequentially step after step, you'd want to split and pad the observations at every done so you don't pass in the a newly reset environment observation with the hidden states from last time step. But I wonder if this issue is already avoided since RolloutStorage class stores the previous timestep's hidden states along with the current timestep's observation, i.e.

hidden states	observation	actions	...
$$\text{hiddenStates}_{t-1}$$	$$obs_t$$	$$a_t$$	...
...	...	...	...