Confusion about the intention

[zichunxx]

Hi! @andrew-j-levy

I'm new to reinforcement learning and interested in your work.

After I read your article thoroughly, I'm confused about the intention to solve the long horizon task with the goal-conditioned reward scheme.

In my opinion, the goal-conditioned reward can be treated as the sparse reward, which performs badly in long horizon tasks.

Thus, why not use the dense reward with differentiable functions which can lead the training process to convergence? Sometimes, some tasks don't require a lot of goals.

I don't know if I'm on the right point and this may seem meaningless to you, but I'd like to get a response from you.

Thanks!