We have seen how experience replay is used in DQN to avoid a correlated experience. Also, we learned that prioritized experience replay is an improvement to the vanilla experience replay as it prioritizes each experience with the TD error. Now we will look at a new technique called hindsight experience replay (HER), proposed by OpenAI researchers for dealing with sparse rewards. Do you remember how you learned to ride a bike? On your first try, you wouldn't have balanced the bike properly. You would have failed several times to balance correctly. But all those failures don't mean you didn't learn anything. The failures would have taught you how not to balance a bike. Even though you did not learn to ride the bike (goal), you learned a different goal, that is, you learned how not to balance a bike. This is how we humans learn, right? We learn from failure, and this is the idea behind hindsight experience replay.

Let's consider the same example given in the paper. Look at the FetchSlide environment as shown in the diagram; the goal in this environment is to move the robotic arm and slide a puck across the table to hit the target, a small red circle (diagram from https://blog.openai.com/ingredients-for-robotics-research/):

In few first trails, the agent could not definitely achieve the goal. So the agent only received -1 as the reward, which told the agent it was doing wrong and that it did not achieve the goal: 

But this doesn't mean that agent has not learned anything. The agent has achieved a different goal, that is, it has learned to move closer to the actual goal. So instead of considering it a failure, we consider that it has a different goal. If we repeat this process over several iterations, the agent will learn to achieve our actual goal. HER can be applied to any off-policy algorithms. The performance of HER is compared by DDPG without HER and vice versa and it is seen that DDPG with HER converge quicker than DDPG without HER. You can see the performance of HER in this video: https://youtu.be/Dz_HuzgMxzo.