Hindsight Experience Replay (HER)

Re-examined failed trajectories in multi-goal RL by relabeling the achieved final state as the intended target goal, learning successfully from sparse binary failure rewards.

Serves as qualified theoretical and systems foundation for task-reinforcement-learning, task-robotics.

Assumptions

• Markovian state dynamics and stationary reward functions hold in target evaluation environments

Limitations

• Sample efficiency, exploration stability, and real-world sim-to-real transfer gaps require specialized tuning