Instant Episode Repetition Boosts Reinforcement Learning Sample Efficiency

In human learning, repetition of successful experiences is a fundamental mechanism for strengthening memory and consolidating skills. Drawing inspiration from this biological principle, researchers have developed Instant Episode Repetition (IER), a new technique designed to improve sample efficiency in reinforcement learning (RL). IER operates by directly influencing the data collection process. When an agent completes a high-reward episode, its action sequence is immediately repeated for a set number of subsequent episodes. This active repetition differs from conventional methods like Experience Replay or Self-Imitation Learning, which passively reuse past experiences during training updates. By re-engaging with the environment using successful sequences, IER effectively reinforces valuable behaviors. The effectiveness of IER was evaluated by integrating it into state-of-the-art SAC and TD3 algorithms. Experiments on continuous-control benchmarks, including MuJoCo, DeepMind Control Suite, and a real-world robotic manipulation task, demonstrated that this simple mechanism significantly improves learning performance compared to standard and self-imitation-based baselines.