OnDeFog Enhances Online RL Performance in Frame-Dropping Environments

In real-world reinforcement learning scenarios, agents often face challenges like communication delays or sensor failures, leading to "frame dropping" where critical state and reward information is lost. While the offline Decision Transformer under Random Frame Dropping (DeFog) was developed to mitigate this, its offline nature limits its ability to generalize to new, unseen states. To address this, researchers have proposed OnDeFog, an innovative online reinforcement learning approach. OnDeFog integrates the robust frame-dropping mechanisms of DeFog with the Online Decision Transformer (ODT), allowing it to learn policies through direct interaction with the environment. Extensive experimental evaluations show that OnDeFog significantly outperforms ODT in environments characterized by high frame dropping rates. Furthermore, it demonstrates superior performance compared to DeFog when trained on datasets rich in low-reward data, highlighting its enhanced adaptability and generalization capabilities in challenging, dynamic settings.