In this study, we examine the robustness of Decision Transformer on Atari games. By replacing the state data of offline reinforcement learning with noise-augmented data, we evaluate the learning performance under these conditions. In the evaluation experiments, we compared the performance of four different Atari games (Breakout, Pong, Qbert, and Seaquest) across five evaluation tests (Clean, Gaussian, Shot, Impulse, and Speckle) scores. Experimental results showed that game scores were lower for the normal data (Clean) in all noise evaluation tests. However, when noise-augmented data learning was introduced as a countermeasure, game scores tended to improve for the Pong game. These results indicate that decision transformers in Atari games are vulnerable to noise evaluation tests and that noise-augmented data training can improve the robustness, particularly for the Pong games.