Temporal Augmentations for Streamed Video Games: Supplementary Material

July 11, 2026 - 06:23

This supporting website is linked to the research paper "Augmentations for Robust and Efficient Imitation Learning in Streamed Video Games," which was presented at the Conference on Games 2026. The study explores how spatiotemporal augmentations that simulate typical streaming artifacts such as pixelation, blur, scrubs, and ghosting can boost the sample efficiency and robustness of imitation learning agents trained from limited data.

The paper investigates whether artificially adding these common video stream distortions during training helps agents better handle real-world streaming conditions. By mimicking the visual noise and temporal inconsistencies found in live game streams, the researchers aimed to make imitation learning models more resilient. The supplementary material provides additional experimental results, implementation details, and visual examples of the augmentations used. It includes ablation studies that break down the contribution of each augmentation type, as well as comparisons against baseline methods without these temporal adjustments. The findings suggest that carefully designed augmentations can significantly reduce the amount of demonstration data needed while improving the agent's ability to generalize to unseen streaming scenarios. This work has potential applications in game AI, cloud gaming, and any domain where video data is transmitted over unreliable networks.