New Augmentations Boost AI Agent Robustness in Streamed Video Games

Training AI agents for complex 3D video games using imitation learning from human demonstrations is a promising approach. However, collecting these demonstrations is costly, and modern game streaming introduces visual artifacts like pixelation, blur, and ghosting due to network delay and compression. These artifacts can cause a significant shift in data distribution, leading to performance degradation for agents at test time. Researchers propose novel "streaming augmentations" designed to simulate these common network-induced visual artifacts. By training agents with these augmentations, the goal is to make them more resilient to the real-world conditions of streamed gameplay. The method was applied to predictive inverse dynamics models (PIDM) and tested across three different 3D video game tasks. Results show that agents trained with these spatiotemporal augmentations achieved up to 41% higher performance under stable streaming conditions compared to those without augmentations, given the same data budget. Crucially, when network lag was introduced, augmented agents experienced only a 7.45% performance drop, whereas unaugmented agents saw a nearly 50% reduction. This highlights the augmentations' effectiveness in creating robust and efficient game-playing AI.