The Bleeding Edge Gameplay Sample dataset was released by Microsoft Research in 2025. It includes operation trajectories, video frames, and corresponding action data for the multiplayer online game "Bleeding Edge," aiming to provide high-quality synthetic gaming data for reinforcement learning and computer vision research.

This dataset comprises 1,024 sixty-second gameplay clips (totaling approximately 75 GB), preprocessed into videos at a resolution of 300x180 pixels sampled at 10 fps.

Dataset Composition

The dataset primarily consists of 1,024 sixty-second gameplay trajectory records used for model evaluation.

Data Instances

Each data instance typically consists of one .npz file and one .mp4 video file:

• .npz files contain decompressed video frames, human player actions (actions) and game states (states). Specifically, actions have dimensions (610, 16), states have dimensions (610, 56), and images have dimensions (610, 3, 180, 300).
• .mp4 files correspond to the contents of their respective .npz counterparts, providing a convenient visual representation over 60 seconds.