CS2-10k: A Large-Scale Egocentric Counter-Strike 2 Dataset (reka.ai)

CS2-10k has been launched as a groundbreaking dataset tailored for research in embodied AI and world-model training, harnessing the resources from professional matches of Counter-Strike 2. This extensive dataset includes over 600,000 player-round videos, collecting more than 10,000 hours of first-person footage at high resolution, paired with detailed per-frame annotations that capture keyboard inputs, mouse movements, and player trajectories. The precise synchronization of video and action signals enables researchers to examine how visual changes occur in response to player actions in a controlled environment, demystifying the interaction dynamics crucial for training effective AI systems. The significance of CS2-10k lies in its potential to streamline the development of interactive world models, which are essential for creating AI that can understand and manipulate complex environments. By providing a consistent and richly annotated dataset, it opens up numerous avenues for experimentation in AI research and development, especially within the gaming domain, where rich visual and behavioral data is typically difficult to obtain. Additionally, the release of the open-source cs2-dem-renderer pipeline further empowers the community to generate their own datasets tailored to specific needs, making it easier to explore the nuances of embodied intelligence in action.