EgoCS-400K: An Egocentric Gameplay Dataset for World Models

Researchers have released EgoCS-400K, a dataset of more than 400,000 first-person Counter-Strike videos spanning 10,000 hours of professional gameplay, designed to train interactive world models that require temporally aligned video, action, and state data [1]. The dataset was built from public professional CS and CS2 match demos that preserve human gameplay trajectories, enabling parsing, replaying, rendering, and temporal alignment [1]. It covers 13 maps and captures 10 player viewpoints per round, drawing from more than 1,000 matches and 40,000 rounds [1]. The creators extracted player states, view directions, movements, keyboard and button inputs, view-angle changes, weapon usage, game events, and round-level context, then rendered clean first-person videos from those same trajectories [1]. The project addresses a structural gap in data for world models. The authors note that web video datasets offer broad visual coverage but lack executable actions and reliable states, while robotic datasets provide action and state supervision but are costly and limited in scene diversity [2]. Existing simulators often lack large-scale human-driven interaction trajectories [2]. EgoCS-400K is positioned as a bridge connecting visual observations with human actions, camera motion, game states, and events at scale [2]. Tasks the dataset supports include action-conditioned future prediction, state- and event-aware scene rollout, replay-grounded captioning, and agent egocentric action understanding [1]. The paper was submitted to arXiv on 16 June 2026 under the Computer Vision and Pattern Recognition category [1]. On the Hugging Face Hub, papers can be linked to models, datasets, and demo Spaces, and the platform extracts arXiv IDs to create paper pages that aggregate related artifacts [4]. The Hub also supports embedding demos directly alongside papers on arXiv abstract pages, allowing users to try models in a browser without writing code [5]. A daily trending papers section on Hugging Face surfaces new submissions to the community [6].