EgoCS-400K: 400K Counter-Strike Trajectories for World Model Training
June 15, 2026
EgoCS-400K provides 400K egocentric video-action-language trajectories extracted from professional CS and CS2 match replays, targeting the data gap for interactive world models that require temporally aligned actions, camera motion, and game state. Unlike web video datasets, trajectories include executable actions and reliable state annotations.
HOW THIS AFFECTS YOU
●
researcherDirectly usable as a benchmark or pretraining corpus if you're working on action-conditioned video generation or game world models; the replay-grounded extraction pipeline is also worth examining for other simulator datasets.