Note CS:GO deathmatch video with frame-aligned action labels + metadata. A canonical behavioral-cloning / imitation-learning FPS benchmark (basis of the 'Counter-Strike from pixels' work). ~690 downloads, 17 likes.

Note Large egocentric Counter-Strike 2 video dataset from professional matches, 600+ hours of POV footage. Recent (2026-06) and heavily used (27k downloads) for FPS agent/world-model training.

Note CrossFPS: multi-game first-person-shooter dataset, ~69k five-second 20fps clips paired with frame-aligned actions. Good for cross-game FPS imitation learning / generalization.

Note SolarisEngine multi-agent Minecraft world-modeling corpus: 12.64M frames at 20 FPS with actions. Large-scale (3.2k downloads) for action-conditioned world models and Minecraft agents.

Note 73 Minecraft trajectories from single- and multi-agent scenarios (collected via GPT-4o / LLM agents). Compact set of action-labeled Minecraft gameplay for agent training.

Note Frame-action pairs from 17,886 early-game Minecraft videos, processed with Lumine's 5 FPS method. Formatted for vision-language-action (VLA) game agent pretraining.

Note 10M frames + actions from the Atari Pong environment (Bellemare et al. ALE). Part of the p-doom per-game Atari suite (Alien, Breakout, Battle Zone, Demon Attack, etc.) for RL / world-model / IL research.

Note 1.3M human gameplay demonstrations across Atari games (sourced from Zenodo), formatted for vision-language-action training. Human demos rather than RL-agent rollouts.

Note 118,166 frames from 8 Super Mario Bros TAS playthroughs as NumPy .npz with action labels. Small, clean action-conditioned dataset for platformer world models / imitation learning.

Note Dreamer4-formatted Flappy Bird gameplay: frame shards paired with action and reward signals. Ready-to-use for model-based RL / world-model (Dreamer) experiments.

Note Assetto Corsa racing sim: 64M steps from human and Soft Actor-Critic policies in an autonomous-racing Gym. Mixes human demonstrations with RL rollouts for driving-agent training.

Note 56 self-recorded Crusader Kings III sessions capturing raw mouse/keyboard events + screen telemetry. Grand-strategy gameplay for behavior cloning / imitation learning (recent, 2026-07).

Note 234 synchronized video episodes of Pokémon Legends: Z-A gameplay with 26-dimensional control signals. Action-conditioned console-game footage for imitation learning.

Note PokeDreamer v2: 200,000 high-resolution RGB trajectories of Pokémon Red played in a PyBoy emulator. Emulator-generated trajectories for world-model / game-agent training.

Note Synchronized, action-conditioned gameplay videos with frame-accurate input logs. General-purpose gameplay recording aimed at action-conditioned world-model training.

Note 9,385 professional Dota 2 matches parsed from .dem replays into per-second decision context (hero states, cooldowns, combat events, objectives). For Transformer/RL models of MOBA game state; replay-derived rather than pixel gameplay.

Note 1v1 Generals.io replays featuring at least one 70+ star player, formatted for reinforcement-learning agent training on this real-time strategy game.

Note Expert PPO trajectories across all 16 Procgen environments.