rocket-science
收藏资源简介:
该数据集由Kyutai等机构联合创建,专门用于多智能体世界模型的研究,聚焦于高度动态的物理交互环境。数据集包含长达1万小时的《火箭联盟》游戏对战视频,涵盖四名玩家的渲染画面、控制器操作流及特权游戏状态数据,为模型训练提供了丰富的多视角交互样本。数据采集通过强化学习策略在私人比赛中完成,确保了动作日志的精确性和大规模可行性,旨在解决复杂多智能体环境中动作归因与物理一致性建模的挑战,为可控仿真、多智能体训练及规划决策等应用奠定基础。
This dataset was co-created by Kyutai and other institutions, specifically designed for research on multi-agent world models and focused on highly dynamic physical interaction environments. It contains 10,000 hours of competitive gameplay footage from Rocket League, including rendered frames, controller input streams, and privileged in-game state data for four-player matches, providing rich multi-view interaction samples for model training. The data was collected using reinforcement learning policies in private matches, ensuring the accuracy and large-scale feasibility of action logs. This dataset aims to address the challenges of action attribution and physical consistency modeling in complex multi-agent environments, laying a solid foundation for applications such as controllable simulation, multi-agent training, and planning and decision-making.




