世界アクションモデル、ロボットが動く前に行動をシミュレーション可能に

researchers from fudan university and the national university of singapore published the first systematic framework for world action models, categorizing about 100 papers into cascaded and joint architectures. cascaded wams predict future video and derive commands from it, while joint models like nvidia's cosmos policy process visual input and actions simultaneously. the primary advantage is the ability to utilize unlabeled video data, addressing the bottleneck of expensive teleoperation datasets like open x-embodiment. by simulating outcomes, these models achieve better performance on unfamiliar objects compared to standard vision-language-action models.