Zero-WAM：从人类视频上下文学习世界动作模型

Original title: Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization

Source: arXiv cs.RO·2026-08-28 12:00Publish time unverified

arXiv:2608.26103v2 Announce Type: replace Abstract: Zero-shot cross-task generalization, where a policy must execute manipulation tasks never seen during training, remains a central challenge in robot learning. In large language models, a novel task can be performed simply by specifying it in the context, without any parameter update. This form of in-context learning (ICL) turns generalization into a problem of task specification.