迈向物理接地 JEPA 世界模型的目标条件机器人规划

Original title: Toward Physically Grounded JEPA World Models for Goal-Conditioned Robotic Planning

Source: arXiv cs.RO·2026-09-04 12:00Publish time unverified

arXiv:2609.03565v1 Announce Type: new Abstract: Action-conditioned JEPA world models enable planning toward visually specified goals without reconstructing future pixels, yet latent prediction alone does not explicitly encourage the learned representations to retain information relevant to robotic control. We introduce an end-to-end JEPA world model that augments latent prediction with inverse dynamics (IDM) and state alignment (SA).