INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

作者：Junhan Sun, Hao Zhao, Guofeng Zhang · 单位：State Key Laboratory of CAD&CG, Zhejiang University, Institute for AI Industry Research (AIR), Tsinghua University, RoboParty Lab, collapse and preserves a unified multi-task intent space. Center: one shared, JEPA that turns action-labeled, reward-free trajectories into a deployable, sampling remains available for diversity or optional verification rather than, centered on the Direct plan reaches 96.86% macro success using 384 instead of · 会议/期刊：arXiv preprint · 方向：cs.RO · 发布日期：2026-07-29