INTACT: End-to-End JEPA Enables Search-Free Intent-to-Action World ModelsINTACT：端到端 JEPA 实现无搜索意图到动作的世界模型
S 1.7T11 sources1 个来源R7-research
Junhan Sun, Hao Zhao, and Guofeng Zhang (submitted 28 Jul 2026) introduce INTACT (INtent-To-ACTion), an end-to-end JEPA world model that turns action-labeled, reward-free trajectories into a deployable intent-to-action interface, removing the expensive test-time search that prior forward latent world models require to recover actions for a desired scene change.
INTACT derives two training signals from the same trajectories — a "physical intent" z_{t+1}-z_t from each observed transition and a "deployment intent" sg(z_g)-z_t from a future goal — processed through an isomorphic four-slot graph architecture shared between the local-motion and goal-motion branches.
For autonomous driving, cutting test-time search out of world-model-based planning matters because real-time vehicle control needs low-latency inference, making this search-free intent-to-action mapping directly relevant to trajectory planning under tight decision-latency budgets.