INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models
INTACT maps a desired latent scene change directly into robot actions, avoiding the usual test-time search loop.
The paper frames each transition as an intent signal and trains a JEPA-style model on action-labeled, reward-free trajectories. Its direct policy reaches 85.78%, 100.00%, 97.67%, and 97.89% success on four official LeWM tasks with one epoch and zero search. With an optional local CEM step, it reports 96.86% macro success while using 384 candidate sequences instead of 9,000. Direct inference is reported at 2.9-5.5 ms. HF Daily Papers' note
The paper frames each transition as an intent signal and trains a JEPA-style model on action-labeled, reward-free trajectories. Its direct policy reaches 85.78%, 100.00%, 97.67%, and 97.89% success on four official LeWM tasks with one epoch and zero search. With an optional local CEM step, it reports 96.86% macro success while using 384 candidate sequences instead of 9,000. Direct inference is reported at 2.9-5.5 ms. HF Daily Papers' note
score 5