INTACT swaps latent world-model search for millisecond policy

A quiet trend in robotics research is the push to make latent world models actually controllable without paying for expensive rollouts at inference, and a new preprint on arXiv from Junhan Sun, Hao Zhao and Guofeng Zhang leans hard on that idea. Their system, INTACT, converts action-labeled trajectories into what the authors call an intent-to-action interface, so a desired change in the latent scene maps directly to an action instead of being recovered by test-time search.

The headline numbers, taken from the paper's own abstract, are that one-epoch, zero-search models reach 85.78%, 100.00%, 97.67% and 97.89% success on the four official LeWM tasks. Keep search around as optional verification and a local CEM step centered on the direct plan hits 96.86% macro success using 384 candidate sequences instead of 9,000, which the paper reports as a 23.44x reduction in sampling and a 16.00-point gain over pure CEM. Direct inference takes 2.9 to 5.5 milliseconds.

Why that combination matters: latent world models have been useful research artifacts but awkward controllers, because recovering an action for a desired scene change usually meant running expensive test-time search. If the intent-to-action mapping can be read in a few milliseconds, the same models start to look deployable inside real-time control loops rather than as planners you consult occasionally.

The honest caveat is that these are self-reported LeWM benchmark results in a single preprint by one team, with no independent replication yet and no hardware demonstration in the abstract. The paper also leans on a distributional action law whose conditional mean serves as the policy, and how gracefully that fails on genuinely multi-modal action distributions is not something the abstract answers.

Still, the direction is the interesting part. If search-free intent-to-action generalizes off LeWM, robotics teams that already train latent world models get a millisecond-latency controller for close to free, and the cost of putting a world model on real hardware drops in a way that plain scale-and-search approaches have struggled to match.