Forward latent world models predict how actions change a scene, but recovering actions for a desired change typically requires expensive test-time search. Researchers from Hugging Face introduce INTACT (INtent-To-ACTion), an end-to-end JEPA that turns action-labeled, reward-free trajectories into a deployable intent-to-action interface.
INTACT learns the missing intent-to-action correspondence end to end. A shared conditional operator maps both realized local transitions and deployable goal intents to their induced action distributions. The two intent instances share the same predictor and action supervision, while using deliberately asymmetric gradient paths. This grounds goal-conditioned control in observed physical transitions without forcing pointwise latent alignment or globally linear dynamics.
After only one training epoch, INTACT achieves 95.33% Direct macro success rate across four visual-control tasks, with zero candidate search and 2.9–5.5 ms inference latency. This reduces planning latency by approximately 300× relative to broad CEM control. INTACT also supports joint four-task training with a shared visual encoder and lightweight task-specific heads.
Optional local verification with CEM centered on the Direct plan reaches 96.86% macro success using only 384 candidate sequences instead of 9,000, reducing sampling by 23.44× while improving pure CEM by 16.00 points. A shared four-task encoder reaches 89.39% E5 Direct macro and improves every task over jointly trained LeWM, while predicted–expert action-family kNN tracks Direct success at r=0.954.
The architecture is isomorphic between local and goal motion-intent backbone-input graphs through an identical four-slot grammar and shared parameters, and between supported intent families through action-law semantics induced by the same predictor. INTACT provides intact transfer from RGB evidence to action-effective latent intent coordinates and from intent families to their corresponding action-law families.