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

2026-07-28

AI Summary

This work addresses the inefficiency of traditional forward latent-variable world models, which require costly test-time search to infer actions from target states. The authors propose INTACT, a novel framework featuring an intention-action isomorphic architecture that leverages end-to-end Joint Embedding Predictive Architecture (JEPA) to directly map reward-free trajectories into intention-to-action policies. By exploiting the isomorphic graph structure between physical dynamics and goal intentions—combined with a four-slot syntactic graph, action-law semantics, and asymmetric gradients—the method enables search-free control. Evaluated on four LeWM tasks, INTACT achieves up to 100% success rate after a single training round, with an average success rate of 89.39% under search-free inference, requiring only 2.9–5.5 ms per decision and improving sample efficiency by 23.44×.

Abstract

Forward latent world models predict how actions change a scene, but recover actions for a desired change only through expensive test-time search. We introduce INTACT (INtent-To-ACTion), an end-to-end JEPA that turns action-labeled, reward-free trajectories into a deployable intent-to-action interface. Each transition supplies physical intent $z_{t+1}-z_t$, while a future goal supplies deployment intent $\operatorname{sg}(z_g)-z_t$. The architecture is isomorphic between the local and goal motion-intent backbone-input graphs through an identical four-slot grammar and shared parameters, and between supported local and goal motion-intent families through action-law semantics induced by the same predictor rather than pointwise latent equality. INTACT also provides intact transfer from RGB evidence to action-effective latent intent coordinates and from intent families to their corresponding action-law families. Asymmetric endpoint gradients ground physical successors and fix future goals as anchors, joining representation learning and control without pointwise latent matching or globally linear dynamics. The resulting coordinates support a robust distributional action law: its conditional mean serves directly as a search-free policy, while sampling remains available for diversity or optional verification. On the four official LeWM tasks, one-epoch, zero-search models reach 85.78\%, 100.00\%, 97.67\%, and 97.89\% success. Optional local CEM centered on the Direct plan reaches 96.86\% macro success using 384 instead of 9,000 candidate sequences, reducing sampling by $23.44\times$ while improving pure CEM by 16.00 points. One 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 $r=0.954$. Direct inference takes 2.9--5.5 ms.

Problem

• world models
• search-free control
• intent-to-action
• latent dynamics
• action inference

Innovation

• search-free control
• isomorphic intent-to-action
• joint embedding predictive architecture
• latent intent coordinates
• action-law semantics