AHA-WAM:Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing

2026-06-08

AI Summary

Existing world-action models struggle to simultaneously achieve long-horizon scene understanding and real-time control due to their coupling of perception and action at the same temporal resolution. This work proposes an asynchronous, horizon-adaptive world-action modeling framework that decouples the temporal rhythms of perception and action through a dual-diffusion Transformer architecture. The approach incorporates observation-guided video context routing (OVCR), rolling key-value memory, and inter-layer joint attention mechanisms. Evaluated on RoboTwin, the method achieves an average success rate of 92.80%, with 78.3% success across four real-world tasks, operating at a closed-loop frequency of 24.17 Hz—4.59× faster than Fast-WAM—without requiring any pretraining on robotic data.

Abstract

World-action models have emerged as a promising paradigm for robot manipulation, jointly modeling visual scene dynamics and actions to inject physical priors into policy learning. However, existing world-action models couple world prediction and action execution at the same temporal resolution, forcing the world branch to model near-term frame variations that are redundant and weakly informative. We posit that strictly binding world prediction and action execution to the same temporal rhythm may underutilize the potential of the video branch for embodied control. Therefore, we propose AHA-WAM, an Asynchronous Horizon-Adaptive World-Action Model built on a dual Diffusion Transformer (DiT) architecture that reorganizes world-action modeling around this temporal asymmetry. AHA-WAM instantiates the video DiT as a low-frequency world planner that maintains rolling key-value memory over past observations and exposes reusable layerwise latent context encoding long-horizon scene evolution, while a high-frequency action DiT executes short action chunks in closed loop by querying this context through layerwise joint attention. To support asynchronous execution, we introduce horizon-adaptive offset training and Observation-Guided Video-Context Routing (OVCR), which together let the action expert exploit long-horizon world context while remaining responsive to real-time execution state without rerunning the video DiT. Experiments on RoboTwin and real-world manipulation tasks show that AHA-WAM achieves state-of-the-art performance without any robot-data pretraining, attaining 92.80% average success on RoboTwin and 78.3% success across 4 real-world tasks, while reaching 24.17 Hz closed-loop control with a 4.59x speedup over Fast-WAM.

Problem

• world-action modeling
• temporal resolution
• asynchronous execution
• robot manipulation
• video context

Innovation

• Asynchronous World-Action Modeling
• Horizon-Adaptive Planning
• Observation-Guided Context Routing
• Dual Diffusion Transformer
• Embodied Control