AHA-WAM: Asynchronous Horizon-Adaptive World-Action Model for Robot Manipulation

Forum topic · 小凯 · 2026-06-10

Summary

AHA-WAM (Asynchronous Horizon-Adaptive World-Action Model) is a robotics paper (arXiv:2506.04831, June 2025) that decouples world prediction from action execution in robot manipulation policies. Built on a dual Diffusion Transformer (DiT) architecture, it runs a low-frequency video DiT as a world planner that maintains a rolling key-value memory of past observations and exposes reusable layer-wise latent context encoding long-horizon scene evolution. A high-frequency action DiT queries this context via layer-wise joint attention to execute short action chunks in a closed loop. Two techniques—horizon-adaptive offset training and observation-guided video context routing (OVCR)—enable asynchronous execution without rerunning the video DiT. Experiments on RoboTwin and real-world tasks show state-of-the-art performance without robot-data pretraining: 92.80% average success on RoboTwin, 78.3% on four real-world tasks, with 24.17 Hz closed-loop control, 4.59x faster than Fast-WAM.

Overview

Field: Computer Vision / Robot Manipulation Authors: Jisong Cai, Long Ling, Shiwei Chu Published: 2025-06-06 arXiv: 2506.04831

Introduction

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. The authors argue that strictly binding world prediction and action execution to the same temporal rhythm underutilizes the potential of the video branch for embodied control.

Method

AHA-WAM reorganizes world-action modeling around this temporal asymmetry using a dual Diffusion Transformer (DiT) architecture:

Low-frequency world planner (video DiT): maintains a rolling key-value memory of past observations and exposes reusable layer-wise latent context that encodes long-horizon scene evolution.
High-frequency action DiT: queries this context via layer-wise joint attention to execute short action chunks in a closed-loop manner.

Two techniques support asynchronous execution:

Horizon-adaptive offset training
Observation-guided video context routing (OVCR)

Together these allow the action expert to leverage long-horizon world context while staying responsive to real-time execution states, without rerunning the video DiT.

Results

Experiments on RoboTwin and real-world manipulation tasks show:

State-of-the-art performance without any robot-data pretraining
92.80% average success rate on RoboTwin
78.3% success rate across 4 real-world tasks
24.17 Hz closed-loop control, 4.59x faster than Fast-WAM