HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models

2026-07-05

AI Summary

This work addresses the challenge of end-effector misalignment or insertion failure in high-precision robotic manipulation caused by calibration inaccuracies, perception errors, and contact dynamics within world-action (WA) models. To overcome these limitations, the authors propose a hybrid attention-based latent-variable-guided online reinforcement learning framework. The approach integrates latent features and action priors from the WA model generation process through a lightweight actor-critic adapter, augmented with a hybrid attention mechanism that jointly captures task-relevant information from visual context and end-effector correction demands while preserving temporal consistency in actions. Evaluated on four real-world high-precision tasks, the method achieves an average success rate of 87.1%, surpassing the strongest baseline by 19.2 percentage points, with only 45–75 minutes of online training per task.

Abstract

World-action (WA) models can generate long-horizon action chunks for general-purpose robotic manipulation, but they remain vulnerable to calibration, perception, and contact-dynamics errors in real-world precision tasks, often failing in the final few millimeters of alignment or insertion. We propose HALO-WA, a hybrid-attention latent-guided online reinforcement learning (RL) framework for WA models, which leverages latent features and action priors from the WA generation process through a lightweight actor-critic adapter to enable fast online adaptation to real deployment errors. HALO-WA introduces a hybrid-attention structure that preserves the temporal consistency of action chunks while reading task-relevant information from WA latents conditioned on visual context and end-stage correction requirements, thereby producing refined action chunks. We validate HALO-WA on four real-world precision manipulation tasks, where it improves the average success rate from 26.4\% for WA-base to 87.1\%, outperforming the strongest baseline by 19.2 percentage points while requiring only 45--75 minutes of online training per task. To facilitate reproducibility, we further conduct supplementary simulation experiments in RoboTwin and release the code at https://github.com/YeanRoot/HALO-WA.

Problem

• world-action models
• precision manipulation
• real-world deployment errors
• contact-dynamics errors
• perception errors

Innovation

• hybrid-attention
• latent-guided
• online reinforcement learning
• world-action models
• precision manipulation