IronMan: Information-Constrained Video-Action Learning for Robot Manipulation

2026-10-06

AI Summary

This study addresses the problem that redundant details in video representations impair the generalization of robotic manipulation policies. Grounded in the information bottleneck principle, this work proposes a robust video-action learning framework. Its core innovation lies in a dynamics-aware bottleneck mechanism that distills noisy, entangled single-step video features into a compact world representation, effectively suppressing irrelevant visual distractions while preserving action-relevant dynamic cues. The proposed method demonstrates superior performance across both simulation benchmarks, including LIBERO and RoboTwin, and real-world robot experiments. Notably, it achieves an out-of-distribution (OOD) robustness improvement of 10.4 percentage points over state-of-the-art baselines, significantly enhancing policy generalization.

Abstract

Video Action Models (VAMs) couple visual dynamics modeling with action generation for robot manipulation. However, video representations are not naturally suited to action generation, as exposing the action policy to excessive visual detail can impair its generalization ability. Therefore, we introduce IronMan (Information-constRained videO-actioN learning for robot MANipulation), a robust video-action learning framework built on the information bottleneck principle. The core principle of this framework is to impose information constraints that suppress irrelevant visual information while preserving action-relevant dynamics cues. IronMan employs a dynamics-aware bottleneck that distills noisy, entangled one-step video features into compact world representations. Extensive simulation and real-world experiments demonstrate strong in-distribution (ID) performance and out-of-distribution (OOD) robustness while maintaining efficient inference. IronMan achieves success rates of 99.0% on LIBERO and 79.4% on RoboTwin clean2clean, outperforming all the evaluated baselines. Under OOD shifts, IronMan achieves a success rate of 79.1% on LIBERO-Plus, exceeding the strongest baseline by 10.4 percentage points. Project page: https://youngsoul0731.github.io/ironman-project-page/

Problem

• Robot Manipulation
• Video Action Models
• Generalization
• Visual Dynamics
• Information Bottleneck

Innovation

• Information Bottleneck
• Video-Action Learning
• Robot Manipulation
• Dynamics-Aware Bottleneck
• Out-of-Distribution Robustness