Video World Models · 2026-10-07
IronMan: Information-Constrained Video-Action Learning for Robot Manipulation
Yuanshuo Zhang, Wenzhe Zhao, Zixing Lei, Bin Chen, Siheng Chen
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
IronMan is an information-constrained video-action learning framework for robot manipulation. It distills noisy one-step video features into compact “world tokens” using a dynamics-aware bottleneck based on the information bottleneck principle. Video and action are trained jointly with flow-matching and a variational IB objective. The authors report strong in-distribution performance and out-of-distribution robustness while keeping efficient inference.
Why it matters
The paper targets a mismatch between video representations and action generation: exposing action policies to irrelevant visual detail can hurt generalization under distribution shifts. IronMan’s dynamics-aware, information-constrained interface aims to preserve action-relevant dynamics while suppressing redundant visual cues, and the authors evaluate this across multiple simulation benchmarks (ID/OOD) and real-world tasks.
Method
- Uses a video diffusion transformer (Video DiT) with Wan2.2-5B to extract one-step hidden states from a specified layer, avoiding full video denoising overhead during action inference.
- Introduces a dynamics-aware bottleneck: a spatio-temporal aggregation module with Qformer-style blocks plus a Gaussian head to produce compact world tokens via a variational information bottleneck.
- Trains with a variational IB objective combined with action learning framed as flow matching; during action inference, latents/world tokens are computed once and reused across action denoising steps.
Limitation
The provided text does not state a limitation of the method.
Abstract (from arXiv)
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/
Related papers
- OpenWAM: An Open Framework for Composable World-Action Models
- Video Prediction Policy 2: Predict Better, Act Better
- WorldLine: Action-Driven Visual Simulation for Robotic Manipulation
- Adapting Video Models for Robot Control via Motion-guided Asymmetric Prediction
- Being-M0.7: A Latent World-Action Model for Humanoid Robots
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-06 · Indexed by Pub-AI: 2026-10-07 UTC
IronMan: Information-Constrained Video-Action Learning for Robot Manipulation
Preprint posted: 2026-10-06 · Original record ↗
- arxiv
- 2610.07961
Explicit source versions (1)
- v1 · 2026-10-06
- Preprint available
Source date: 2026-10-06 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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/
