Rolling-WAM: World Action Models with Rolling Imagination

arXiv 2026 29.6 method, application

TLDR

Rolling-WAM distributes joint video-action denoising across replanning cycles, enabling faster closed-loop robotic manipulation with competitive performance and a 4.5x replanning speedup.

Reasoning

The paper presents a clear latency-reduction mechanism for World Action Models and validates it on simulated and real robots. However, the abstract lacks detailed limitations or comparisons, and the world-model framing is more incremental than foundational.

Read-first score

Read-first score 29.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation

Topical relevance 25%

47.1

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 214.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

interactive world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Distributes joint video-action denoising across successive replanning cycles instead of denoising the full horizon at each cycle
Maintains a sliding window of video-action chunks at staggered noise levels
Uses a rolling noise schedule that fully denoises the imminent action chunk for execution while partially refining farther-future chunks
Carries an evolving visual-action context across chunk boundaries as retained future chunks continue denoising with new camera observations

Methodology

Rolling-WAM introduces a rolling formulation of World Action Models in which a sliding window of video-action chunks is maintained at staggered noise levels. At each replanning step, a rolling noise schedule fully denoises the imminent action chunk for execution and partially refines future chunks, while newly observed camera frames advance the window and retained chunks continue denoising. The method is evaluated on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid against standard joint WAMs using manipulation performance and replanning speedup.

Key Results

Rolling-WAM achieves competitive manipulation performance on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid. It provides a 4.5x steady-state replanning speedup over standard joint WAMs by avoiding denoising the entire prediction horizon from scratch.

Resources

Paper