GigaWorld-Policy: An Efficient Action-Centered World--Action Model

arXiv 26.3 2026 62.8 method, application

TLDR

GigaWorld-Policy is an action-centered World-Action Model that efficiently predicts future actions and optionally generates videos for robot policy learning.

Reasoning

The paper introduces a novel action-centered World-Action Model that reduces inference overhead and decouples visual and motion representations, with a causal design making video generation optional. Strengths include addressing key bottlenecks and using a large-scale robot dataset, but the abstract lacks experimental results and comparisons to baselines.

Read-first score

Read-first score 62.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 42.

Recency 8%

100

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

Methodology quality 25%

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

Topical relevance 42%

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 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 245.

Keyword Scores

video world model

world model

world dynamics prediction

generative world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Action-centered World-Action Model (WAM) that learns 2D pixel-action dynamics with efficient action decoding and optional video generation.
Two coupled components: future action sequence prediction conditioned on current observation, and future video generation conditioned on predicted actions and same observation.
Causal design preventing future-video tokens from influencing action tokens, making explicit video generation optional at inference for faster deployment.
Curated diverse, large-scale robot dataset for pre-training an action-centered video generation model, then adapted for robot policy learning.

Methodology

GigaWorld-Policy formulates policy training as two coupled components: the model predicts future action sequences conditioned on the current observation, and simultaneously generates future videos conditioned on the predicted actions and the same observation. The policy is supervised by both action prediction and video generation losses. A causal design ensures future-video tokens do not influence action tokens, allowing optional video generation at inference. The model is pre-trained on a curated large-scale robot dataset as an action-centered video generation model, then adapted for policy learning.

Key Results

On real-world robotic platforms, GigaWorld-Policy runs 9x faster than the leading WAM baseline Motus while improving task success rates by 7%. Compared with pi-0.5, it improves performance by 95% on RoboTwin 2.0.

Resources

Paper