⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

GigaBrain-0: A World Model-Powered Vision-Language-Action Model

arXiv 25.10 2025 66.6 method, application

TLDR

GigaBrain-0 uses world model-generated data to train VLA robots, reducing real data needs and improving generalization.

Reasoning

The paper presents a novel approach leveraging world models for data generation, which reduces reliance on expensive real-world data and shows strong generalization in real-world tasks. However, the abstract lacks details on the world model architecture and potential limitations of synthetic data.

Read-first score

Read-first score 66.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.

Recency 8%
86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%
72.9

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

Methodology quality 25%
70

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

Reproducibility 25%
46

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 115.

Keyword Scores

world model
10
generative world model
9
world simulator
8
video world model
8
world dynamics prediction
7
interactive world model
5
model-based reinforcement learning world model
4

Deep Analysis

Innovations

  • Leveraging world model-generated data (video generation, real2real transfer, human transfer, view transfer, sim2real transfer) to train VLA models, reducing reliance on real robot data.
  • Incorporating RGBD input modeling and embodied Chain-of-Thought (CoT) supervision for spatial reasoning and long-horizon task execution.
  • Introduction of GigaBrain-0-Small, a lightweight variant optimized for edge devices like NVIDIA Jetson AGX Orin.

Methodology

GigaBrain-0 is a Vision-Language-Action (VLA) foundation model trained on diverse data generated by world models, including video generation, real2real, human transfer, view transfer, and sim2real transfer. The model uses RGBD input and embodied Chain-of-Thought supervision to reason about spatial geometry, object states, and long-horizon dependencies. It is evaluated on real-world dexterous, long-horizon, and mobile manipulation tasks, with baselines implied by comparisons to prior VLA systems.

Key Results

GigaBrain-0 achieves superior generalization across variations in appearances (textures, colors), object placements, and camera viewpoints, with substantial gains in real-world performance on dexterous, long-horizon, and mobile manipulation tasks.

Limitations

  • Reliance on world model-generated data may introduce domain gaps or artifacts that affect real-world transfer.
  • The approach still requires some real robot data (though significantly reduced), and the scalability of world model generation for all task variations is not fully addressed.
  • The lightweight variant GigaBrain-0-Small may have reduced capacity or performance compared to the full model.

Resources

PaperProject

Tags