GigaWorld-1 - The Embodied World Model Open-Sourced by GigaAI

What is GigaWorld-1

GigaWorld-1 is an embodied world model launched by GigaAI, which ranked first globally in the WorldArena evaluation. The model adopts the AC-WM (Action-Conditioned World Model) architecture, combining explicit action modeling with a differentiable physics engine to achieve geometrically consistent and physically accurate video generation. The model leads Google and NVIDIA by a wide margin on key metrics such as 3D accuracy (97.02 points) and physics adherence, and is trained on tens of thousands of hours of real robot data.

Main Features of GigaWorld-1

• High-Fidelity Video Generation: Generates geometrically consistent and physically accurate embodied interaction videos based on action instructions, supporting simulation of complex scenarios such as robotic arm manipulation.
• Action-Conditioned Control: Through an explicit action modeling mechanism, ensures that generated videos strictly follow the input action trajectories and physical interaction logic.
• 3D Spatial Understanding: The model possesses high-precision three-dimensional spatial cognition, scoring close to full marks (97.02 points) in 3D accuracy evaluation.
• Physics Simulation: The model has a built-in differentiable physics engine, faithfully reproducing physical interaction processes such as object collisions and grasping.
• Data Generation and Augmentation: As a "digital sandbox" platform, it can generate diverse synthetic data to enhance the generalization ability of robot policy models.

Technical Principles of GigaWorld-1

• AC-WM Architecture: GigaWorld-1 adopts the Action-Conditioned World Model (AC-WM) architecture, designed specifically for embodied intelligence scenarios. The architecture takes action instructions as core conditional input, enabling the model to predict future visual observations based on specific robot actions, achieving closed-loop interaction between action and perception.
• Explicit Action Modeling: The model introduces an explicit action modeling mechanism, embedding action information into the video generation process in a structured manner. This fundamentally ensures geometric spatial consistency of generated videos, avoiding the action-vision misalignment issues common in traditional implicit modeling.
• Differentiable Physics Engine: GigaWorld-1 innovatively integrates a differentiable physics engine, capable of obtaining precise physical parameters of robotic arms and simulating complex physical interaction processes. This enables the model to generate visually realistic videos that strictly follow physical laws, achieving accurate modeling of dynamic processes such as collisions, friction, and grasping.
• Large-Scale Real Data Training: The model is trained on tens of thousands of hours of high-quality real robot operation video data accumulated by GigaAI over time. The data covers rich open scenarios and fine manipulation, significantly enhancing the model's generalization ability and action adherence accuracy in real environments.

How to Use GigaWorld-1

• Access the Open-Source Platform: Visit HuggingFace or the official GitHub repository to obtain the code and datasets.
• Download Resources: Download the model weights, inference code, and the CVPR-2026-WorldModel-Track dataset (containing tens of thousands of hours of real robot operation videos).
• Configure the Environment: Install dependencies according to the official documentation and configure a hardware environment that supports deep learning inference (GPU acceleration is recommended).
• Load the Model: Use the provided interface to load the pre-trained GigaWorld-1 model weights to a local or cloud server.
• Input Action Instructions: Encode the robotic arm's motion trajectories or operation instructions into a format recognizable by the model, and use it as conditional input for generation.
• Generate predicted video: Run the inference script. The model will generate video predictions of future frames based on the input actions, outputting geometrically consistent and physically accurate interactive scenes.
• Evaluation and fine-tuning: Use the WorldArena evaluation tool to verify generation quality, or fine-tune the model for domain adaptation based on your own data.
• Integration and application: Integrate the model into a robot simulation system or policy training pipeline for downstream tasks such as synthetic data augmentation and action planning validation.

Key information and usage requirements of GigaWorld-1

• Positioning: An AC-WM (Action-Controlled World Model) built specifically for embodied intelligence, ranking first globally with an overall score of 62.34 in the WorldArena evaluation.
• Core advantages: 3D accuracy of 97.02 points, approaching a perfect score; physical adherence capability improved by 16% compared with the second place; industry-leading visual quality.
• Technical approach: Inherits the EmbodiedDreamer architecture, integrates explicit action modeling with a differentiable physics engine, and is trained on over ten thousand hours of real robot data.
• Development team: GigaAI (Tsinghua-affiliated, the first company in China to work on world models); founder Huang Guan is a PhD from Tsinghua's Department of Automation and former head of perception technology at Horizon Robotics.
• Hardware environment: A GPU-accelerated deep learning inference environment is required; for specific VRAM requirements, refer to the official documentation.
• Data foundation: The pretrained model can be used directly for inference; if fine-tuning is needed, prepare robot manipulation video data that meets the format requirements.
• Technical skills: Experience with deep learning frameworks (such as PyTorch) is required, along with familiarity with research backgrounds related to embodied intelligence or world models.

Core advantages of GigaWorld-1

• Ranked first globally in evaluation results: WorldArena overall score of 62.34, the only embodied world model to exceed 60 points, surpassing giants such as Google and NVIDIA.
• Industry-leading 3D accuracy: The model scores 97.02, approaching a perfect score, achieving high-precision 3D spatial cognition and geometric consistency.
• Breakaway lead in physical adherence capability: Improved by 16% compared with the second place, capable of realistically simulating complex physical interaction processes such as collisions and grasping.
• Comprehensive lead in visual quality: Also significantly outperforms competitors in the visual quality dimension, generating images that combine realism and stability.
• Technical architecture innovation: The first to combine explicit action modeling with a differentiable physics engine, fundamentally solving the action-vision misalignment problem.

Project addresses of GigaWorld-1

• GitHub repository：https://github.com/open-gigaai/CVPR-2026-Workshop-WM-Track
• HuggingFace model library：https://huggingface.co/collections/open-gigaai/cvpr-2026-worldmodel-track

Comparison of GigaWorld-1 with similar competitors

Application Scenarios of GigaWorld-1

• Robot Policy Training: As a high-fidelity simulator, it generates diverse synthetic data to train VLA models, achieving nearly 300% improvement in generalization performance for new textures, new viewpoints, and new object positions.
• Action Planning Validation: It previews robotic arm operation trajectories in a virtual environment, validates action feasibility, avoids real hardware wear and tear, and improves R&D efficiency by 10-100 times.
• Embodied Intelligence Research: It provides academia with a world model baseline, supporting cutting-edge research topics such as action prediction, physical reasoning, and long-horizon task planning.
• Simulation-to-Reality Transfer: Through physically accurate video generation, it builds a Sim2Real bridge, reducing the cost and risk of real robot deployment.
• Enhancement for Data-Scarce Scenarios: For real-world scenarios that are difficult to collect, it generates high-quality training data, solving the bottleneck problem of robot data acquisition.