GigaWorld-1 - An open-source embodied world model with excellent vision.

What is GigaWorld-1?

GigaWorld-1, developed by Excellent Vision, is an embodied world model that topped the WorldArena rankings. The model employs the AC-WM (Motion Controlled World Model) architecture, combining explicit motion modeling with a differentiable physics engine to achieve geometrically consistent and physically accurate video generation. The model significantly outperforms Google and NVIDIA in 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 functions of GigaWorld-1

• High-fidelity video generationIt generates geometrically consistent and physically accurate embodied interactive videos based on action commands, supporting simulation of complex scenarios such as robotic arm operations.
• Action condition controlThrough explicit motion modeling, we ensure that the generated video strictly follows the input motion trajectory and physical interaction logic.
• 3D spatial understandingThe model possesses high-precision three-dimensional spatial cognition capabilities, achieving a near-perfect score (97.02 points) in the 3D accuracy evaluation.
• Physical law simulationThe model has a built-in differentiable physics engine that realistically recreates physical interactions such as object collisions and grabbing.
• Data generation and augmentationAs a "digital sandbox" platform, it can generate diverse synthetic data and improve the generalization ability of robot strategy models.

Technical Principles of GigaWorld-1

• AC-WM architectureGigaWorld-1 adopts an Action-Conditioned World Model architecture, specifically designed for embodied intelligence scenarios. The architecture uses action commands as core conditional inputs, enabling the model to predict future visual observations based on specific robot actions, achieving a closed-loop interaction between action and perception.
• Explicit motion modelingThe model introduces an explicit motion modeling mechanism, embedding motion information into the video generation process in a structured manner. This fundamentally ensures the consistency of the generated video in geometric space, avoiding the motion-visual misalignment problem commonly found in traditional implicit modeling.
• Differentiable physics engineGigaWorld-1 innovatively integrates a differentiable physics engine, enabling it to acquire precise physical parameters for robotic arms and simulate complex physical interactions. This allows the model to generate visually realistic videos and strictly adhere to physical laws, achieving accurate modeling of dynamic processes such as collisions, friction, and grasping.
• Large-scale real-data trainingThe model is trained on tens of thousands of hours of high-quality real robot operation video data accumulated over a long period by Excellent Vision. The data covers a wide range of open scenes and detailed operations, significantly enhancing the model's generalization ability and motion-following accuracy in real-world environments.

How to use GigaWorld-1

• Accessing open source platformsAccess the official HuggingFace or GitHub repository to obtain the code and dataset.
• Download resourcesDownload the model weights, inference code, and the CVPR-2026-WorldModel-Track dataset (which contains tens of thousands of hours of real robot operation videos).
• Configure environmentInstall dependencies according to the official documentation and configure a hardware environment that supports deep learning inference (GPU acceleration is recommended).
• Loading ModelUse the provided interface to load pre-trained GigaWorld-1 model weights to a local or cloud server.
• Input action commandsThe robotic arm's motion trajectory or operation instructions are encoded into a format that the model can recognize, and used as input for generation.
• Generate predictive videoRun the inference script, and the model will generate video predictions for future frames based on the input actions, outputting geometrically consistent and physically accurate interactive scenes.
• Assessment and fine-tuningUse WorldArena's evaluation tool to verify the quality of the generated data, or fine-tune the model for domain adaptation based on your own data.
• Integrated ApplicationsThe model is integrated into a robot simulation system or strategy training process for downstream tasks such as synthetic data augmentation and motion planning verification.

Key information and usage requirements for GigaWorld-1

• positionAC-WM (Motion Control World Model), designed specifically for embodied intelligence, topped the global rankings with an overall score of 62.34 in WorldArena's evaluation.
• Core advantagesThe 3D accuracy score of 97.02 points is close to the perfect score, and the physical conformity is 16% better than the second place, making its visual quality industry-leading.
• technical routeIt inherits the EmbodiedDreamer architecture, integrates explicit motion modeling and a differentiable physics engine, and is trained on tens of thousands of hours of real robot data.
• Development TeamExcellent Vision (a Tsinghua-affiliated company, the first in China to develop a global model), founded by Huang Guan, a PhD in Automation from Tsinghua University and former head of perception technology at Horizon Robotics.
• Hardware environmentA GPU-accelerated deep learning inference environment needs to be configured. For specific GPU memory requirements, please refer to the official documentation.
• Data FoundationPre-trained models can be used for direct inference; if fine-tuning is required, robot operation video data that meets the format requirements must be prepared.
• Technical capabilitiesApplicants should have experience using deep learning frameworks (such as PyTorch) and be familiar with research backgrounds related to embodied intelligence or world models.

GigaWorld-1's core advantages

• The world's number one in evaluation resultsWorldArena scored 62.34 overall, making it the only embodied world model to break the 60 mark, surpassing giants like Google and Nvidia.
• Industry-leading 3D accuracyThe model scored 97.02, close to the perfect score, achieving high-precision three-dimensional spatial cognition and geometric consistency.
• Physical compliance capability fault leadingIt improved by 16% compared to the second place, and can realistically simulate complex physical interaction processes such as collision and grabbing.
• Leading in visual qualityIt also significantly outperforms competitors in terms of visual quality, generating images that are both realistic and stable.
• Technical architecture innovationIt pioneered explicit motion modeling combined with a differentiable physics engine, fundamentally solving the motion-visual misalignment problem.

GigaWorld-1 project address

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

Comparison of GigaWorld-1 with similar competing products

Application scenarios of GigaWorld-1

• Robot Strategy TrainingAs a high-fidelity simulator, it generates diverse synthetic data to train VLA models, achieving a nearly 300% improvement in generalization performance for new textures, new perspectives, and new object positions.
• Motion planning verificationVirtual environments allow for the rehearsal of robotic arm trajectories, verification of action feasibility, avoidance of real hardware wear and tear, and improved R&D efficiency by 10-100 times.
• Embodied Intelligence ResearchIt provides the academic community with a global model baseline and supports cutting-edge research topics such as motion prediction, physics reasoning, and long-term task planning.
• Simulation to Reality MigrationBy generating physically accurate videos, a Sim2Real bridge is built, reducing the cost and risk of deploying real robots.
• Enhanced Data Scarcity ScenariosIt generates high-quality training data for real-world scenarios that are difficult to collect, thus solving the bottleneck problem of robot data acquisition.