GigaWorld-0

GigaAI's framework that uses a world model as a “data engine” to mass-produce robot training data.

GigaWorld-0 was released by GigaAI in November 2025, positioned as a world-model framework for producing training data for vision-language-action (VLA) models. It has two parts: GigaWorld-0-Video uses a large-scale video-generation model to produce texture-rich robot manipulation videos, with fine-grained control over appearance, camera viewpoint, and action semantics; GigaWorld-0-3D combines 3D generation, 3D Gaussian splatting reconstruction, differentiable physics system identification, and motion planning to keep results geometrically consistent and physically plausible. Its companion training framework, GigaTrain, uses FP8 precision and sparse attention to save compute. GigaBrain-0, trained on this synthetic data, performs well on real robots. Code and models are open-source; in July 2026 the team also released GigaWorld-1, aimed at policy evaluation.

ExampleTake a real-robot manipulation video and change the tabletop texture, lighting, or camera viewpoint, generating multiple versions of training data that look different but share the same action, used to improve a VLA's generalization to appearance changes.

Also called

GigaWorld-0-Video, GigaWorld-0-3D, World Models as Data Engine to Empower Embodied AI

Related

GigaBrain-0 · GigaAI · World Model · Synthetic Data · 3D Gaussian Splatting · World-Model-based Policy Evaluation

Sources

GigaWorld-0 (arXiv:2511.19861)
GigaWorld-0 项目主页 (Chinese)
GigaWorld-1 (arXiv:2607.02642)

As of

2026-07