GigaBrain-0

A VLA model series from GigaAI trained mainly on data generated by a world model.

GigaBrain-0 is a vision-language-action (VLA) model released by GigaAI in October 2025. To address the fact that real-robot data is expensive and scarce, it replaces a large share of training data with data generated by a world model, including video generation, appearance transfer, human-video transfer, viewpoint transfer, and sim-to-real transfer. The model takes RGB-D images as input to strengthen spatial awareness, first produces an embodied chain of thought (writing out intermediate reasoning in language) before outputting an action, and uses knowledge insulation to keep the reasoning training from interfering with the action output. There is also a GigaBrain-0-Small variant that can run on a Jetson AGX Orin. Later versions followed: GigaBrain-0.5, trained on more than 10,000 hours of real-robot data; 0.5M*, which adds world-model-based reinforcement learning; and 0.7, from August 2026, which uses more than 37,000 hours of data and a three-system architecture.

ExampleThe paper gradually raised the share of world-model-generated data in the training mix from 0% to 90%, and the model's generalization to changes in object appearance, placement, and camera viewpoint improved markedly as a result.

Also called

GigaBrain, GigaBrain-0.5, GigaBrain-0.5M*, GigaBrain-0.7, A World Model-Powered Vision-Language-Action Model

Related

GigaAI · GigaWorld-0 · Vision-Language-Action Model · Embodied Chain-of-Thought · Knowledge Insulation · Synthetic Data

Sources

GigaBrain-0 (arXiv:2510.19430)
GigaBrain-0.5M* (arXiv:2602.12099)
GigaBrain-0.7 (arXiv:2608.15875)

As of

2026-08