arxiv:2608.15875

GigaBrain-0.7: Scaling Embodied Foundation Models to Emergent Capabilities with a Three-System Architecture

Published on Aug 16

· Submitted by

Yang Wang on Aug 26

GigaAI

Upvote

Authors:

Guangqing Ding ,

Abstract

GigaBrain-0.7 is a vision-language-action model that improves embodied generalization via a three-system architecture, large-scale heterogeneous pretraining, and joint alignment training.

Generated by thinkingmachines/Inkling-Small

Vision-language-action (VLA) models have become a dominant paradigm for generalist embodied agents, demonstrating strong complex and long-horizon task completion in structured settings. Yet it remains an open question whether current VLA systems can benefit from more effective architectural design, scale to substantially larger and more heterogeneous data regimes, and achieve broader generalization across tasks and embodiments. To this end, we present GigaBrain-0.7, an embodied foundation model with substantially improved generalization across diverse robot embodiments. Specifically, GigaBrain-0.7 unifies understanding, prediction, and action through a three-system architecture, scales pretraining to over 37,000 hours of heterogeneous embodied data, and introduces one-stage alignment training that jointly optimizes vision-language understanding and multi-embodiment action generation. Compared with the preceding GigaBrain-0 series and prior state-of-the-art models including π_{0.5}, GigaBrain-0.7 achieves substantial improvements in foundation zero-shot capabilities, language-conditioned instruction following, and post-training task success rates. In particular, on our in-house Maker H01 platform and mainstream robot embodiments, GigaBrain-0.7 demonstrates strong task adaptability and completion ability across both home and industrial scenarios. All training code and pretrained model weights will be released.

View arXiv page View PDF Project page GitHub 2.61k Add to collection

Community

supermodelteam

Paper author Paper submitter 1 day ago

This comment has been hidden (marked as Off-Topic)

supermodelteam

Paper author Paper submitter about 18 hours ago

•

edited about 17 hours ago

Vision-language-action (VLA) models have become a dominant paradigm for generalist embodied agents, demonstrating strong complex and long-horizon task completion in structured settings. Yet it remains an open question whether current VLA systems can benefit from more effective architectural design, scale to substantially larger and more heterogeneous data regimes, and achieve broader generalization across tasks and embodiments. To this end, we present GigaBrain-0.7, an embodied foundation model with substantially improved generalization across diverse robot embodiments. Specifically, GigaBrain-0.7 unifies understanding, prediction, and action through a three-system architecture, scales pretraining to over 37,000 hours of heterogeneous embodied data, and introduces one-stage alignment training that jointly optimizes vision-language understanding and multi-embodiment action generation. Compared with the preceding GigaBrain-0 series and prior state-of-the-art models including 𝜋0.5, GigaBrain-0.7 achieves substantial improvements in foundation zero-shot capabilities, language-conditioned instruction following, and post-training task success rates. In particular, on our in-house Maker H01 platform and mainstream robot embodiments, GigaBrain-0.7 demonstrates strong task adaptability and completion ability across both home and industrial scenarios.

Highlights

1. Proposing System-3: Integrates a world model into the robot’s real-time decision loop, enabling it to simulate and evaluate before acting.
2. Ready Out of the Box: The dual-pyramid framework enables a single pretrained model to perform many tasks, demonstrating the "One Model, Many Tasks" capability that characterizes foundation models.
3. Decisive Benchmark Leadership: GigaBrain-0.7 leads task success rates by a wide margin on Maker H01, our purpose-built embodied robot, and ranks first in all four evaluations on RoboColiseum, a platform with strong sim-to-real alignment.
4. Real-World Tasks in One Continuous Take: An industry-first demonstration of over 20 minutes of long-horizon, complex precision operations spanning more than 10 tasks, all in one continuous take.