Alibaba-DAMO-Academy/RynnBrain-8B
RynnBrain-8B by Alibaba-DAMO-Academy is an 8 billion parameter embodied foundation model built upon Qwen3-VL-8B-Instruct, designed for physics-aware egocentric scene understanding. It excels at spatial comprehension, spatiotemporal localization of objects and areas, and physical-space grounded reasoning. The model's primary application is to support robotic systems with reliable localization and precise planning outputs.
Loading preview...
RynnBrain-8B: An Embodied Foundation Model
RynnBrain-8B, developed by Alibaba-DAMO-Academy, is an 8 billion parameter model based on Qwen3-VL-8B-Instruct, serving as a physics-aware embodied brain. It processes egocentric scenes, grounds language in physical space and time, and provides outputs for robotic systems.
Key Capabilities
- Comprehensive Egocentric Understanding: Demonstrates strong spatial comprehension and cognition in embodied QA, counting, OCR, and fine-grained video analysis.
- Diverse Spatiotemporal Localization: Capable of locating objects, target areas, and predicting trajectories across extended episodic contexts, fostering global spatial awareness.
- Physical-Space Grounded Reasoning: The broader RynnBrain family, including "Thinking" variants, integrates textual reasoning with spatial grounding to ensure real-world relevance.
- Physics-Aware Precise Planning: Incorporates localized affordances, areas, and objects into planning outputs, delivering precise instructions for downstream VLA models.
Use Cases
This model is particularly well-suited for applications requiring:
- Robotics: Providing reliable localization and planning for autonomous systems.
- Spatial AI: Tasks involving detailed spatial understanding and object interaction within dynamic environments.
- Embodied AI Research: Exploring advanced cognition, localization, and reasoning in simulated or real-world embodied agents.
Cookbooks are available showcasing RynnBrain's abilities in cognition, localization, reasoning, and planning, including spatial understanding, object understanding, OCR, object/area/affordance/trajectory location, and grasp pose prediction.