Alibaba-DAMO-Academy/RynnBrain1.1-2B
RynnBrain 1.1-2B by Alibaba-DAMO-Academy is a 2.3 billion parameter embodied foundation model, part of a family extending to 122B-scale sparse-MoE models. It is specifically designed for embodied intelligence, excelling in spatial reasoning, 3D grounding, and real-robot visual-language-action (VLA) transfer. This model introduces explicit 3D-grounded training and contact point prediction, enabling metric 3D understanding and action-relevant interaction grounding for robotic control.
Loading preview...
RynnBrain 1.1-2B: An Embodied Foundation Model
RynnBrain 1.1-2B, developed by Alibaba-DAMO-Academy, is a 2.3 billion parameter model representing a systematic upgrade for embodied intelligence. It is part of the RynnBrain 1.1 series, which includes models scaled up to 122B, all trained under a unified recipe. This model is built upon the Qwen3.5-2B base.
Key Capabilities & Innovations
- Unified Embodied Scaling: RynnBrain 1.1 establishes a unified training recipe across its 2B, 9B, and 122B-A10B scales, allowing for systematic study of embodied cognition evolution with scale.
- Real-Robot VLA Transfer: It bridges perception and action, translating embodied understanding into real-robot control. The model demonstrates strong cross-platform generalization on various robot types (humanoid, bimanual, dexterous-hand tasks) including Unitree G1, Astribot, and Tianji-Wuji.
- Native 3D and Contact Point Grounding: RynnBrain 1.1-2B introduces explicit 3D-grounded training and a novel contact point prediction task. This extends its capabilities from image-plane localization to metric 3D understanding and action-relevant interaction grounding.
Use Cases & Strengths
RynnBrain 1.1-2B is particularly well-suited for applications requiring advanced embodied intelligence, such as:
- Robotics: Enabling robots to understand and interact with their environment through precise 3D grounding and contact point prediction.
- Spatial Reasoning: Performing complex spatial understanding tasks within video scenes.
- Object Interaction: Locating specific objects, understanding their attributes, and predicting affordances and trajectories for manipulation.
Performance evaluations highlight its capabilities in general embodied understanding, real-robot VLA success rates, and 3D grounding accuracy against baselines.