alibaba-damo-academy/RynnBrain

RynnBrain 1.1: Towards More Capable and Generalizable Embodied Foundation Models

What it solves

RynnBrain 1.1 is designed to create more capable and generalizable embodied foundation models. It addresses the gap between high-level vision-language understanding and the physical execution of tasks by robots, enabling them to better understand 3D space, identify interaction points, and translate instructions into real-world actions.

How it works

It uses a unified decoder-only vision-language architecture available in three scales (2B, 9B, and 122B-A10B). The model processes omni-vision inputs and language instructions to produce aligned outputs such as text, pointing sequences, 3D perception data, and contact signals. It bridges perception and action through RynnBrain-VLA, which translates these understandings into control signals for various robot platforms.

Who it’s for

This project is for robotics researchers and developers working on embodied intelligence, specifically those building Vision-Language-Action (VLA) models for humanoid, bimanual, and dexterous-hand robots.

Highlights

  • Unified Scaling: Provides models ranging from 2B to a 122B sparse-MoE model to study how embodied cognition evolves with scale.
  • 3D and Contact Grounding: Supports native 3D bounding box prediction and contact point prediction for action-relevant interaction.
  • Real-Robot Transfer: Demonstrates strong cross-platform generalization across different robot hardware like Unitree G1 and Astribot.
  • Comprehensive Capabilities: Includes cookbooks for spatial understanding, object grounding, affordance location, and trajectory inference.

Related

  • Project
    X-Square-Robot/wall-x

    An open-source framework for embodied foundation models that enables robots to perform complex physical manipulations by bridging vision-language understanding with action prediction.

  • Project
    dexmal/opendm

    OpenDM is a Vision-Language-Action (VLA) model framework for open-world robot control, providing tools for training, fine-tuning, and deploying models that can handle complex, multi-embodiment tasks.

  • Project
    BeingBeyond/Being-H

    A family of human-centric embodied foundation models, including VLA and WAM architectures, designed to enable robots to learn and generalize actions from human videos.

  • Dispatch
    Qwen-VLA: Unifying Vision-Language-Action Modeling for Embodied Intelligence

    Qwen-VLA is a general-purpose Vision-Language-Action model that unifies robotic manipulation, vision-language navigation, and cross-embodiment control into a single generalist policy model.

  • Project
    dexmal/dexbotic

    A VLA (Vision-Language-Action) development toolbox for embodied intelligence that provides a unified framework for pretraining, fine-tuning, and evaluating robotic policies.