Introduction

RynnBrain 1.1 is a notable step in embodied AI because it is not framed as a generic multimodal model with robot support added later. Instead, it is trained with a unified spatio-temporal and physically grounded design, aiming to make perception, spatial reasoning, localization, and planning part of the same competence stack. That matters for robotics, where language-level understanding is not enough unless it can be tied to actionable geometry and control.

Key points

• Multiple scales: The family spans 2B, 9B, and 122B-A10B variants, offering a range of deployment options.
• More manipulation-aligned outputs: RynnBrain 1.1 adds contact-point prediction across the model family, which helps bridge the gap between scene understanding and physical interaction.
• Native 3D grounding in smaller models: The 2B and 9B versions gain direct 3D grounding, making their outputs more usable for embodied tasks.
• RynnBrain-VLA for different robots: The team also builds RynnBrain-VLA with a unified cross-embodiment action space and embodiment-specific masking, then deploys it on Unitree G1, Astribot-S1, and Tianji-Wuji.
• Strong evaluation claims: The 122B-A10B model reportedly outperforms evaluated proprietary and open-source systems on VSI-Bench, MMSI, and RefSpatial-Bench. In real-robot experiments, RynnBrain-initialized policies beat Qwen-based and representative generalist VLA baselines, while joint multi-task and multi-embodiment training improves process scores and success rates.

Why it matters

The bigger significance of RynnBrain 1.1 is that it treats embodiment as a first-class modeling problem. In robotics, the hardest part is often not recognizing objects, but understanding where the robot is relative to those objects, what contact points are feasible, and how the same action should transfer across different machines. RynnBrain 1.1 is designed around exactly those friction points.

Its emphasis on physically grounded representations suggests a broader trend in embodied AI: models are becoming less like passive observers and more like pre-action planners. The addition of 3D grounding and contact-point prediction is especially important because these are the kinds of signals that can make a policy more reliable for grasping, reaching, and manipulation.

The cross-embodiment setup is also practical. Real-world robotics rarely happens on a single platform, so a model that can adapt across robot bodies without rebuilding the whole action interface is more attractive for deployment. At the same time, the paper’s results are still benchmark- and experiment-based, so the usual questions remain: how well does the system hold up in clutter, under long-horizon tasks, or when the environment shifts unexpectedly?

Even with those caveats, RynnBrain 1.1 shows where embodied foundation models are heading: toward stronger geometric alignment, more transferable action spaces, and a tighter connection between what a model sees and what a robot can actually do.

Source: Hugging Face Daily Papers