Hugging Face has released RynnBrain 1.1, a family of embodied foundation models designed to advance robot perception, reasoning, and manipulation. The models come in three sizes: 2B, 9B, and a mixture-of-experts 122B-A10B variant, all trained using a unified spatio-temporal and physically grounded framework.
Compared to its predecessor RynnBrain 1.0, version 1.1 introduces contact-point prediction across all model sizes and native 3D grounding for the 2B and 9B models. These enhancements aim to produce representations and outputs that are more directly aligned with real-world robot manipulation tasks.
The team also developed RynnBrain-VLA, a vision-language-action model that employs a unified cross-embodiment action space with embodiment-specific masking. This model has been deployed on three different robot platforms: Unitree G1, Astribot-S1, and Tianji-Wuji.
On benchmarks, the 122B-A10B model outperforms all evaluated proprietary and open-source models on VSI-Bench, MMSI, and RefSpatial-Bench. Real-robot experiments show that policies initialized with RynnBrain outperform those based on Qwen and other representative generalist vision-language-action models. Furthermore, joint multi-task and multi-embodiment training improves both process scores and success rates compared to per-task training.
The code and models are available on GitHub at alibaba-damo-academy/RynnBrain.