HuggingFace Papers
8/10 signal
RynnBrain 1.1: Towards More Capable and Generalizable Embodied Foundation Model
agenticmodelsresearch
Summary
The paper introduces RynnBrain 1.1, a family of embodied foundation models at 2B, 9B, and 122B-A10B scales, built on Qwen3.5. The models are trained with a unified spatio-temporal and physically grounded framework, adding new capabilities for contact-point prediction and native 3D grounding to better align with robot manipulation. The largest model, RynnBrain 1.1-122B-A10B, outperforms all evaluated proprietary and open-source models on VSI-Bench, MMSI, and RefSpatial-Bench. A vision-language-action variant, RynnBrain-VLA, demonstrates superior performance in real-robot experiments compared to Qwen-based policies.
Problem
Previous embodied foundation models, including RynnBrain 1.0, did not fully explore how to make model representations and outputs more directly applicable to robot manipulation. It was also unclear how effectively these models could serve as an initialization for downstream Vision-Language-Action (VLA) policy training. This work addresses these gaps by introducing new pretraining tasks and a unified VLA framework to bridge perception with real-world action.
Method
RynnBrain 1.1 is a family of three models (2B, 9B, 122B-A10B) based on the Qwen3.5 architecture. The models were pretrained on a large corpus of multimodal data with two new tasks: contact-point prediction and native 3D grounding (for 2B and 9B models). A VLA variant, RynnBrain-VLA, was developed using a unified cross-embodiment action space with embodiment-specific masking. The models were evaluated on a suite of embodied cognition and localization benchmarks, and RynnBrain-VLA was deployed and tested on three real-world robots: Unitree G1, Astribot-S1, and Tianji-Wuji.
Details
Embodied Cognition and Localization Results (Large-Scale Models):
| Benchmark | RynnBrain 1.1 122B-A10B | Qwen3.5 122B-A10B | Gemini 3 Pro | GPT 5.4 | HY-Embodied 0.5 |
|---|---|---|---|---|---|
| VSI-Bench | 75.0 | 66.6* | 48.8* | 49.2* | 68.3 |
| MMSI | 52.0 | 9.2* | 49.2 | 37.5* | 39.2 |
| ERQA | 54.3 | 62.0 | 70.5 | 47.5* | 62.3 |
| RefSpatial-Bench | 79.1 | 69.3 | 65.5 | 26.7* | 57.2 |
* denotes results from authors' reproduction.
Embodied Cognition and Localization Results (9B-Scale Models):
| Benchmark | RynnBrain 1.1 9B | RynnBrain 8B | Molmo2-ER 5B | Qwen3.5 9B |
|---|---|---|---|---|
| VSI-Bench | 74.9 | 70.9 | 74.5 | 61.0 |
| MMSI | 47.0 | 39.6 | 43.8 | 13.4 |
| MindCube | 86.9 | 56.6 | 57.0 | 33.1 |
| RefSpatial-Bench | 67.2 | 59.2 | 52.5 | 37.3 |
Embodied Cognition and Localization Results (2B-Scale Models):
| Benchmark | RynnBrain 1.1 2B | RynnBrain 2B | Cosmos 3-Edge 2B | Qwen3.5 2B |
|---|---|---|---|---|
| VSI-Bench | 72.9 | 70.5 | 59.2 | 44.1 |
| MMSI | 40.5 | 34.1 | 32.3 | 28.3 |
| MindCube | 61.7 | 50.1 | - | 37.5 |
| RefSpatial-Bench | 58.5 | 52.7 | 48.4 | 30.0 |
Scaling Analysis:
- A comparison of RynnBrain 1.1 models (2B, 9B, 122B-A10B) against their corresponding Qwen3.5 baselines revealed three distinct scaling regimes for embodied capabilities.
- General embodied cognition: Both RynnBrain 1.1 and Qwen3.5 improve monotonically with scale, with the performance gap narrowing.
- Reasoning-intensive cognition: RynnBrain 1.1 improves steadily (+38.6%), while Qwen3.5 exhibits negative scaling (-39.2%). The performance gap widens from 18.2 to 50.8 points.
- Embodied localization: RynnBrain 1.1 improves significantly (+24.4%), while Qwen3.5 stagnates (+1.3%). The performance gap widens from 24.3 to 45.4 points.
3D Grounding and Contact Point Prediction:
- The 2B and 9B models were trained with explicit 3D supervision. On the WildDet3D-Val benchmark, the 9B model achieved an AP of 15.3 and mAP of 10.1, outperforming the 2B model's 13.9 AP and 9.1 mAP.
- On the FoundationPose-Val benchmark, the 9B model achieved an AP of 41.9 and mAP of 31.9, compared to the 2B model's 38.6 AP and 28.7 mAP.
- For contact point prediction on a held-out test set, the 9B model achieved a success rate of 84.2%, while the 2B model achieved 78.3%.
Real-Robot VLA Evaluation:
- RynnBrain-VLA policies were deployed on Astribot-S1 and Tianji-Wuji robots.
- Policies initialized from pretrained RynnBrain 1.1 consistently outperformed VLA policies built from the base Qwen models.
- Joint multi-task and multi-embodiment training improved the average process score and final success rate over separately fine-tuned per-task policies.
What's new
This work introduces two novel pretraining tasks for embodied foundation models: contact-point prediction and native 3D grounding. These tasks yield representations and outputs more directly aligned with robot manipulation. The paper also presents RynnBrain-VLA, a vision-language-action model using a unified cross-embodiment action space with embodiment-specific masking, allowing a single policy to be trained on and deployed across heterogeneous robots.
Limitations
The native 3D grounding capability was only implemented and evaluated for the 2B and 9B models, not the largest 122B-A10B model. The real-robot VLA evaluations, while conducted on three distinct platforms, are specific to those embodiments and the tasks demonstrated.
Conclusion
The authors present RynnBrain 1.1, a family of embodied foundation models that demonstrate a clear upward trend in capabilities with increasing scale. The 122B-A10B model achieves state-of-the-art results on several embodied cognition and localization benchmarks. The introduction of contact point prediction and native 3D grounding improves alignment with robotic manipulation. Finally, the RynnBrain-VLA variant shows strong real-world performance and benefits from joint multi-task, multi-embodiment training, highlighting the value of embodied pretraining for downstream robotic policies.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.