RynnBrain 1.1: Towards More Capable and Generalizable Embodied Foundation Model
Kehan Li, Bohan Hou, Minghao Zhu, Tianyi Zhang, Zesen Cheng, Zhikai Wang, Sicong Leng, Xin Li, Xiao Lin, Biying Yao, Minghua Zeng, Jiangpin Liu, Ronghao Dang, Jiayan Guo, Siteng Huang, Haoyu Zhao, Heng Ping, Yaxi Zhao, Kexiang Wang, Tong Lu, Shengke Xue, Jiahao Tang, Yulei Wang, Zejing Wang, Jianwei Gao, Shijian Lu, Chengju Liu, Jianfei Yang, Mingxiu Chen, Deli Zhao
cs.RO
2026-07-20
An embodied foundation model family (2B/9B/122B-A10B) natively outputs 3D grounding and grasp contact points; the 122B variant beats all closed models on spatial benchmarks, and embodied pretraining turns negative scaling on reasoning tasks into positive.
General multimodal LLMs are strong on image-text QA, but real robots must additionally understand 3D structure, reason across viewpoints, ground language to physical locations, and serve as initialization for downstream manipulation. Plugging a general VLM into VLA training underperforms because it lacks physical priors. RynnBrain 1.1 asks two questions: how do the model's representations and outputs get closer to robot manipulation, and how useful is such a model as VLA post-training initialization?
A unified spatio-temporal, physically grounded framework on a decoder-only architecture (vision encoder, projector, Qwen3.5 backbone), with three scales (2B, 9B, 122B-A10B) sharing architecture and recipe for a clean scaling study. Semantic reasoning and spatial localization are learned jointly, not in separate heads.
Two changes over RynnBrain 1.0. First, the grasp representation moves from grasp rectangles to a contact point (p, theta), because a target admits a set of functionally valid grasps rather than a unique rectangle. Second, the 2B and 9B models gain native 3D grounding, predicting a 3D box (center, size, orientation) from a single image via heavy 3D modeling during training.
Training data spans general MLLM data, multi-dimensional cognition, spatio-temporal localization, 3D-grounded data, contact-point data, and physics-aware planning. A view-conditioned rendering pipeline converts simulation grasp annotations into reliable 2D supervision, filtering views that are too near, distant, truncated, or ambiguous. The model outputs text, boxes, points, trajectory waypoints, 3D perception, and contact signals in one multimodal set.
On embodied cognition benchmarks, the 122B variant reaches VSI-Bench 75.0 (surpassing all open and closed models) and RefSpatial-Bench 79.1 versus Gemini 3 Pro 65.5. The 9B model hits MindCube 86.9 (RynnBrain-8B 56.6, plus 30.3) and MMSI 47.0 versus Qwen3.5-9B 13.4. On 3D grounding, the 9B reaches WildDet3D-Bench AP3D 23.44, beating the specialized WildDet3D detector (22.6); on SUN RGB-D it reaches AP@15 41.12 versus Gemini Robotics-ER 48.3.
The scaling finding is the headline. Three regimes: general cognition improves for both; on reasoning-intensive tasks RynnBrain 1.1 improves steadily (plus 38.6%) while Qwen3.5 shows negative scaling (minus 39.2%); localization improves for both. Embodied pretraining determines whether scaling helps or hurts on reasoning: it turns a capability that degrades under pure VLM scaling into one that benefits consistently.
On real robots (three long-horizon tasks, 20 trials each), RynnBrain-VLA averages 86.67% success, beating GR00T N1.7 (73.33%) and pi0.5 (65.00%); a cross-embodiment generalist reaches 91.67% average success, above single-task policies.
For embodied-AI builders, the evidence is that VLM scaling will not carry spatial reasoning for free; embodied pretraining is not optional. Unified contact-point and 3D outputs also let downstream policies drop a separate 3D detector. Three open scales make it usable as a baseline.
Native 3D grounding is only on 2B and 9B, not 122B. There is no agreed metric for contact-point prediction, so functional validity is hard to score. 3D grounding requires camera intrinsics as input. The contact point is an action-grounding interface, not a full executable grasp (depth, approach, aperture, collision, kinematics go to the downstream policy), so "end-to-end" overstates it. WildDet3D synthetic data inherits residual noise from monocular lifting. The paper has no dedicated limitations section; these are read out of the text and evaluation difficulties.