RynnBrain 1.1 is a family of embodied vision-language models (2B, 9B, 122B-A10B) that add contact-point prediction and native 3D bounding-box output to the pretraining mix, then post-trains a shared Vision-Language-Action model policy across three real robots using a masked unified action space, with the joint multi-embodiment policy beating per-task fine-tuning by roughly +5 pp success rate.
Suppose you’re shipping a robot manipulation stack. You’ve probably picked a general vision-language model like Qwen or a specialist Vision-Language-Action model like π0.5, and you’re either fine-tuning one policy per robot or hand-aligning action vectors when data comes from a different arm. Both paths hurt: the general VLM doesn’t know 3D geometry or where a gripper should actually touch an object, and per-embodiment policies can’t share what they learn about “reach, grasp, place.” This paper is the DAMO Academy group’s answer to both problems in one release: a base model pretrained specifically on embodied signals (3D boxes, contact points, trajectories, affordances), plus a post-training recipe that lets one policy drive a humanoid, a bimanual arm, and a dexterous-hand rig from the same weights.
The base model is a standard decoder-only VLM built on Qwen3.5, but the output vocabulary is extended so the model can autoregressively emit spatial tokens alongside text: 2D boxes, points, trajectory waypoints, and now full 9-parameter 3D boxes (center xyz, size, pitch/yaw/roll) with all continuous values discretized into integer bins in [0, 1000]. Two pretraining tasks are new versus version 1.0. First, native 3D grounding on the WildDet3D corpus plus synthetic FoundationPose dataset data teaches the 2B and 9B variants to predict metric 3D boxes in camera coordinates from a single image. Second, contact-point prediction replaces the older four-corner grasp rectangle with a compact (x, y, θ) tuple: where to touch, and at what in-plane gripper angle. The authors argue rectangles penalize functionally-valid alternate grasps and bake in gripper-specific extents.
On top of the base model, RynnBrain-VLA is a flow-matching policy where the VLM itself acts as the Diffusion Transformer (DiT) denoiser over action chunks. The trick is the unified 81-dimensional action space, partitioned into semantic groups (Arm-Joint, Arm-EEF, Gripper, Hand, Torso, Head). Each robot activates only its physically-available dimensions via a binary mask; loss is computed only on active dims. Shared components like arm motion get gradient signal from every embodiment; embodiment-specific dims only from their own data.
# Per training sample
embodiment = sample.robot_id # G1, Astribot, Tianji-Wuji
mask = MASKS[embodiment] # 81-dim binary
tokens = pack(instruction, cameras, state, noisy_action_chunk)
pred_chunk = vla_denoise(tokens) # flow matching
loss = mse(pred_chunk[mask], target[mask]) # active dims only
At deploy time they use Real-Time Chunking: predict a 32-step chunk, start executing, kick off the next inference 5 steps in with the unexecuted tail as a guidance signal so chunk boundaries stay smooth.
The prevailing move for robotics foundation models is to scale a general VLM and hope spatial reasoning emerges. This paper shows the opposite. On reasoning-heavy embodied tasks, scaling the base VLM alone actively hurts: raw Qwen3.5 loses ~39% going from 2B to 122B on multi-view spatial benchmarks, while the same architecture with embodied pretraining gains ~39% over the same scale range. The load-bearing evidence is the Qwen3.5-vs-RynnBrain scaling comparison, not the headline leaderboard numbers. Bigger language priors override weak visual-spatial signal unless you supervise the spatial signal explicitly.
The scaling ablation is the finding that anchors everything else. On reasoning-intensive cognition (multi-view, temporal), Qwen3.5 regresses with scale (-39.2% from 2B to 122B), while RynnBrain 1.1 improves (+38.6%). The gap widens from 18.2 to 50.8 points. That’s the mechanism claim: embodied pretraining doesn’t just add capability, it changes the sign of scaling.
Secondary results follow from that:
•
On VSI-Bench, MMSI-Bench, and RefSpatial-Bench, the 122B-A10B model beats every proprietary and open baseline the paper tests (including Gemini 3 Pro, GPT-5.4, Claude Sonnet 4.6).
•
Native 3D grounding at 9B hits 41.12 AP@15 on SUN RGB-D, closing much of the gap to closed-source Gemini Robotics-ER (48.3), and 23.44 AP3D on WildDet3D-Bench (beating the specialized detector at 22.6).
•
On real robots, RynnBrain-VLA vs. a Qwen-initialized VLA with identical post-training recipe: average success rate jumps from 60.00% to 86.67% across three long-horizon tasks. On “Grab the Spatulas” specifically, 50% → 95%.
•
The Generalist policy (one set of weights, all tasks, all robots) beats per-task fine-tuning: 91.28% → 94.14% process score, 86.67% → 91.67% success. Cross-embodiment joint training helps rather than interferes.
Reach for this work when you’re training manipulation policies across more than one robot and today you either maintain N separate checkpoints or hand-map action vectors between them. The unified-action-space-with-masks pattern is directly liftable: define semantic body-part groups once, mask per embodiment, compute loss on active dims only, and let the shared dims (arm motion, grasping) benefit from all data. The finding that this helps rather than causes interference is the concrete takeaway for your training pipeline.
On artifacts: the paper is a technical report and does not link a public code or weights release. It positions itself as an upgrade over RynnBrain 1.0 (also from the same group) and evaluates against GR00T N1.7 and π0.5 as generalist Vision-Language-Action model baselines. Datasets used in pretraining are largely public (Grasp-Anything, GraspNet-1B, WildDet3D, AgibotWorld, Open X-Embodiment); several “RynnBrain-*” subsets are self-collected and the paper doesn’t specify release plans.
For embodied models, supervise the spatial signal explicitly or scale will work against you. Bigger language backbones sharpen priors that overwrite weak visual-spatial cues, so a 122B general VLM can score worse than a 2B model that saw coordinate supervision during pretraining. If your product depends on 3D geometry, contact points, or multi-view reasoning, budget for embodied pretraining data before you budget for parameters.
•
The Qwen-vs-RynnBrain scaling comparison uses the same architecture but different training data, so “embodied pretraining” and “the specific data mix the authors curated” are entangled. A different curated general-purpose mix might close some of the gap.
•
Contact-point evaluation is qualitative only. The authors explicitly note no standardized metric exists for their (x, y, θ) representation, so the claim that it beats rectangles rests on argument and pictures, not numbers.
•
The real-robot comparison against π0.5 and GR00T N1.7 uses the authors’ own tasks and their own teleoperation data. Those baselines weren’t trained on this data. The Qwen-vs-RynnBrain comparison (same recipe, same data, different init) is the cleaner signal; the generalist-baseline numbers are suggestive at best.