Alibaba DAMO AcademyChina2026

RynnBrain-VLA

An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.

RobotsG1by Unitree Robotics DS1by Astribot D

Key figures as stated by the developer

4 of 6 stated · 2 gaps

Model type ⁠B vision-language-action
Controls ⁠A arms and hands whole body
Inputs ⁠A camera language robot state
Access not stated
Parameters not stated
Base model ⁠B RynnBrain-2B

How to read the figures

14kgA30.9lbA: The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
74.5lb: Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.

The four states

Versions and notes in the key figures (5)

Model type

Developer wording: "vision-language-action modeling (RynnBrain-VLA)".

Controls

The report describes the tested robots as "covering gripper-based whole-body manipulation, dexterous grasping, bimanual coordination, and humanoid whole-body control". Humanoid whole-body motion on Unitree G1 goes through a separate whole-body controller (SONIC).

Inputs

RynnBrain 1.1 technical report: "The model takes a packed sequence of language instruction, multi-camera observations, robot state, and noisy action chunks as input, and iteratively denoises to predict an action chunk."

Access

Note The RynnBrain 1.1 model zoo lists three base models and no RynnBrain-VLA checkpoint.

Base model

RynnBrain 1.0 report: "We build RynnBrain-VLA upon RynnBrain-2B". RynnBrain 1.1 report: "RynnBrain-VLA in RynnBrain-1.1 largely follows that in RynnBrain-1.0".

Full specifications

6 of 12 fields stated

The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field	Value	Source and caveats
Input and output
Model type	`vision-language-action`	B Developer wording: "vision-language-action modeling (RynnBrain-VLA)".
Inputs	`camera` `language` `robot state`	A RynnBrain 1.1 technical report: "The model takes a packed sequence of language instruction, multi-camera observations, robot state, and noisy action chunks as input, and iteratively denoises to predict an action chunk."
Controls	`arms and hands` `whole body`	A The report describes the tested robots as "covering gripper-based whole-body manipulation, dexterous grasping, bimanual coordination, and humanoid whole-body control". Humanoid whole-body motion on Unitree G1 goes through a separate whole-body controller (SONIC).
Degrees of freedom controlled	not stated	Note The shared action space is larger than any one robot uses: "Actions from heterogeneous robot embodiments are represented within a shared 81-dimensional action space partitioned into semantically aligned body-part groups". ⁠A
Size and speed
Parameters	not stated
Control rate	not stated	Note Model layer: "Runs the VLA policy to produce a 32-step action chunk in the standard action space." Control layer of the deployment framework: "Two independent loops: a low-frequency loop (30 Hz) coordinating inference timing, the user interface, and action source switching; and a high-frequency loop (200 Hz) that interpolates the model’s action chunk to the target control frequency and sends commands to the robot." ⁠A
Where it runs and what it is built on
Runs where	not stated	Note In the robot tests, "inference is performed locally on a workstation equipped with a single NVIDIA GeForce RTX 4090 GPU. The workstation is directly connected to the robot and transmits the predicted actions for real-time execution." ⁠A
Base model	RynnBrain-2B	B RynnBrain 1.0 report: "We build RynnBrain-VLA upon RynnBrain-2B". RynnBrain 1.1 report: "RynnBrain-VLA in RynnBrain-1.1 largely follows that in RynnBrain-1.0".
Openness
Access	not stated	Note The RynnBrain 1.1 model zoo lists three base models and no RynnBrain-VLA checkpoint. ⁠C
License	not stated
The developer’s own claims
Learns new tasks	needs retraining	A Fine-tuning setup in the report: "We collect teleoperation demonstrations on the three target embodiments described above, covering a range of pick-and-place and manipulation tasks".
Works across robots	yes	A Developer claim: "RynnBrain-VLA achieves stable performance across multiple robot embodiments, including the Astribot bimanual platform, the Tianji-Wuji dexterous-hand system, and the Unitree G1 humanoid." "jointly training multiple tasks and embodiments within a single VLA does not introduce obvious interference". Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Notes

Alibaba DAMO Academy builds RynnBrain-VLA on its RynnBrain base model to translate fine-grained plans into executable robot actions.
RynnBrain-VLA appears in both the RynnBrain 1.0 and the RynnBrain 1.1 technical reports; the 1.1 report extends it to several robots through one shared action space.
Each robot uses only its own part of the shared action space through a mask.

Sources

A Manufacturer product page. arxiv.org/html/2607.17977 · Retrieved 2026-10-13
B arxiv.org/html/2602.14979 · Retrieved 2026-10-13
C github.com/alibaba-damo-academy/RynnBrain · Retrieved 2026-10-13
D alibaba-damo-academy.github.io/RynnBrain/ · Retrieved 2026-10-13

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.