16B UNIFIED EMBODIED MODEL

RynnUnify

Towards Unified
Embodied Brain

Understand the present. Imagine what comes next.
Connect visual reasoning with action.

Explore the model Paper Code Models

8B Reasoner+8B GeneratorOne unified architecture

Explore RynnUnify ↓

Real-robot highlights

Understand & groundImagine & reasonPlan over timePredict & act

Authors

Minghao Zhu^*,‡
Bohan Hou^*,‡
Kun Xiang^*
Zixuan Wang^*
Yinzhou Tang^*,‡
Yifei Yang^*
Zhikai Wang^*
Tianyi Zhang^*
Kehan Li^*
Haoyu Zhao
Siteng Huang
Leigang Qu
Sicong Leng
Dongchi Huang
Mingxiu Chen
Shijian Lu
Yue Wang
Xiaodan Liang
Zhenyu Zeng
Deli Zhao
Chengju Liu^†
Jianfei Yang^†
Xin Li^‡

DAMO Academy, Alibaba Group Hupan Lab

* Core contributors‡ Project leaders† Co-corresponding authors

RESULTS · UNIFIED MULTIMODAL MODELS

Understand. Imagine.
Reason. Act.

One foundation for embodied understanding,
visual planning, and anticipatory reasoning.

OVERALL

General benchmarks

Best among compared UMMs

28.0

BAGEL

44.9

RxBrain

53.2

Cosmos 3 Nano

57.9

RynnUnify

Score ↑ · unified models

OVERALL

Rynn-GEU

DirectRollout / Agentic

11.9

10.8

BAGEL

31.4

28.6

RxBrain

37.3

33.5

Cosmos 3 Nano

44.0

50.0

RynnUnify

BAGEL / RxBrain / Cosmos: interleaved rollout. RynnUnify: Agentic.

RYNN-VP · OVERALL

Visual planning

Best among compared UMMs

20.32

BAGEL

8.00

RxBrain

45.05

Cosmos 3 Nano

54.80

RynnUnify

Score ↑ · unified models

UNDERSTANDING

VSI-Bench

Best among compared UMMs

34.5

BAGEL

43.5

RxBrain

54.9

Cosmos 3 Nano

67.1

RynnUnify

Score ↑ · unified models

UNDERSTANDING

RoboSpatial

Best among compared UMMs

41.0

BAGEL

60.5

RxBrain

62.1

Cosmos 3 Nano

72.8

RynnUnify

Score ↑ · unified models

SPATIAL GROUNDING

RefSpatial

Best among compared UMMs

8.6

BAGEL

34.1

RxBrain

53.1

Cosmos 3 Nano

60.3

RynnUnify

Score ↑ · unified models

Full general benchmark results+

Method	VSI-Bench	MMSI-Bench	MindCube	ERQA	RoboSpatial	EmbSpatial	RefSpatial	ShareRobot Affordance	ShareRobot Trajectory ↓	Where2Place	Average
Qwen3-VL-8B	60.3	29.8	36.8	44.8	58.2	78.5	53.4	37.9	0.37	64.7	52.7
Qwen3.5-9B	61.0	13.4	33.1	41.5	54.1	76.9	37.3	7.1	0.53	48.8	42.0
RynnBrain-8B	71.0	39.6	56.6	46.8	73.1	80.4	59.2	44.7	0.35	75.5	61.2
MiMo-Embodied-7B	48.5	30.2	43.1	46.8	70.3	76.2	48.0	35.6	0.41	63.6	52.1
Embodied-R1.5-8B	56.1	29.8	27.9	46.0	69.7	78.1	54.2	25.2	0.29	74.0	53.2
GPT-5.6 Terra	58.9	36.9	62.6	56.5	57.2	81.1	31.4	31.3	0.29	39.3	52.6
Claude Sonnet 5	37.3	28.1	49.6	46.2	41.3	74.3	15.1	2.9	0.47	3.3	35.1
BAGEL	34.5	31.0	34.7	39.5	41.0	67.0	8.6	3.2	0.88	8.8	28.0
RxBrain	43.5	35.8	47.5	42.8	60.5	82.3	34.1	11.7	0.49	39.8	44.9
Cosmos 3 Nano	54.9	36.2	40.5	46.0	62.1	81.4	53.1	28.0	0.34	64.0	53.2
RynnUnify	67.1	35.9	44.2	43.8	72.8	83.0	60.3	41.8	0.27	56.9	57.9

Full Rynn-GEU results+

Model	MMSI-Bench	MindCube	ERQA	Public Avg.	GEU: Trajectory	GEU: Affordance	GEU: Area	GEU: Position	GEU: Distance	GEU: Size	GEU: Direction	GEU: Action reasoning	GEU: Physics	GEU: Mess	GEU Overall
Qwen3-VL-8B	29.8	36.8	44.8	37.1	67.3	62.3	31.7	27.5	39.6	52.5	20.0	45.5	15.0	7.3	36.9
Qwen3.5-9B	13.4	33.1	41.5	29.3	73.8	47.5	17.5	42.0	9.2	57.5	17.0	45.0	7.2	15.8	33.2
RynnBrain-8B	39.6	56.6	46.8	47.7	79.2	60.5	28.1	8.0	43.5	63.0	34.5	40.0	23.0	9.5	38.9
MiMo-Embodied-7B	30.2	43.1	46.8	40.0	64.8	16.0	15.1	34.0	45.0	53.5	13.0	47.5	5.3	18.2	31.2
Embodied-R1.5-8B	29.8	27.9	46.0	34.6	77.9	63.5	35.6	28.0	43.1	52.5	19.5	44.0	21.9	5.3	39.1
GPT-5.6 Terra	36.9	62.6	56.5	52.0	77.6	53.2	42.9	44.0	46.5	60.0	24.0	56.0	20.0	26.3	45.1
Claude Sonnet 5	28.1	49.6	46.2	41.3	70.9	20.6	16.0	31.5	47.5	58.0	22.2	63.0	24.0	27.1	38.1
BAGEL	31.0	34.7	39.5	35.1	4.4	26.9	3.0	11.5	15.0	41.0	7.0	8.0	0.0	2.0	11.9
BAGEL + Rollout	30.7	39.6	44.8	38.4	1.4	22.8	3.5	18.5	10.8	33.5	5.5	10.0	0.5	1.9	10.8
RxBrain	35.8	47.5	42.8	42.0	42.8	44.9	27.0	24.5	40.2	44.5	21.5	33.5	13.7	21.0	31.4
RxBrain + Rollout	31.8	45.3	37.0	38.0	33.3	46.0	27.0	24.0	37.9	42.5	20.5	26.0	9.6	19.5	28.6
Cosmos 3 Nano	36.2	40.5	46.0	40.9	76.8	61.3	31.8	27.0	39.5	42.3	17.5	40.5	20.0	16.8	37.3
Cosmos 3 Nano + Rollout	30.1	36.4	41.4	36.0	73.7	53.1	23.9	32.0	28.4	39.1	21.7	40.0	4.5	19.0	33.5
RynnUnify	35.9	44.2	43.8	41.3	78.9	65.7	35.9	38.5	52.9	57.0	22.0	57.5	22.6	8.5	44.0
RynnUnify + Agentic	36.3	52.6	44.8	44.6	81.5	67.0	66.2	36.0	51.8	56.5	25.5	61.0	25.0	29.1	50.0

Full visual-planning results+

Method	Output	PI	VR	VQ	Overall
GPT-5.6 Terra + MiniMax-H3	Composed video system	59.82	69.56	79.73	65.71
BAGEL	Step images	9.46	24.36	58.50	20.32
RxBrain	Step images	1.28	11.35	28.15	8.00
Cosmos 3 Nano	Video	30.50	54.58	79.69	45.05
RynnUnify	Video	45.66	60.54	77.53	54.80

01 / UNIFIED ARCHITECTURE

A unified model.
Memory across time.

An 8B reasoner and 8B generator, connected through joint attention.

MULTIMODAL CONTEXTLanguageImagesVideo / history

RynnUnify 16B · MIXTURE OF TRANSFORMERS

Reasoner

Autoregressive reasoning

···

Generator

Rectified-flow video prediction

···

Joint Attention

MODEL OUTPUTS

Answers & spatial grounding

Conditioned video predictions

Subplan · video prompt · duration

01 / UNDERSTAND

Interpret language, images, and video for embodied answers and spatial grounding.

Paper architecture↗

What the generator sees

TextSubtask · scene changes · duration

···

Ref. 0Initial frame

Ref. 1History frame

Ref. 2History frame

Ref. 3Latest endpoint

Scene memory

Main video

Fixed prefix

···

Noisy future

···

Fixed visual conditionNoisy prediction tokens · video loss

02 / BUILDING THE FOUNDATION

Learn to reason.
Learn to imagine.

Train the reasoner, then the memory-conditioned video generator.

PRETRAINING

Reasoner fine-tuning

Train reasoner

Embodied understandingPhysical groundingNext-subtask planningVideo-prompt writing

Autoregressive lossAssistant tokens only

PRETRAINING

Generator fine-tuning

Reasoner frozen

Cross-chunk scene memory + self-forcing

↳

Selected visual conditionsNoise + optional blur, before VAE encoding

Initial

GT anchor

History 1

Selected history

History 2

Selected history

Latest

Selected condition

Same reference selection & VAE encoding

Start / prefix

→GT future target unchanged

Flow-matching lossPredicted video positions only

Paper training diagram↗

03 / LONG-HORIZON VISUAL PLANNING

Reason a step.
Imagine its outcome.

Predict a subtask, imagine its outcome, and continue.

MODEL-GENERATED VISUAL PLAN

Example

Goal

Current subtask

Reason → Imagine → Continue01 / 10

Model-predicted subtasks. Select a step to inspect its scene.

04 / AGENTIC REASONING

Look beyond
the observed frame.

Answer directly, or generate visual evidence before answering.

Rynn-GEU · Spatial

Archived example · schematic flow

Agentic Reasoner Additional 8B agent · initialized from the reasoner

D · answer directlyGenerated visual evidence returns to the agent

01 / 07

OBSERVE

Agentic RL training +

TRAINING ONLY · BCPO

Reasoning + routing

Only the agent is updated

01 / TASK REWARDr = F × A

Valid format × task-specific answer quality

02 / WITHIN EACH BRANCHGSPO

Compare trajectories under the same D or G path.

03 / BETWEEN BRANCHESSoft D/G routing

Mean reward differences teach when generation helps.

Paper RL diagram↗

Model-generated clips · abridged prompts · predicted grounding points

DIRECT ANSWERING

Sometimes, the observation is enough.

D · Answer from the given observations.

05 / WAM

Visual prediction.
Physical execution.

Video hidden states guide robot actions.

REASONER OUTPUTContinuous subtask latent

GENERATOR OUTPUTFuture video

ACTION OUTPUTAction chunk

···

RynnUnify UMM SHARED MULTIMODAL BACKBONE

Frozen

Reasoner

Task + visual history → latent plan

Trainable

Generator

Subtask-conditioned video prediction

Joint Attention

Video hidden states

···

Trainable

Action head

Flow matching

Cross-attends to video features

Task instructionVisual historyCurrent RGB views

↳ Camera-pose headAuxiliary geometry supervision · training only

↑ Noisy action

FEATURES → ACTIONS

The action head reads video hidden states, not decoded frames.

ROBOT POST-TRAINING

Three training stages

At inferenceVideo features are reused during action sampling. The camera-pose head is omitted.

Paper WAM diagram↗

Robot benchmark results

RynnUnify WAM · success rate (%) ↑

SUCCESS RATE (%) ↑ · 0–100

Selected-20

Cosmos 345.0

Xiaomi-Robotics-146.0

CMA47.0

RynnUnify51.5

RoboCasa365 · 200 episodes

SUCCESS RATE (%) ↑ · 0–100

Composite-Seen

Cosmos 349.6

Xiaomi-Robotics-157.1

CMA50.6

RynnUnify52.0

RoboCasa365

SUCCESS RATE (%) ↑ · 0–100

Composite-Unseen

Cosmos 339.0

Xiaomi-Robotics-132.1

CMA44.0

RynnUnify50.0

RoboCasa365

SUCCESS RATE (%) ↑ · 0–100

Real robot

π₀.₅25.0

GR00T N1.726.7

RynnUnify46.7

Astribot S1 · three tasks

Full RoboCasa365 table+

Method	Composite-Seen	Composite-Unseen	Selected-20
π₀	6.1	1.1	3.0
π₀.₅	7.1	1.2	2.0
GR00T N1.5	14.8	2.7	7.5
GR00T N1.6	9.4	1.7	6.5
Cosmos 3	49.6	39.0	45.0
Xiaomi-Robotics-1	57.1	32.1	46.0
CMA	50.6	44.0	47.0
RynnUnify	52.0	50.0	51.5

Explore the world-action model & real-robot tasks ↗

Task setups & execution sequences +

06 / OUR BENCHMARKS

Rynn-VP & Rynn-GEU

Long-horizon visual planning.
Reasoning beyond observed states.

Reference videos illustrate the benchmarks; planning text is illustrative.Original PDF ↗

RynnUnify

Towards UnifiedEmbodied Brain

Authors

Understand. Imagine.Reason. Act.

A unified model.Memory across time.

Two experts. One visual foundation.

Reasoner

Generator

What the generator sees

Learn to reason.Learn to imagine.

Reason a step.Imagine its outcome.

Look beyondthe observed frame.

When should the model imagine?

Reasoning + routing

Sometimes, the observation is enough.

Visual prediction.Physical execution.

From latent plans to robot actions.

Reasoner

Generator

Action head

Three training stages

Robot benchmark results

Selected-20

Composite-Seen

Composite-Unseen

Real robot

Rynn-VP & Rynn-GEU

Towards Unified
Embodied Brain

Understand. Imagine.
Reason. Act.

A unified model.
Memory across time.

Learn to reason.
Learn to imagine.

Reason a step.
Imagine its outcome.

Look beyond
the observed frame.

Visual prediction.
Physical execution.