Robotics @ XIAOMI

Overview

Xiaomi-Robotics-U0 is a 38B autoregressive world foundation model continually trained for embodied intelligence.

Unlike previous embodied world models trained solely on robot trajectories, Xiaomi-Robotics-U0 unifies the following within a single autoregressive framework.

Text-to-Image Generation
Any-to-Image Editing
Multi-view Embodied Scene Generation
Embodied Transfer
Embodied Video Generation

By transferring the rich semantic knowledge of foundation image generation into robotics, Xiaomi-Robotics-U0 achieves state-of-the-art performance across both single-step and sequential embodied generation tasks while significantly improving downstream robot policies.

Model

Xiaomi-Robotics-U0 unifies multimodal tokenization, embodied data training, and accelerated autoregressive inference into a single foundation model for robot-centric generation.

Data

Xiaomi-Robotics-U0 is continually trained using two complementary stages.

01

Single-step Generation

Text-to-Image
Any-to-Image
Embodied Scene Generation
Embodied Transfer

02

Sequential Generation

Interleaved Subtask-Subgoal
Embodied Videos
Multi-FPS (1 / 3 / 5 FPS)

This unified training strategy preserves foundation generation capability while introducing robot-centric reasoning and interaction modeling.

Architecture

Xiaomi-Robotics-U0 is built on a multimodal autoregressive architecture and uses IBQ as the image tokenizer, representing multiple modalities in the same space.

We propose FlashAR+, a high-speed inference acceleration scheme that extends FlashAR to X2I tasks. At inference time, FlashAR+ uses lightweight post-training and acceleration adaptation, then combines with vLLM to further improve efficiency. Generation efficiency increases by up to 82.9x, with native support for Anything-to-Image multi-reference image editing scenarios.

Acceleration82.9xPeak generation efficiency gain

Before450.77sOriginal 1024 * 1024 single-sample latency

After FlashAR+5.44sAccelerated 1024 * 1024 single-sample latency

With diagonal-parallel decoding and vLLM paged KV-cache batch scheduling, 1024 * 1024 single-sample throughput improves from 450.77 seconds per image to 5.44 seconds per image, balancing high generation quality with engineering deployment speed.

AR and FlashAR+ comparison for embodied transfer and embodied scene generation — FlashAR+ substantially optimizes inference efficiency while preserving generation quality. When further integrated with vLLM at the engineering level, it achieves nearly 83x inference acceleration.

Capability

Across scene synthesis, controllable transfer, video rollout, and general image generation, the model turns language and visual context into scalable embodied data generation.

Capability I: Multi-view Embodied Scene Generation

Generate physically plausible multi-view robot observations directly from language.

Comparison with GPT-Image-2

GPT-IMAGE-2OURS

Embodied Scene Generation

Easy

62.0%

Hard

68.5%

For Capability I: Multi-view Embodied Scene Generation, this benchmark reports the overall win rate of embodied scene generation. Xiaomi-Robotics-U0 outperforms GPT-Image-2 in most comparisons, achieving clear advantages in multi-view consistency while maintaining comparable instruction-following performance.

Capability II: Controllable Embodied Transfer

Transfer an existing embodied scene into a new one while preserving robot interaction.

Structured Scene Control

Xiaomi-Robotics-U0 disentangles embodied scenes into five independently controllable dimensions.

🏢Workspace

🌄Background

📦Foreground Irrelevant Objects

🎯Target Objects

💡Lighting

Each factor can be independently recombined, producing exponentially many robot scenes from compact descriptions without affecting robot pose or manipulation state.

Transfer Results

OriginalTransferred

Transfer Comparison with GPT-Image-2

Original

GPT-Image

Ours

Original

GPT-Image

Ours

Original

GPT-Image

Ours

Original

GPT-Image

Ours

Embodied Transfer

Easy

82.0%

Hard

79.3%

For Capability II: Controllable Embodied Transfer, this benchmark reports the overall win rate of embodied transfer. Xiaomi-Robotics-U0 shows stronger transfer robustness than GPT-Image-2 while preserving comparable instruction-following performance across edited embodied scenes.

Capability III: Embodied Video Generation

Starting from a single observation, Xiaomi-Robotics-U0 predicts future robot interactions across diverse scenarios.

WorldArena

Ranks 1st among over 100 models on WorldArena.

Initial Frame+Language+Action→Video Rollout

Model	EWMScore_P	Rank	Visual Quality	Motion Quality	Content Consistency	Physics Adherence	3D Accuracy	Controllability	Aesthetic Quality	Background Consistency	Depth Accuracy	Dynamic Degree	Flow Score	Image Quality	Instruction Following	Interaction Quality	JEPA Similarity	Motion Smoothness	Perspectivity	Photometric Consistency	Semantic Alignment	Subject Consistency	Trajectory Accuracy
UNIS (Xiaomi-Robotics-U0)	73.64	1	61.78	85.08	55.87	64.60	92.05	91.60	40.79	86.44	85.25	73.70	86.02	53.94	93.86	87.30	90.60	95.51	98.84	2.13	89.35	79.05	41.89
SisyphusWorld	73.06	2	60.18	89.17	58.11	58.28	92.60	85.93	38.38	86.52	94.85	76.27	99.82	45.57	82.68	71.98	96.58	91.42	90.34	5.69	89.18	82.11	44.58
BWM-Fast	72.71	3	63.08	79.67	58.22	62.39	91.78	89.69	40.15	90.17	86.41	69.58	75.11	51.22	90.22	79.88	97.87	94.33	97.14	3.08	89.15	81.42	44.89
SACWM	72.67	4	59.80	85.58	58.86	59.62	93.67	85.41	37.30	87.90	94.99	70.35	95.79	45.83	81.66	72.28	96.26	90.60	92.34	6.50	89.16	82.17	46.95
DexWorldEngine	72.66	5	59.70	79.25	58.24	66.22	91.89	91.04	40.96	88.85	86.19	69.16	74.31	51.44	92.62	81.04	86.71	94.27	97.58	4.42	89.45	81.46	51.41
PAIWorld	72.34	6	63.12	80.49	57.82	61.75	91.59	87.07	40.28	89.03	86.81	68.81	77.22	52.22	85.24	77.70	96.85	95.43	96.36	3.02	88.91	81.42	45.80
FlowWAM-FiveAges	72.00	7	63.54	80.14	57.34	60.54	90.94	86.96	41.94	88.43	85.61	70.12	75.72	53.43	85.24	76.66	95.26	94.59	96.26	2.64	88.68	80.96	44.42
EvoPhysWorld	71.73	8	60.07	84.15	58.04	59.75	89.20	85.62	39.34	88.46	82.37	79.61	77.27	58.73	82.84	76.00	82.15	95.58	96.02	3.59	88.39	82.07	43.49
NoxWM	71.70	9	60.73	87.28	55.22	58.42	87.83	86.68	38.92	87.49	80.08	75.76	87.51	53.66	89.24	79.92	89.61	98.57	95.58	1.10	84.12	77.07	36.91
LeviosaWM	71.64	10	59.28	88.26	55.08	56.34	88.48	88.57	35.82	85.36	81.40	77.17	89.31	57.24	88.40	77.42	84.77	98.29	95.56	1.26	88.73	78.63	35.26

Data source: WorldArena

Generated Scene Rollout

Xiaomi-Robotics-U0 connects image generation with world modeling — generated scenes become starting points for trajectory generation.

Language→Embodied Scene+Language→Video Rollout

Multi-view Prediction

Consistent multi-view embodied prediction across different camera perspectives.

Multi-view Initial Frame+Language→Video Rollout

Capability IV: General T2I & X2I

Beyond embodied generation, Xiaomi-Robotics-U0 retains broad visual synthesis and editing capability across text-guided creation and instruction-driven image transformation.

General Text-to-Image

High-fidelity visual synthesis across scenes, objects, and styles

Text-to-Image sample: Flying Whale — Flying Whale

Text-to-Image sample: Macro Bee — Macro Bee

Text-to-Image sample: Misty Mountain Lake — Misty Mountain Lake

Text-to-Image sample: Impressionist Room — Impressionist Room

Text-to-Image sample: Xiaomi-Robotics-U0 Robot — Xiaomi-Robotics-U0 Robot

Text-to-Image sample: Church Light — Church Light

Text-to-Image sample: Golden Dog — Golden Dog

Text-to-Image sample: Floating Citadel — Floating Citadel

Text-to-Image sample: Snow Taxi — Snow Taxi

Text-to-Image sample: Raspberry Dessert — Raspberry Dessert

Text-to-Image sample: Future City — Future City

Text-to-Image sample: Whiskey Glass — Whiskey Glass

Text-to-Image sample: Neon Rain — Neon Rain

Text-to-Image sample: Teddy Basket — Teddy Basket

Text-to-Image sample: Mushroom Macro — Mushroom Macro

Any-to-Image Editing

Instruction-following edits that preserve structure while changing intent

Add a small wooden boat, with a dark red hull.

Change the watch strap to a dark brown crocodile-embossed leather strap.

Replace the tiny house on wheels with a vintage car.

Replace the bird in the image with a small rabbit.

Add morning dew drops.

Change the house to a coastal setting with beach and ocean.

Add a sprig of cilantro to the painting.

Add a family of ducks swimming near the waterfalls.

Add a person walking along the dirt path toward the ocean.

Change the background to matte black.

Extract the yellow necktie worn in the image.

Remove the colorful inflatable bounce house in the foreground.

Make it look like abstract art.

Add warm sunset lighting.

Transfer the image into a hand-sculpted claymation style.

Replace the pipe-cleaner dragon with a bright yellow rubber duck.

Remove the camper van and make the forest seamless.

Make it look like an architectural sketch.

Change the main object color while preserving details.

Add soft overhead factory lighting.

Add a small wooden boat, with a dark red hull.

Change the watch strap to a dark brown crocodile-embossed leather strap.

Replace the tiny house on wheels with a vintage car.

Replace the bird in the image with a small rabbit.

Add morning dew drops.

Change the house to a coastal setting with beach and ocean.

Add a sprig of cilantro to the painting.

Add a family of ducks swimming near the waterfalls.

Add a person walking along the dirt path toward the ocean.

Change the background to matte black.

Extract the yellow necktie worn in the image.

Remove the colorful inflatable bounce house in the foreground.

Make it look like abstract art.

Add warm sunset lighting.

Transfer the image into a hand-sculpted claymation style.

Replace the pipe-cleaner dragon with a bright yellow rubber duck.

Remove the camper van and make the forest seamless.

Make it look like an architectural sketch.

Change the main object color while preserving details.

Add soft overhead factory lighting.

Real Robot Improvement

Xiaomi-Robotics-U0-generated style-transfer data is added to real demonstration data for policy post-training, enabling the policy to learn robustness to unseen backgrounds, lighting, and visual interference.

Base Settings

Store Earphones

Fold Towel

Pack Box

Varying object layouts under nominal lab lighting and seen tablecloths.

Interference Settings

Store Earphones

Fold Towel

Pack Box

Held-out tablecloths and lighting, including low, colored, and patterned light.

Progress

Task Completion Progress (%)

30.6

-45.8

76.4

54.2

-27.7

81.9

51.1

-48.9

100.0

73.3

-26.7

100.0

28.9

-37.8

66.7

62.2

-2.2

64.4

36.9

-44.1

81.0

63.2

-18.9

82.1

Store EarphonesFold TowelPack BoxAverage

Interference Progress36.9% -> 63.2%+26.3 pts under held-out backgrounds and lighting.

Drop from Base44.1 pts -> 18.9 pts25.2 pts smaller degradation with Xiaomi-Robotics-U0-Aug.

Real Robot Demos

Store Earphones

Green Interference Scene

Original

Xiaomi-Robotics-U0-Aug

Pink Interference Scene

Original

Xiaomi-Robotics-U0-Aug

Fold Towel

Green Interference Scene

Original

Xiaomi-Robotics-U0-Aug

Pink Interference Scene

Original

Xiaomi-Robotics-U0-Aug

Pack Box

Green Interference Scene

Original

Xiaomi-Robotics-U0-Aug

Blue Interference Scene

Original

Xiaomi-Robotics-U0-Aug

See Our Video

Citation

@misc{li2026xiaomiroboticsu0,
  title         = {{Xiaomi-Robotics-U0}: Unified Embodied Synthesis with World Foundation Model},
  author        = {Xinghang Li and Jun Guo and Qiwei Li and Long Qian and Hang Lai and Yueze Wang and Hongyu Yan and Jiahang Cao and Xi Chen and Jingen Qu and Jiaxi Song and Nan Sun and Hanye Zhao and Futeng Liu and Wanli Peng and Heyun Wang and Yunhong Wang and Caoyu Xia and Jack Zhao and Diyun Xiang and Hangjun Ye and Heng Qu and Huaping Liu and Jason Li},
  year          = {2026},
  eprint        = {2607.11643},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2607.11643}
}

Acknowledgements

Built upon Qwen3-32B & EMU3.5.

Thanks to all contributors.