Paper deep dive
Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model
Xinghang Li, Jun Guo, Qiwei Li, Long Qian, Hang Lai, Yueze Wang, Hongyu Yan, Jiahang Cao, Xi Chen, Jingen Qu, Jiaxi Song, Nan Sun, Hanye Zhao, Futeng Liu, Wanli Peng, Heyun Wang, Yunhong Wang, Caoyu Xia, Jack Zhao, Diyun Xiang, Hangjun Ye, Heng Qu, Huaping Liu, Jason Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 82%
Last extracted: 7/18/2026, 2:52:01 PM
Summary
The paper introduces Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive model for unified embodied synthesis. It extends foundation world models to handle multi-view consistency, geometric coherence, and robot embodiment constraints by jointly optimizing text-to-image, image editing, embodied scene generation, embodied transfer, and embodied video generation. The model outperforms GPT-Image-2.0 in human evaluations and significantly improves downstream robot policy success rates.
Entities (24)
Relation Signals (23)
Xiaomi-Robotics-U0 → improves → pi_0.5
confidence 90% · improving the out-of-distribution success rate of pi_0.5 from 36.9% to 63.2%
Xiaomi-Robotics-U0 → outperforms → GPT-Image-2.0
confidence 90% · outperforming GPT-Image-2.0 in human evaluations of embodied scene generation and transfer
Xiaomi-Robotics-U0 → ranksfirston → World Arena
confidence 90% · ranking first on World Arena for embodied video generation
Emu3.5 → basedon → Qwen-3-32B
confidence 80% · built upon the Qwen-3-32B decoder-only Transformer
Xiaomi-Robotics-U0 → initializedby → Emu3.5
confidence 80% · We adopt the open-sourced EMU3.5 [16] as the initialization model
Xiaomi-Robotics-U0 → trainedon → RoboTwin2.0
confidence 80% · we incorporate data from RoboTwin2.0 [11]
Xiaomi-Robotics-U0 → trainedon → GenieSim
confidence 80% · GenieSim [69]
Xiaomi-Robotics-U0 → trainedon → InternData-A1
confidence 80% · InternData-A1 [57]
Xiaomi-Robotics-U0 → trainedon →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recent foundation image and video generation models offer strong generalization and controllability, but their direct application to embodied scenarios is limited by requirements for multi-view consistency, geometric coherence, and robot embodiment constraints. Existing methods typically adapt foundation models with limited robot data, often sacrificing visual knowledge acquired during large-scale pre-training. We present Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive model for unified embodied synthesis. It treats embodied generation as an extension of foundation image and video generation and jointly optimizes text-to-image generation, image editing, embodied scene generation, embodied transfer, and embodied video generation. This unified framework preserves the generalization of the pre-trained world foundation model while adapting it to embodied settings. Xiaomi-Robotics-U0 is the first model to support high-quality multi-view scene generation across multiple robot embodiments and to introduce structured, controllable embodied transfer for fine-grained editing while preserving multi-view consistency and interaction dynamics. It achieves state-of-the-art results on single-step and sequential generation tasks, outperforming GPT-Image-2.0 in human evaluations of embodied scene generation and transfer, ranking first on World Arena for embodied video generation, and improving the out-of-distribution success rate of pi_0.5 from 36.9% to 63.2% on challenging real-world manipulation tasks. These results show that foundation world models can serve both as embodied world models and scalable data engines for embodied intelligence. Code and checkpoints are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2607.11643v1
- Canonical: https://arxiv.org/abs/2607.11643v1
Trouble viewing inline? Open PDF directly →
Full Text
112,160 characters extracted from source content.
Expand or collapse full text
Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model Xiaomi Robotics 1 Abstract Recent foundation image and video generation models offer strong generalization and controllability, but their direct application to embodied scenarios is limited by requirements for multi-view consistency, geometric coherence, and robot embodiment constraints. Existing methods typically adapt foundation models with limited robot data, often sacrificing visual knowledge acquired during large-scale pre-training. We presentXiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive model for unified embodied synthesis. It treats embodied generation as an extension of foundation image and video generation and jointly optimizes text-to-image generation, image editing, embodied scene generation, embodied transfer, and embodied video generation. This unified framework preserves the generalization of the pre-trained world foundation model while adapting it to embodied settings. Xiaomi-Robotics-U0is the first model to support high-quality multi-view scene generation across multiple robot embodiments and to introduce structured, controllable embodied transfer for fine-grained editing while preserving multi-view consistency and interaction dynamics. It achieves state-of-the-art results on single-step and sequential generation tasks, outperforming GPT-Image- 2.0 in human evaluations of embodied scene generation and transfer, ranking first on World Arena for embodied video generation, and improving the out-of-distribution success rate ofπ 0.5 from 36.9% to 63.2% on challenging real-world manipulation tasks. These results show that foundation world models can serve both as embodied world models and scalable data engines for embodied intelligence. Code and checkpoints are available athttps://robotics.xiaomi.com/ xiaomi-robotics-u0.html. 1 Introduction Recent foundation image and video generation models [4,13,16,27,42,45,55,60,64] have made remarkable progress in semantic understanding, controllable generation, and visual reasoning through training with data on the Internet. Large-scale generative models are now capable of synthesizing highly realistic images and videos from various multimodal inputs, demonstrating impressive generalization far beyond the distribution of their training data. Such capabilities make foundation generative models an attractive starting point for embodied intelligence [34,50,66,70], where robots are required to reason about complex environments and imagine future interactions before acting. However, embodied generation [30,32,36] introduces challenges that differ fundamentally from conventional image and video synthesis. Unlike natural image generation, embodied scenarios require strict multi-view consistency, accurate geometric and physical coherence across cameras, explicit robot embodiment constraints, 1 See Contributions section for full author list. Please send correspondence to mi-robotics@xiaomi.com. 1 arXiv:2607.11643v1 [cs.RO] 13 Jul 2026 Scene Gen Embodied Transfer Image Gen Video Gen Figure 1 Embodied and general capabilities ofXiaomi-Robotics-U0. The rectangle corresponds to the initial observations for the same embodiment, the pairwise transfer sample, and the keyframes within a video for the embodied capabilities. All frames are referenced and generated images. and temporally consistent interaction dynamics. The generated observations must remain compatible with robot kinematics, camera calibration, and downstream manipulation policies rather than merely appearing visually realistic. Consequently, directly applying existing foundation image or video generation models to embodied scenarios often leads to inconsistent geometry, implausible robot states, and poor compatibility with robot control. Recent embodied world models [1,50,72] attempt to bridge this gap by continually adapting pre-trained foundation models using robot trajectories [5,15,24,57,58,69] or egocentric manipulation videos [2,21]. While these approaches have demonstrated encouraging progress, their adaptation paradigm remains fundamentally different from that of foundation model pre-training. Instead of preserving the diverse image and video generation tasks that originally endow foundation models with strong semantic understanding and controllable generation, post-training is typically performed solely on robot-specific datasets. Since embodied datasets are considerably smaller, less diverse and highly repetitive compared with internet-scale visual corpora [18,38,48], such adaptation inevitably weakens the generalization capability inherited from foundation models and limits the diversity, controllability, and scalability of the embodied generation. In this work, we presentXiaomi-Robotics-U0, a Unified Embodied Synthesis Model that unifies foundation generation and embodied generation within a single training paradigm. As shown in Figure 1, rather than specializing a foundation model into a robot-specific generator,Xiaomi-Robotics-U0reformulates embodied synthesis as a natural extension of foundation image and video generation. Starting from a world foundation model,Xiaomi-Robotics-U0is continually trained on both general-domain and embodied datasets under a unified autoregressive objective, jointly learning text-to-image generation, image editing, embodied scene generation, embodied transfer, and embodied video generation. Figure 2 exhibits the above-mentioned capabilities ofXiaomi-Robotics-U0. This unified formulation preserves the rich semantic knowledge and controllable generation capability of foundation models while effectively adapting them to robot-centric multi-view reasoning and interaction modeling. Beyond single-step generation,Xiaomi-Robotics-U0further extends the embodied synthesis to sequential world modeling. By jointly learning interleaved embodied sequences and manipulation videos across multiple temporal resolutions, the model captures both long-horizon task progression and fine-grained interaction 2 AÄ ® hÈĭ dÝČÄt ėü ėÚÈ tëij澗 TĜČo ėÚÈ Ýó®ÔÈ Ýoėü ®o üÝn s®ÝoėÝoÔ tėijnÈ澗 CÚ®oÔÈ ėÚÈ tÚ®sÈ üh ėÚÈ hnü¾ë üh dÝČÄt澗 EIJėÈoÄ ėÚÈ nÈhė ®Čó ®oÄ ÈIJȾĜėÈ ® sÝo¾Ú ÔČ®ts üo ėÚÈ tó®n üdèȾė oÈ®Č ėÚÈ ė®dnÈ ÈÄÔÈ R®ÝtÈ ėÚÈ nÈhė®Čó ėü ÔČ®ts ėÚÈ düėnÈ LÝhė ėÚÈ ČÈÄ dÈyÈČ®ÔÈ ¾®o ĭÝėÚ ¾ĜČyÈÄ ėüs澘 sĜė Ýė Ýo ėÚÈ dnĜÈ sn®tėݾdüIJ hüČ tėüČÝoÔ ėÚÝoÔt澘 ėÚÈo Ú®oÄnÈ ėÚÈ tóüėÚ ĭÚÝėÈ ¾®o ĭÝėÚ ®¾Èoėt tÝóÝn®Čnij澗 Rüdüė AČó TijsÈ澙 AÔÝBüė G溷 WüČëts®¾È澙 PünÝtÚÈÄ ÔünÄÈo IoÄÝ®o dČ®t ėÚ®nÝ sn®ėÈ ĭÝėÚ ÄÝyÝÄÈÄ tȾėÝüot OdèȾėt澙 TĜČóÈČݾ süĭÄÈČ澘 ČÈÄ ¾ÚÝnÝ澘 ÈoÔČ®yÈÄ tÝnyÈČ tsüo澘 ¾ü¾üoĜė ¾ÚĜėoÈij澘 ¾Ýn®oėČü tsČÝÔ澘 ÄČÝÈÄ tė®Č ®oÝtÈ LÝÔÚėÝoÔ澙 W®Čó üyÈČÚ騀 Ýo¾®oÄÈt¾Èoė nÝÔÚė澘 ÔünÄÈo ÔnÝoėt澘 tühė tÚ®Äüĭt B®¾ëÔČüĜoÄ澙 CünüČhĜn tÝnë ÄČ®sÈt澘 s®ėÈČoÈÄ ČĜÔ 煥 ĭüÄÈo ¾ĜtÚÝüot澘 dČ®t üÝn n®óst澘 ¾®ČyÈÄ ĭüÄÈo s®ČėÝėÝüo Rüdüė AČó TijsÈ澙 AÔÝdüė G溸 WüČëts®¾È澙 DÈė®ÝnÈÄ J®s®oÈtÈ NÝtÚÝèÝo nüó ĭÝėÚ ė®Ĝė tÝnë ėÚČÈ®Ät OdèȾėt澙 GünÄ ėÚČ騀 tsünt澘 ĭüÄÈo tÚĜėnÈt澘 o®ėĜČ®n ÝoÄÝÔü ÄijÈ è®Čt LÝÔÚėÝoÔ澙 Sühė ÄÝhĜtÈÄ oüČėÚ nÝÔÚė澘 ¾nÈ®Č ¾ünüČ ÄÝhÈČÈoėÝ®ėÝüo B®¾ëÔČüĜoÄ澙 P®ėÈČoÈÄ ëÝóüoü ėÈIJėÝnÈt tė®¾ëÈÄ üo ĭüÄÈo Č®¾ët CÚ®oÔÈ ėÚÈ ĭüČëts®¾Èėü ® ĭ®Čó ü®ë ¾ÝČ¾Ĝn®Č ė®dnÈ ĭÝėÚ ® o®ėĜČ®n ĭüÄ ÔČ®Ýo tĜČh®¾È澗 CÚ®oÔÈ ėÚÈ ė®ČÔÈė üdèȾėtėü ® ÔünÄ hüÝn瀨ĭČ®sÈÄ ¾üoė®ÝoÈČ ĭÝėÚ ® ¾ČÝoënÈÄ ėÈIJėĜČÈ澗 CÚ®oÔÈ ėÚÈ ė®të üdèȾėtėü ė®o nÈ®ėÚÈČ ÚÝÄÈ ĭÝėÚ Ä®Čë tėÝė¾ÚÈÄ tÈ®ót澗 CÚ®oÔÈ ėÚÈ nÝÔÚė ėü ĜnėČ®yÝünÈė瀨tėijnÈ nÝÔÚėÝoÔ ÄČÈo¾ÚÈt ėÚÈ t¾ÈoÈ澗 CÚ®oÔÈ ėÚÈ ĭüČëts®¾Èėü ® ĭüÄÈo ė®dnÈ ĭÝėÚ óÈÄÝĜó dČüĭo hÝoÝtÚ ®oÄ yÝtÝdnÈ ĭüÄ ÔČ®Ýo ėÈIJėĜČÈ澗 Figure 2 Overview of all tasks in Xiaomi-Robotics-U0, Gray: inputs; Orange: outputs. 3 dynamics. More importantly, this enables embodied generation to evolve from static scene synthesis into a scalable trajectory generation engine: generated embodied scenes can be directly rolled out into temporally coherent manipulation videos, providing abundant synthetic trajectories for downstream policy learning. We believe that this capability represents an important step toward agentic embodied data generation, where foundation world models continuously generate increasingly diverse robot experiences through autonomous interaction rollout. Extensive experiments demonstrate thatXiaomi-Robotics-U0establishes a new state of the art across both single-step and sequential embodied generation tasks. It is the first unified embodied synthesis model capable of high-quality multi-view embodied scene generation, structured controllable embodied transfer, and zero-shot embodied video generation within a single autoregressive framework.Xiaomi-Robotics-U0outperforms GPT- Image-2.0 in human evaluation on embodied scene generation and embodied transfer, achieves state-of-the-art performance on the World Arena benchmark, and significantly improves the out-of-distribution robustness of downstream robot policies using generated data. These results suggest that world foundation models can serve not only as embodied world models, but also as scalable data engines for embodied intelligence. Our contributions are summarized as follows: • We introduceXiaomi-Robotics-U0, the first unified embodied synthesis model that seamlessly integrates foundation image generation, image editing, embodied scene generation, embodied transfer, and embodied video generation within a single autoregressive framework. By jointly optimizing general-domain and embodied generation tasks under a unified continual training paradigm,Xiaomi-Robotics-U0effectively transfers the generalization capability of foundation world models to embodied generation while preserving their original visual generation ability. •We propose the first unified framework for multi-view embodied generation, including embodied scene generation and controllable embodied transfer. By jointly training with large-scale text-to-image and image editing data,Xiaomi-Robotics-U0achieves state-of-the-art performance with superior multi-view consistency, fine-grained controllability, and instruction following, outperforming GPT-Image-2.0 in human evaluations. Furthermore, we introduce a structured control formulation that disentangles workspace, background, foreground irrelevant objects, target objects, and lighting into independent control dimensions, enabling scalable embodied video augmentation while preserving geometric consistency and interaction dynamics. The generated data substantially improves the robustness of downstream robot policies on challenging real-world manipulation tasks. •We extend embodied generation from static scene synthesis to sequential embodied world modeling through interleaved subtask-subgoal learning and multi-FPS embodied video generation.Xiaomi-Robotics-U0 achieves state-of-the-art performance on the World Arena benchmark and demonstrates strong zero- shot multi-view embodied video generation capability. By enabling coherent long-horizon rollout from synthesized embodied scenes,Xiaomi-Robotics-U0transforms foundation world models into scalable embodied data engines, providing a practical path toward agentic generation of diverse embodied interaction trajectories for future policy learning. 2 Xiaomi-Robotics-U0 2.1 Problem Formulation Our goal is to develop a unified embodied world model that supports diverse generation and prediction tasks under a single autoregressive framework. Given a multi-modal context C =x 1 ,x 2 ,· ,x n ,(1) where each element can be a text token, an image token, or a robot-related control token, the model learns the joint distribution over the subsequent multi-modal sequence P(Y|C) = T Y t=1 P(y t |y <t ,C),(2) 4 using standard next-token prediction (NTP). To support embodied intelligence,Xiaomi-Robotics-U0is trained on several complementary data formats with different temporal structures. Specifically, we categorize all training samples into two groups. Single-step samples describe a one-shot mapping between inputs and outputs without explicit temporal evolution. These samples include: (i)Text-to-Image(T2I), where the model generates an image conditioned on textual descriptions; (i)Any-to-Image(X2I), where one to three reference images together with text are used for image editing and generation; (i)EmbodiedSceneGeneration, where the model generates the initial multi-view robot observations conditioned on robot embodiment and scene descriptions; and (iv) EmbodiedTransfer, where the model predicts transferred multi-view observations given the current embodied observations and a target scene description. Sequential samples model long-horizon interactions and future evolution. We consider two representative formats: (i) image-text interleaved subtask-subgoal sequences, where each subtask description is followed by the corresponding multi-view observations after completion; and (i) embodied manipulation videos collected at multiple frame rates, enabling the model to capture both sparse task-level transitions and dense interaction dynamics. Despite their different semantics, all tasks are represented as unified multi-modal token sequences and optimized under the same autoregressive objective, allowing knowledge to be naturally shared across image generation, scene understanding, embodied reasoning, and future prediction. ÔÝĬÈ 濮Büt濯 ½ÝČėÚėü ėüÔÝĬȽÝČėÚ ėÚÈ XR ėÚÈ XR U溶 濮Eüt濯 ...... ...... 濮EüĐ濯 Sė®õÄ®ČÄ AR IõhÈČÈõ¾È XÝ®üóÝ NÈIJė TüëÈo PČÈÄݾėÝüo TČ®ÝõÝõÔ IõhÈČÈõ¾È U溶 XÝ®üóÝ Fn®tÚAR IõhÈČÈõ¾È 炄 Figure 3 Overview structure ofXiaomi-Robotics-U0. The model is trained end-to-end at scale with a unified next-token prediction objective. An efficient inference extension is provided along. 2.2 Model Architecture Figure 3 illustrates the overall architecture ofXiaomi-Robotics-U0. We adopt the open-sourced EMU3.5 [16] as the initialization model, which is built upon the Qwen-3-32B decoder-only Transformer and has demonstrated strong capabilities in image generation and image-text interleaved modeling. Images are first tokenized by the IBQ Tokenizer [51] with a spatial compression ratio of 16×16. By extending the original Qwen vocabulary with the IBQ codebook, all modalities are represented within a unified discrete vocabulary and modeled using next-token prediction. Compared with conventional embodied world models that introduce task-specific prediction heads, Xiaomi-Robotics-U0 formulates all generation tasks—including image generation, embodied scene synthesis, scene transfer, sequential subtask prediction, and video generation—as a single autoregressive sequence 5 modeling problem. This unified formulation enables efficient multi-task co-training while preserving the strong visual generation capabilities inherited from the large-scale pre-trained model. To improve inference efficiency, we further adopt an inference-accelerating adaptation inspired by FlashAR [73]. Specifically, an additional vertical prediction head is introduced to decode image tokens in an anti-diagonal order, allowing multiple visual tokens to be generated simultaneously. Combined with optimized KV-cache management and continuous batching provided by vLLM [26],Xiaomi-Robotics-U0achieves up to 82.9×faster image generation at 1024×1024 resolution compared with the original NTP pattern, while preserving most generation quality. 2.3 Dataset 2.3.1 Dataset Curation Data Collection. To endowXiaomi-Robotics-U0with broad visual understanding and generation capabilities, we assemble a diverse training corpus spanning six distinct domains, each contributing complementary visual and semantic priors. • General image-text data provides foundational visual-semantic alignment. We include ShareGPT4V [9] and similar large-scale vision-language benchmarks to establish a strong baseline for text-conditioned generation. •Embodied manipulation data bridges the gap between passive visual understanding and active physical interaction. This domain combines both real-world and simulated sources. On the real side, we leverage AgiBotWorld-Beta [8], Open X-Embodiment [15], and a proprietary multi-view robot manipulation dataset collected in house, which we refer to as MiBot. On the simulation side, we incorporate data from RoboTwin2.0 [11], GenieSim [69], and InternData-A1 [57], enabling coverage of scenarios that are difficult or dangerous to collect physically. •Autonomous driving data captures structured outdoor scenes with complex spatial layouts and dynamic objects. We source this from Cosmos-Drive-Dreams [47], which provides high-quality driving scenarios with diverse weather, lighting, and traffic conditions. • Egocentric data, drawn from EgoWalk [2], supplies first-person perspective recordings that reflect natural human viewpoint and motion patterns, enriching the model’s understanding of human-centric visual experience. •3D reconstruction data offers dense multi-view correspondences and accurate geometry. We utilize ScanNet++ v2 [68] and DL3DV-10K [31], both of which provide high-quality indoor and outdoor 3D scene captures. •Game data, sourced from Open-P2P [71], introduces stylized and procedurally generated environments that expand the diversity of visual distributions beyond real-world captures. Data Filtering. Raw data collected at scale inevitably contains artifacts that can impair generation quality. Inspired by Abot Physical World [14], we implement an autonomous data filtering pipeline including motion blur detection, image quality detection, and gripper-state alignment. To sum up, we collect 9.5 million samples for single-step generation, which corresponds to 56.4 billion tokens. For sequential generation, we select 2.6 million video clips, which corresponds to 49.6 billion tokens. 2.3.2 Dataset Labeling Unified Annotation Pipeline. Most raw samples lack fine-grained text annotations and only carry coarse task-level descriptions, which cannot support precise text-driven embodied generation. We thus construct a unified annotation pipeline to standardize supervision signals across heterogeneous data. Unified Annotation Pipeline. We leverage Qwen3-VL-235B [3], a powerful vision-language model (VLM), to generate dense, structured annotations for generic frames, multi-source video sequences, and robotic manipulation trajectories. The pipeline consists of four decoupled branches with distinct annotation objectives: 6 C®ó瀨CėČn EÄÝėTÝóÈ瀨CėČn EÄÝė SėČĜ¾ėĜČÈ EIJėČ®¾ėÝüo EódüÄÝÈÄ S¾ÈoÈ GÈoÈČ®ėÝüo GÈoÈČ®n T溸I 煥 X溸I EódüÄÝÈÄ TČ®otÓÈČ Figure 4 The illustration of the constructed single-step generation dataset. Except for the embodied transfer and the embodied scene generation data, we additionally introduce Text-to-Image(T2I) and Any-to-Image(X2I) data for co-training. Concretely speaking, we add camera-control and time-control data to improve the spatial and temporal understanding capability. For structure and geometry perception, we construct skeleton, depth, edge, and semantic mask datasets with advanced computer vision tools. To maintain the ability of the pre-trained model, we introduce general T2I and X2I datasets for co-training. 7 •General Scene Annotation All video data spanning general visual corpora and embodied robot sequences are processed without domain restrictions. Each video clip undergoes temporal downsampling to select high-clarity keyframes. The VLM produces holistic captions describing overall scene layouts, object interactions, and cross-object spatial relationships for every sampled frame. • Structured Embodied Scene Annotation Annotation is only performed on embodied manipulation data. Decomposed subtask outputs from trajectory processing guide scene partitioning into five orthogonal semantic dimensions: workspace, task objects, irrelevant objects, lighting, and background. Separate descriptions are generated for each dimension according to task relevance, supporting compositional reasoning and fine-grained controllable generation during inference. •Geometric Modality Extraction Temporally consistent inverse depth maps are extracted for all frames and videos via Video Depth Anything [10] to supply geometric conditioning signals for cross-domain embodied transfer tasks. Brighter pixels correspond to closer surfaces and darker pixels to distant regions within this depth representation, allowing the model to synthesize photorealistic RGB outputs conditioned on geometric priors. •Trajectory-level Subtask Decomposition As illustrated in Figure 5, HDBSCAN clustering [37] segments robot trajectories using end-effector poses and gripper signals to identify motion and grasp transition boundaries. Keyframes with dense physical interactions are sampled at these boundaries and sent to VLM to generate temporally aligned subtask plans with frame boundaries. Transition frames with rich physical interactions are processed with scene annotation to produce training samples for embodied scene generation and embodied transfer tasks. Task-Oriented Data Organization. To support the three training objectives ofXiaomi-Robotics-U0, we reorganize all annotated data into task-specific formats: •Embodied Transfer. Each training sample consists of a system instruction, a scene description, and multi-view depth maps as input, with multi-view RGB images as the supervision target. This task trains the model to hallucinate photorealistic appearance from geometric cues. • Scene Generation. Each sample pairs a system instruction with a scene description, supervised by multi-view RGB images. This task learns text-to-multi-view-scene synthesis without geometric priors. •Video Generation. Each sample comprises a system instruction, a task instruction, and an initial observation frame, supervised by a sequence of 15–25 RGB frames. For datasets that provide robot action labels, a sequence of robot mask images rendered from the robot action and URDF is additionally appended after the initial observation frame, serving as action guidance. This task learns to predict future visual states conditioned on task semantics and initial context. Data Balancing. Our training corpus contains heterogeneous multi-domain data, which easily leads to domain bias and overfitting toward dominant data distributions. We adopt sample-level reweighting conditioned on data domains and generation task types (multi-view generation, scene editing, single-view generation), with the core target of maintaining full coverage of robotic manipulation scenarios. Specifically, we first eliminate invalid embodied episodes, including sequences with corrupted visuals, misaligned timestamps, missing proprioceptive or action signals, failed execution records, and static trajectories lacking robot-object interactions. All remaining samples are categorized via hierarchical annotations derived from subtask and structured scene labeling. Grouping criteria cover task goals, scene contexts, target objects, spatial relations, subtask phases, and atomic manipulation skills. Such fine-grained grouping supports more precise balancing than simple dataset-level weighting: redundant samples sharing identical embodiment-task-object combinations are downsampled, whereas clips with novel visuals, object layouts, manipulation skills or temporal stages are prioritized. Higher sampling weights are assigned to long-tail categories, including rare objects, seldom-used manipulation skills, and core interaction phases (grasping, releasing, placing, and object state transitions). The resulting training distribution eliminates redundant repeated demonstrations while retaining diverse visual scenes and comprehensive manipulation skill coverage. 8 S¾ÈoÈ DÈt¾ČÝsėÝüo GÈoÈČ®ėÝüo WüČëts®¾ÈOdèȾėtLÝÔÚėÝoÔB®¾ëÔČüĜoÄ A nÝÔÚė dČüĭo ĭüÄÈo ė®dnÈ ĭÝėÚ yÝtÝdnÈ ĭüÄ ÔČ®Ýo ėÈIJėĜČÈ ®oÄ ó®ėÈ tĜČh®¾È澗 TÚÈĭüČëts®¾ÈÝt®ė®dnÈ ¾üyÈČÈÄĭÝėÚ®sĜČsnÈ ®oÄĭÚÝėÈ澛sČÝoėÈÄoÈ®Č ėÚȾÈoėÈČ澗 T®të ü½èȾėt澙 A dn®¾ë t¾®oÈČ ĭÝėÚ澛 IČÈnÈĬ®oė ü½èȾėt澙 A ČÈÄ ëÈė¾ÚĜs düėnÈ澛 T®të ü½èȾėt澙 Adn®¾ëtó®ČėsÚüoÈ澛 IČÈnÈĬ®oė ü½èȾėt澙 NüĜoČÈn®ėÈÄüdèȾė澛 BČÝÔÚė üyÈČÚ騀 nÝÔÚėÝoÔ ¾ČÈ®ėÝoÔ tÚ®Čs tÚ®Äüĭt üh ėÚÈ Čüdüė ®Čót ®oÄ üdèȾėt üo ėÚÈ ĭüČëts®¾È澗 TÚÈnÝÔÚėÝoÔÝtdČÝÔÚė澘 ÄÝhĜtÈ澘®oÄoÈĜėČ®nÝo ¾ünüČ澛 ėÚÈüdèȾėtüo ėÚÈė®dnÈtĜČh®¾È澗 A nÝÔÚė ÔČ®ij ėÝnÈÄ hnüČ ĭÝėÚ tsȾënÈÄ ėÈIJėĜČÈ dÈnüĭ 澛 ®Čüĭt Ýo ėÚÈ h®Č ČÝÔÚė d®¾ëÔČüĜoÄ澗 OoėÚÈnÈhėtÝÄÈ澘ėÚÈČÈ®ČÈ ĭÚÝėÈnȾėČݾ®nüĜėnÈėt ®oÄ澛sÝsÈČĜotÄüĭo ėÚÈĭ®nnoÈ®ČėÚȾÈoėÈČ澗 SĜ½ė®të DȾüósütÝėÝüo CnÝĉ SÈÔóÈoė®ėÝüoKÈij瀨ÓČ®óÈ S®óĉnÝoÔSĜdė®të GÈoÈČ®ėÝüo 灃ÝÄ灃澙 溷澘 灃tė®Čė灃澙 溶澘灃ÈoÄ灃澙 溷溺溻 瀶tĜdė®të灃澙 灃AssČü®¾Ú sn®ėÈ Č®¾ë ĭÝėÚ ėÚÈ ČÝÔÚė ®Čó灃 灃ÝÄ灃澙 溸澘 灃tė®Čė灃澙 溷溺溼澘灃ÈoÄ灃澙 溷溽溼 瀶tĜdė®të灃澙 灃GČ®ts sn®ėÈ ĭÝėÚ ėÚÈ ČÝÔÚė ®Čó灃 灃ÝÄ灃澙 溹澘 灃tė®Čė灃澙 溷溽澘灃ÈoÄ灃澙 溸溷溺 瀶tĜdė®të灃澙 灃LÝhė sn®ėÈ ĭÝėÚ ėÚÈ ČÝÔÚė ®Čó灃 瀵ÝÄ瀶澙 溻澘 瀵tė®Čė瀶澙 溹溸溽澘瀵ÈoÄ瀶澙 溹溿溺 瀶tĜdė®të灃澙 灃RÈėĜČo ėü tė®Čė ĭÝėÚ ½üėÚ ®Čót灃 灃ÝÄ灃澙 溺澘 灃tė®Čė灃澙 溸溷溻澘灃ÈoÄ灃澙 溹溸溼 瀶tĜdė®të灃澙 灃Pn®¾È sn®ėÈ ĭÝėÚ ėÚÈ ČÝÔÚė ®Čó灃 MĜnėÝ瀨VÝÈĭO½tÈČĬ®ėÝüo GČÝsÈČSÝÔo®n RÝÔÚėAČóA¾ėÝüoCnĜtėÈČÝoÔ 溶 溷溺溻 溷溽溼 溸溷溺 溹溻溸 溹溿溺 CüódÝoÈÄA¾ėÝüoSÈÔóÈoė®ėÝüo 溶 溷溺溻 溷溽溼 溸溷溺 溹溸溼 溹溻溸 溹溿溺 T炈溶 T炈溷溺溻 T炈溷溽溼 T炈溸溷溺 T炈溹溻溸 T炈溹溿溺 CnütÝoÔ 濂 GČ®ts OsÈoÝoÔ 濂 RÈnÈ®tÈ 溶 溹溸溼 溹溿溺 LÈhėAČó A¾ėÝüoCnĜtėÈČÝoÔ 溼DEESė®ėÈ Figure 5 Overview of the unified data labeling pipeline. The upper part illustrates scene description generation, where each sample is annotated with structured semantic attributes. The lower part shows trajectory-level subtask decomposition, including action-based clip segmentation, key-frame sampling, and subtask generation. This pipeline provides consistent scene, geometry, and action-level supervision across heterogeneous robotic datasets. 2.4 Training To efficiently adapt the pre-trained image generation model into a unified embodied world model while preserving its original visual generation capability, we adapt two continued training setups:single-step training andsequentialtraining. Single-step training. This setup focuses on aligning the model with diverse one-step generation tasks while maintaining the original image generation capability. We jointly train four complementary datasets, including Text-to-Image (T2I), Any-to-Image (X2I), Embodied Scene Generation, and Embodied Transfer. The T2I and X2I datasets continuously reinforce the general visual generation and image editing capabili- ties inherited from the pre-trained model, preventing catastrophic forgetting during embodied adaptation. Meanwhile, Embodied Scene Generation teaches the model to synthesize physically plausible initial robot observations conditioned on robot embodiments and scene descriptions, while Embodied Transfer enables cross-scene multi-view generation that preserves robot interaction states and camera geometry. Co-training these tasks effectively aligns the visual generation prior with embodied scene understanding under a unified autoregressive objective. Sequential training. The sequential training extends the model to long-horizon embodied reasoning and interaction modeling. We first introduce image-text interleaved subtask-subgoal sequences, where each language-described subtask is paired with the corresponding multi-view observations after execution. This data format teaches the model to reason over task decomposition while predicting future embodied observations, further aligning the sequence modeling capability inherited from the base model. 9 To enhance dynamic interaction modeling, we additionally incorporate large-scale embodied manipulation videos collected from InternData-A1 [57], GenieSim [69], AgiBotWorld-Beta [8], RoboCOIN [62], and Robo- MIND [61]. Unlike conventional low-frame-rate trajectory datasets, we construct multi-frame-rate video sequences with FPS values of 1, 3, and 5. Sparse sequences encourage long-horizon planning and future prediction, whereas denser sequences enable the model to capture fine-grained manipulation dynamics and continuous interaction processes. Training across multiple temporal resolutions significantly improves the model’s ability to model both high-level task progression and low-level embodied dynamics within a unified autoregressive framework. 2.5 Inference To further accelerate inference, we introduce a FlashAR + extension toXiaomi-Robotics-U0. The extension keeps the conditional modeling of reference images and other input signals unchanged, while applying parallel autoregressive generation only to the final target image region. This design is intended to preserve the multi-source conditioning capability ofXiaomi-Robotics-U0and reduce the serial decoding cost during X2I inference. 2.5.1 FlashAR + Extension FlashAR baseline. FlashAR [73] proposes a post-training and inference acceleration paradigm for autoregressive text-to-image generation. Its core idea is to preserve the generation capability of the original model while introducing extra vertical prediction, so that image tokens can be generated in parallel along anti-diagonal groups. This reduces the number of serial decoding steps in the image region. Inspired by FlashAR, we perform FlashAR + post-training on top ofXiaomi-Robotics-U0, extending it from single-image T2I generation to broader X2I scenarios, and use the resulting model for X2I inference acceleration inXiaomi-Robotics-U0. Figure 6a demonstrates the decoding steps of the FlashAR + extension. InXiaomi-Robotics-U0, X2I covers tasks such as T2I, I2I, and multi-reference image generation. The input conditions may include text instructions and one or more reference images. u4PM-Vu4PM-Ou4PM-Su4PM-6 u4PM-D u4PM-E u4PM-X UN[]0NH4zd-UPz5 6 P [ 4 ] 7 z d - U P z 5 VO O S S S 6 6 6 6 D D D E E X (a) Diagonal step schedule. txt ! img ! img " s # s ! s " s $ s % s & s ' txt ! img ! img " s # s ! s " s $ s % s & s ' s # s # s ! s ! s " s " txt " txt " s $ s $ (b) Prefix-conditioned step-causal mask. Figure 6Xiaomi-Robotics-U0FlashAR + extension for X2I generation. Target image tokens are grouped by diagonal steps, while all input conditions are kept in the prefix and the target image is generated with a step-causal attention mask. FlashAR + mask and attention. Compared with standard autoregressive generation,Xiaomi-Robotics-U0 FlashAR + Extension introduces the key changes summarized in Table 1. During training, the FlashAR + extension identifies image blocks during data parsing and defines the target image. All preceding content, including text instructions, reference images, control conditions, and the target image header, is treated as the conditional prefix. Parallel autoregressive losses are computed only over the target image region, while reference images serve only as a conditional context. This prevents reference images 10 Table 1 Comparison between standard autoregressive generation and theXiaomi-Robotics-U0FlashAR + Extension. AspectStandard AR Xiaomi-Robotics-U0 FlashAR + Extension Effect Target image generation Serial token-by-token generation in raster order Parallel generation by anti-diagonal steps in the target image Reduces serial decoding steps Target image attention Standard one-dimensional causal attention The conditional prefix is visible; the target image uses two-dimensional step-causal attention internally Prevents information leakage Prediction headsA single AR prediction headHorizontal head, vertical head, and gated fusion Supports anti-diagonal generation Directional auxiliary loss Standard supervisionIntroduces H/V loss constraintsConstrains both directional heads from being incorrectly supervised as generation targets and ensures that the model learns to generate the target image from multi-source conditions. The attention constraint is also restricted to the target image region, as shown in figure 6b. The conditional prefix preserves standard autoregressive visibility. Target image tokens can attend to the entire conditional prefix, but within the target image they can only attend to tokens from earlier anti-diagonal steps. For a target grid position (r,c), its parallel step is defined asr+c; tokens within the same step are mutually invisible, which prevents information leakage during parallel generation. Reference images, text tokens, and control tokens are not involved in the anti-diagonal mask construction inside the target image, preserving the semantic integrity of X2I conditions. This constraint can be expressed as a visibility function within the target image. For a target-image query token q = (r q ,c q ) and key token k = (r k ,c k ), M(q,k) = 1 [s(r k ,c k ) < s(r q ,c q )]. FlashAR + loss.Xiaomi-Robotics-U0FlashAR + Extension introduces a key adjustment to directional auxiliary supervision. We match the horizontal auxiliary loss weight to the vertical auxiliary loss weight: L =L fuse + 0.05L h + 0.05L v + 0.2L distill . Here,L fuse is the target-image prediction loss after H/V gated fusion,L h andL v supervise the horizontal and vertical heads respectively, andL distill is the AR distillation term. For a target position (r,c), letz h (r,c) and z v (r,c) denote the logits from the horizontal and vertical heads, and letg(r,c)∈[0,1] denote the gate value. The fused logits are z fuse (r,c) = g(r,c)z h (r,c) + (1− g(r,c))z v (r,c). The corresponding fused loss is L fuse =− X (r,c)∈Y log softmax (z fuse (r,c)) y r,c . The directional auxiliary lossesL h andL v use the same target tokens to supervisez h andz v , respectively. The H head also needs to adapt to the target-region distribution. IfL h = 0, the H head may fail to converge 11 EmbodiedSceneGenerationEmbodiedTransfer AR FlashAR + (vLLM) 450.77 16.56 5.44 10³ 10² 10¹ 10⁰ 27.22x 3.04x ARFlashAR⁺ FlashAR⁺ (vLLM) Generation Time (s/img) EfficiencyComparision GenEvalImgEdit Evaluation Results 1 0 0.8 0.6 0.4 0.2 0.74 0.71 4.08 4.08 GenEvalandImgEditResults AR FlashAR⁺ Qualitative Comparison Quantitative Comparison 5 0 4 3 2 1 Figure 7 Qualitative and efficiency comparison between AR, FlashAR + , and FlashAR + with vLLM. FlashAR + preserves high visual quality on Embodied Transfer and Embodied Scene Generation while substantially reducing generation latency. sufficiently, and the H/V gate may collapse toward the H branch in later training stages. When the gate collapses to an under-trained H head, the fused prediction degenerates into an unstable single-direction prediction, and further training can lead to degraded generation quality. Adding 0.05L h explicitly supervises the H head and stabilizes H/V fusion. Xiaomi-Robotics-U0FlashAR + inference. During inference,Xiaomi-Robotics-U0first encodes the text instruc- tion and reference images as a conditional prefix, followed by the target image header. The model then performs anti-diagonal generation only on the target image grid: positions are generated step by step in increasing order ofr+c, and all positions within the same step are sampled in parallel. Boundary positions use the available single-direction prediction, while interior positions use H/V gated fusion. This procedure preserves the constraints imposed by multi-source conditions while extending image-region parallel decoding toXiaomi-Robotics-U0FlashAR Extension inference. As a result, the model maintains Any-to-Image conditional generation capability while reducing the serial generation cost in the target image region. 2.5.2 vLLM Integration To further improve the image-generation efficiency ofXiaomi-Robotics-U0, we integrate vLLM on top of FlashAR + . The implementation keeps the FlashAR + decoding rule unchanged: vLLM handles the conditional prefix, batched execution, and paged KV-cache management, while a diagonal scheduler emits the visual tokens of each anti-diagonal step and synchronizes the paired conditional and unconditional requests used for 12 classifier-free guidance. As shown in Figure 7, the accelerated path preserves high generation quality on embodied tasks. On T2I generation at 1024×1024 resolution with a single H20 GPU, standard AR takes 450.77 seconds per image. FlashAR + reduces this latency to 16.56 seconds per image, achieving a 27.22×speedup over AR, and FlashAR + with vLLM further reduces the latency to 5.44 seconds per image withmax_num_seq=28, achieving an additional 3.04× speedup over FlashAR + . 3 Experiments Our model unifies embodied transfer, embodied scene generation, and embodied video generation while retaining general image editing and generation capabilities. This section presents benchmark and qualitative results for our embodied transfer and scene generation tasks, which outperform GPT-Image-2.0 in human evaluation. Specifically, our embodied transfer method substantially enhances robot trajectory diversity while preserving geometric consistency, boosting the success rate ofπ 0.5 from 36.9% to 63.2%. We further validate the model’s general image generation capacity via evaluations on the Geneval and ImageEdit benchmarks, confirming that our method retains dominant general image generation performance. For embodied video generation, our model achieves state-of-the-art results on the World Arena benchmark. It also demonstrates strong generalization to self-synthesized initial scenes, enabling a closed-loop data engine for infinite video generation conditioned on arbitrary scene descriptions and robotic instructions. 3.1 Embodied Transfer Benchmark and protocol. To evaluate embodied transfer, we construct a benchmark of 300 samples, each consisting of a textual scene description together with a set of multi-view depth maps; the model is required to synthesize multi-view-consistent RGB images of the described robot manipulation scene. Each sample additionally provides the multi-view RGB reference images from the original dataset. These references do not fully correspond to the (edited) scene description, but they capture the structural information beyond depth and are used to compute the benchmark metrics. Since the depth maps are drawn from the same domain as the training set, we deliberately edit a subset of the scene descriptions so that they call for content not present in the training data, thereby probing the model’s general-purpose generation ability rather than mere reconstruction. According to how much the edited description deviates from the original scene, we split the benchmark into 150Easysamples (descriptions close to the underlying scene) and 150Hard samples (descriptions requiring substantial novel content). We compare against GPT-Image-2, a strong image generation model. Metrics. We assess generation quality from three complementary aspects, all computed against the reference views. For depth consistency, we extract monocular depth from both the generated image and the reference with a shared estimator and report SI-RMSE (scale-invariant RMSE,↓),δ 1 (the fraction of pixels with relative depth error below 1.25,↑), and AbsRel (absolute relative error,↓); these measure how faithfully the generated image respects the 3D geometry imposed by the input depth. For structural fidelity, we compute Canny edge F1, Precision, and Recall (↑), capturing whether the generated layout preserves the fine-grained contours of the scene. For semantic grounding, we run open-vocabulary detection and segmentation and report Segmentation mIoU and Recall (↑), reflecting whether the described objects are correctly synthesized and localized. As summarized in Table 2,Xiaomi-Robotics-U0outperforms GPT-Image-2 by a large margin across all metrics on both the Easy and Hard splits, indicating markedly better adherence to the input geometry and stronger semantic alignment with the scene description. Qualitative results. Figure 8 presents qualitative examples of embodied transfer. For each example we display the scene description, the original multi-view RGB images from the source dataset, and the images generated byXiaomi-Robotics-U0. Our pipeline first estimates the multi-view depth maps from the original images, then edits the scene description—with the edited spans highlighted—to call for content that departs from the original scene; the extracted depth maps and the edited description are jointly fed to the model for generation. Across diverse scenes, and given a scene description in either natural-language or structured form,Xiaomi-Robotics-U0faithfully follows both the edited textual description and the multi-view depth 13 Table 2 Objective metrics on the embodied transfer benchmark. Values in parentheses denote the change of Xiaomi-Robotics-U0relative to GPT-Image-2 on the same split (signed toward the actual value; a lower value is better for ↓ metrics and a higher value for ↑ metrics). Metric GPT-Image-2 Xiaomi-Robotics-U0 (Ours) Easy HardEasyHard Depth SI-RMSE (↓)0.4007 0.3444 0.1407 (−0.260) 0.1435 (−0.201) Depth δ 1 (↑)0.5307 0.5340 0.8017 (+0.271) 0.7846 (+0.251) Depth AbsRel (↓)0.4350 0.4045 0.1647 (−0.270) 0.1702 (−0.234) Canny F1 (↑)0.0519 0.0507 0.2066 (+0.155) 0.1815 (+0.131) Canny Precision (↑)0.0560 0.0536 0.2037 (+0.148) 0.1770 (+0.123) Canny Recall (↑)0.0572 0.0582 0.2290 (+0.172) 0.2176 (+0.159) Segmentation mIoU (↑) 0.4105 0.4098 0.7621 (+0.352) 0.7436 (+0.334) Segmentation Recall (↑) 0.6208 0.6524 0.8136 (+0.193) 0.8032 (+0.151) !"#$%&'(): A dark gray fabric-covered bed with a wrinkled texture and matte surface. *'%$+ ",-)(.%: A white plastic hanger with a glossy finish and smooth curved shape is held by the robot arm near the center of the workspace... /##)0)1'2.+ ",-)(.%: A brown cotton shirt with a matte surface and visible stitching lies flat on the workspace. A small blue fabric label with white text is stitched on the shirt’s chest area. 3456.425: Warm golden sunlight streams in from the upper right, creating long crisp shadows and glowing surfaces. 7'($5#"829: A frosted glass partition with soft diffused light behind it in the far background. A silver insulated blanket with a crinkled texture on the right edge of the scene. !"#$%&'(): The workspace is a white marble table with soft gray veining and a polished stone texture... *'%$+ ",-)(.%: A white cotton towel with visible looped texture is centered on the workspace, partially grasped by two robotic arms. /##)0)1'2.+",-)(.%: A translucent blue plastic basket with perforated sides is visible on the left side of the workspace... 3456.425: Soft golden light from multiple lamps around the room wraps gently around the robotic arms and towel. The illumination is warm and cozy, with broad diffuse shadows and subtle brightness variation. 7'($5#"829: Distant pale sand with faint wind-rippled bands and a partial view of a weathered wooden crate in the upper left corner. !"#$%&'(): A deep blue felt-covered table with a soft fibrous surface and visible seam lines between panels. *'%$+ ",-)(.%: A brushed stainless steel desk sign with deep blue engraved text 'Bruce' and a satin reflective surface standing upright near the center of the workspace. /##)0)1'2.+",-)(.%: A bouquet of purple feathered flowers with soft layered petals in a woven wicker vase on the left side of the workspace... 3456.425: Bright overhead lighting casting sharp shadows of the robot arm and nameplate on the workspace. 7'($5#"829: A warm cedar wood paneled wall behind the workspace. A low leather lounge chair visible in the far background. A honey-brown plank floor beneath the workspace. InputXiaomi-Robotics-U0 Figure 8 Qualitative results on embodied transfer. For each example we show the scene description, the original multi- view RGB images from which the depth conditions are extracted, and the images generated byXiaomi-Robotics-U0. Starting from the original images, we estimate their multi-view depth maps and edit the scene description (edited text highlighted); the depth maps together with the edited description are then fed toXiaomi-Robotics-U0, which synthesizes multi-view-consistent RGB images that adhere to both the edited text and the input geometry. 14 conditions, correctly rendering the described objects, lighting, and background while respecting the spatial geometry of each view. Crucially, the generated views remain mutually consistent in object identity, appearance, and arrangement, demonstrating that our unified model produces coherent multi-view scenes rather than independently plausible but conflicting images. !"#$%&'(): A white matte table surface with rounded corners. *'%$+ ",-)(.%: A black and gray robotic arm with metallic joints extends from the right side, its gripper positioned above a dark walnut wooden pencil box near the center of the table. The gripper is open, fingers slightly curved, ready to grasp the wooden box. /##)0)1'2.+",-)(.%: A pink analog clock with white face sits to the right of the pouch. A white paper cup with printed patterns is near the top edge of the table... 3456.425: Soft overhead lighting casts subtle shadows. 7'($5#"829: A dark brown wooden shelf with colorful items, a gray concrete floor, and part of a black shelving unit with snacks and supplies on the left. !"#$%&'(): A white rectangular table with a smooth matte surface. *'%$+ ",-)(.%: A white robotic arm with black joints and a black gripper mounted on the right side of the table... /##)0)1'2.+ ",-)(.%: A blue plastic cup sits on the table near the center, slightly to the left of the marker... 3456.425: Soft overhead fluorescent lighting illuminates the scene, casting faint shadows. 7'($5#"829: A rough stone cavern wall with dangling roots and pockets of moss, a damp dark rock floor, a pale limestone storage alcove with silver handles, a black monitor on a desk in the background. A torch-lit tunnel entrance and shelving units with various items are visible in the far background. InputGPT-Image-2Xiaomi-Robotics-U0 !"#$%&'(): A vivid cobalt resin table with swirling synthetic texture and a glossy surface. *'%$+ ",-)(.%: A black robot arm with a matte finish and articulated joints visible on the left side of the workspace. Another black robot arm with a matte finish and articulated joints visible on the right side of the workspace. /##)0)1'2.+",-)(.%: A small orange pumpkin with a matte surface and round shape inside the frying pan at the center of the workspace... 3456.425: Even overhead lighting casting soft shadows beneath objects on the workspace. 7'($5#"829: A plain white wall behind the workspace. A dark gray tiled floor with visible grout lines below the workspace. Figure 9 Qualitative comparison between GPT-Image-2 andXiaomi-Robotics-U0on the embodied transfer benchmark. GPT-Image-2 generates plausible, well-aligned images but violates the input depth and produces multi-view-inconsistent results, whereasXiaomi-Robotics-U0adheres to the geometry of every view and keeps the views mutually consistent. We further compareXiaomi-Robotics-U0with GPT-Image-2 in Figure 9, on three samples from the source datasets, each showing the scene description, the input multi-view depth maps, and the two models’ outputs. GPT-Image-2 produces visually appealing images with strong text–image alignment: the described objects, materials, lighting, and background are all rendered plausibly, and in isolation each view looks convincing. However, it exhibits a poor understanding of the input geometry and of cross-view relationships. Its outputs frequently violate the depth conditions—objects are placed at inconsistent distances, surfaces are warped, and the overall spatial layout drifts away from the imposed geometry—which is reflected in its high SI-RMSE and lowδ 1 in Table 2. Moreover, the views generated for the same scene are largely independent: object identity, pose, and arrangement fail to remain consistent across viewpoints, so the three images do not correspond to a single coherent 3D scene. In contrast,Xiaomi-Robotics-U0faithfully respects the depth of every view while keeping the multi-view outputs mutually consistent, indicating that it not only matches the strong text-following ability of GPT-Image-2 but also genuinely grounds generation in the input geometry. Zero-shot data augmentation. Beyond benchmark evaluation, we leverage the structured text-to-image capability ofXiaomi-Robotics-U0to augment the demonstration trajectories collected for four real-world tasks (pack box, fold towel, pack phone, and pack earphone), aiming to endow downstream policies with better generalization to novel scenes. For each trajectory, we prompt a VLM (Qwen3-VL) to generate, conditioned on the current depth maps and the original scene description, 1000 diverse alternatives for every dimension except the task objects, encouraging maximal diversity across background, workspace, and lighting. We then randomly combine these per-dimension descriptions and feed them toXiaomi-Robotics-U0for generation. As shown in Figure 10 and Figure 11, the resulting scenes exhibit rich diversity, and none of these compositions appear in the training set, demonstrating thatXiaomi-Robotics-U0can substantially expand the visual coverage of the collected trajectories in a zero-shot manner. 15 Figure 10 The zero-shot augmented keyframes of the multi-view manipulation trajectories with the embodied transfer capability ofXiaomi-Robotics-U0. We generate 1000 descriptions for the background, workspace, and lighting condition with Qwen3-VL and replace the original condition randomly. The left panel is the original keyframes, and the right corresponds to the transferred. 16 P®¾ë BüIJP®¾ë PÚüoÈSėüČÈ E®ČĉÚüoÈ Figure 11 More results of the zero-shot augmented keyframes of the multi-view manipulation trajectories with the embodied transfer capability of Xiaomi-Robotics-U0. 17 !"#"$%&'(%)*+,: AgiBot G1. -"'./+01,: A smooth, dark slate tile used for serving sushi acts as a cool, matte black platform. 2#3,1$/: A sharp yanagiba knife with a wooden handle lies on the left. A mound of sticky ... 4567$586: Focused spotlight from a hanging pendant light highlights the fresh ingredients. 901.6'":8;: A minimalist kitchen counter with hidden storage blends into the dark background. A chef ... !"#"$%&'(%)*+,: ARX Robot. -"'./+01,: A slab of dark slate stone, rough-hewn and uneven, used as a serving board ... 2#3,1$/<%A loaf of crusty sourdough bread with a flour-dusted top rests on the left. A wedge of aged ... 4567$586: Warm, inviting light from a hanging Edison bulb above creates a cozy atmosphere with soft ... 901.6'":8;: Exposed brick walls and open wooden shelves filled with jars of spices and dried pasta ... !"#"$%&'(%)*+,: Agilex Piper. -"'./+01,: A flat-topped termite mound in the African savanna serves as an earthen workbench... 2#3,1$/: Specimen jars with insects, field notebooks, and compasses rest on the rough, textured soil... 4567$586: The harsh, direct sunlight of the equator creates high contrast and deep, dark shadows... 901.6'":8;: Acacia trees and grazing zebras are visible against a vast, hazy horizon ... !"#"$%&'(%)*+,: Agibot G2. -"'./+01,: The stainless steel tray of a mobile street food cart specializing in sugary treats... 2#3,1$/: Molds for sugar art, bags of colored sanding sugar, and skewers of candied fruit clusters... 4567$586: Vibrant, multi-colored LED strip lights reflect off the sticky, glossy surfaces... 901.6'":8;: A bustling night market scene with blurred lanterns and crowds of people in motion... Figure 12 Qualitative results on embodied scene generation. 3.2 Embodied Scene Generation Benchmark and Protocol. To evaluate embodied scene generation, we construct a dedicated benchmark using LLM-generated scene prompts. The benchmark contains 400 samples split equally into two difficulty levels: 200 Easy samples for conventional indoor manipulation scenes, and 200 Hard samples for unconstrained open-domain scenes. Each prompt specifies a robot arm type and a structured scene description covering workspace layout, object configuration, lighting, and background. We compareXiaomi-Robotics-U0against GPT-Image-2 as the baseline. To ensure a fair comparison, multi-view reference images corresponding to each target robot arm type are provided to GPT-Image-2 as visual context. Human-Based Pairwise Evaluation. Quantifying cross-view geometric consistency remains an open challenge, as no automated metric reliably evaluates global 3D coherence across viewpoints. We thus adopt a human pairwise preference protocol for quantitative assessment. Annotators are presented with paired outputs fromXiaomi-Robotics-U0and GPT-Image-2 in randomly shuffled left-right orders to mitigate positional preference bias. Annotators select the superior result based on two criteria: (1) multi-view geometric consistency, measuring unified 3D layouts with invariant object positions, scales, and occlusion across views; (2) textual instruction following, measuring alignment with the workspace, objects, lighting, and background in prompts. Figure 14 reports the overall win rate.Xiaomi-Robotics-U0outperforms GPT-Image-2 in most comparisons, achieving clear advantages in multi-view consistency while maintaining comparable instruction-following performance. Qualitative comparisons in Figure 13 show that GPT-Image-2 produces high-quality images but suffers from severe cross-view geometric inconsistencies, with conflicting object positions, scales, and layouts across cameras. By contrast,Xiaomi-Robotics-U0generates geometrically coherent multi-view observations with consistent spatial layouts, while retaining photorealistic quality. This verifies thatXiaomi-Robotics-U0 simultaneously achieves high visual fidelity and robust cross-view consistency. Qualitative Results. Figure 12 presents qualitative results covering four robot arm types.Xiaomi-Robotics-U0 18 !"#"$%&'(%)*+,: AgiBot G1. -"'./+01,: A light brown wooden dining table with a visible wood grain pattern. The surface has a glossy varnish finish that reflects ambient light. It is large and square-shaped, occupying the central foreground. 2#3,1$/: A white ceramic bowl containing blueberries is positioned slightly to the right of center. Two silver spoons lie parallel to each other on the left side. A folded blue linen napkin rests near the back edge of the table. 4567$586: Warm natural sunlight streams in from a window on the right side. The light creates soft gradients and long shadows across the wooden surface. The overall intensity is moderate and diffused. 901.6'":8;: A blurred view of a green garden is visible through a large glass window. Beige curtains hang loosely on either side of the window frame. A section of hardwood flooring is visible in the lower corner. !"#"$%&'(%)*+,: Agibot G2. -"'./+01,: An ancient, cracked scroll of yellowed parchment stretched taut over a bamboo frame acts as the workspace. The paper fibers are visible and rough, stained with centuries of brown ink splatters. 2#3,1$/: A heavy black ink stone with a carved dragon motif sits heavily on the left side. A slender brush with white goat hair and a jade handle lies diagonally across the center. A small red cinnabar seal block rests near the bottom right corner. A pile of gold leaf flakes shimmers loosely near the top edge. An iron paperweight shaped like a mountain range anchors the top left. A tiny ceramic water dropper in the shape of a peach sits beside the ink stone. 4567$586: Soft, warm light from a hanging paper lantern illuminates the scene from directly above, casting diffuse shadows. The light highlights the texture of the paper and the gleam of the gold leaf. 901.6'":8;: Vertical wooden slats form a shoji screen wall, glowing softly from light behind them. Outside the screen, blurred shadows of bamboo stalks sway gently in the wind. The floor is covered with tatami mats featuring a woven grass texture. A low wooden shelf in the far corner holds rows of bound books. The atmosphere is quiet and filled with the scent of old paper. InputGPT-Image-2Xiaomi-Robotics-U0 Figure 13 Qualitative comparison between GPT-Image-2 andXiaomi-Robotics-U0on the embodied scence generation benchmark. EódüÄÝÈÄ S¾ÈoÈ GÈoÈČ®ėÝüoEódüÄÝÈÄ TČ®otÓÈČ 溼溸澗溶炢溼溾澗溻炢溾溸澗溶炢溽溿澗溹炢 E®ĐijH®ČÄE®ĐijH®ČÄ XÝ®üóÝ瀨Rü½üėݾt瀨U溶GPT瀨Ió®ÔÈ Figure 14 Pairwise evaluation results of embodied transfer and embodied scene generation. generates photorealistic and geometrically consistent multi-view initial scenes for diverse scene prompts, demon- strating strong open-ended generalization to arbitrary textual descriptions. This shows that our model extends conventional single-view text-to-image synthesis to multi-view robotic observation generation by learning intrinsic geometric relations between arm configurations and camera projection. Notably, among the four evaluated robot arm types, the AGIBOT G2 is trained exclusively on simulated data, yetXiaomi-Robotics-U0 can still produce photorealistic, real-world-aligned observations for this specific arm type. This validates the model’s promising sim-to-real generalization potential via learned real-world visual priors. 19 GČ®ts C®tÈ OsÈo LÝÄGČ®ts E®ČdĜÄ 溷SėüČÈ E®ČdĜÄ 溷 GČ®ts E®ČdĜÄ 溸SėüČÈ E®ČdĜÄ 溸CnütÈ LÝÄPn®¾È C®tÈ LÝÓė TüĭÈnFn®ėÈo TüĭÈnFÝČtė FünÄ Pn®¾È TüĭÈn SȾüoÄ FünÄ P®¾ë IėÈó 溷P®¾ë IėÈó 溸P®¾ë IėÈó 溹 CnütÈ LÝÄ P®¾ë IėÈó 溺 SėüČÈ E®ČsÚüoÈt FünÄ TüĭÈn P®¾ë BüIJ Figure 15 Stage-level progress visualization for each task. Each row corresponds to one task, and each column shows the head view at the completion of a progress milestone. For Store Earphones, the eight milestones run from grasping the case, through storing both earbuds, to closing the lid and placing the case. For Fold Towel, the milestones cover lifting and flattening the towel, two sequential folds, and placing the folded towel. For Pack Box, they cover four pick-and-place steps followed by closing the box lid. 3.3 Real World Experiments We conduct the real-world evaluation to investigate whether the proposed world model improves downstream policy learning. Specifically, we use the Xiaomi-Robotics-U0 model to augment the expert demonstrations, and compare the VLA policies trained on the original and augmented data across a suite of real-world tasks. Task suite. The benchmark consists of three real-world tabletop manipulation tasks: Store Earphones, Fold Towel, and Pack Box, encompassing diverse challenges such as dexterous manipulation, deformable object handling, and multi-object interaction. Each task is defined by a language instruction and an ordered set of progress milestones. Figure 15 visualizes the milestone states for each task, providing a concrete reference for the progress metric defined later in this section. Policy training and data. For each task in the suite above, we independently post-train aπ 0.5 [23] policy starting from thepi05_basecheckpoint, using its official PyTorch implementation. All training settings are kept identical across compared policies; the only variable is the supervised fine-tuning data mixture. For each task, we collect approximately 40 hours of real-world demonstrations on the robot platform described below, forming the task-specific clean data. To construct augmentation data, we apply the embodied scene generation 20 SėüČÈ E®ČsÚüoÈt FünÄ TüĭÈn P®¾ë BüIJ B®tÈ GČüĜsIoėÈČÓÈČÈo¾È GČüĜs Figure 16 Task layouts for all three tasks. The base group varies the object layout for each task. For Store Earphones, different layouts vary the positions of the case and the earbuds. For Fold Towel, the towel is initialized in a randomized configuration, which is clearer in the left-wrist view shown as an inset at the top-left corner. For Pack Box, a new layout is generated by randomly placing the four objects on the table. The interference group uses the same layout randomization but additionally introduces substantial changes in tablecloth and lighting. pipeline (Section 3.2) to the clean demonstrations, producing approximately 40 hours of style-transferred episodes per task that vary visual appearance—backgrounds, lighting, and textures—while preserving the original robot states and action labels. We compare two resulting per-task policies: Original is post-trained exclusively on the task-specific clean data;Xiaomi-Robotics-U0-Aug is post-trained on the mixture of clean data and its corresponding style-transferred augmentation. No inference-time world-model module is used during robot execution; the world model influences the policy solely through the post-training data mixture. Evaluation schedule. To evaluate robustness beyond the original demonstration distribution while keeping the comparison controlled, we evaluate each policy under two groups of real-world conditions: a base group and an interference group. Each group contains three object-layout variants per task, and both policies are evaluated on an identical condition schedule. In the base group, the three layouts differ only in object placement; the background uses tablecloths seen in the demonstration data, and the scene is evaluated under nominal laboratory illumination. In the interference group, the layouts again vary object placement, but the background is replaced by held-out tablecloths absent from the demonstration data—often with more complex patterns and different textures or surface reflectance. The interference group further introduces held-out illumination, including low light, colored light, and dynamic patterned lighting such as disco-ball effects. The full set of evaluation layouts for every task and group is shown in Figure 16. For every policy, task, test group, and layout, we run three independent real-robot trials. This yields 2×3×3 = 18 trials (two groups × three layouts × three trials) per policy-task pair. Robot setup. We deploy all trained policies on a bimanual ARX robot platform with two 6-DoF arms, each equipped with a 1-DoF parallel-jaw gripper. At each policy query, the observation consists of the current proprioceptive arm state and three RGB views: one head view from the head camera and two wrist views from the left and right wrist cameras. For Store Earphones and Fold Towel, these views are captured with Intel RealSense D405 cameras; for Pack Box, we use fisheye cameras in the same viewpoint configuration to cover the larger packing workspace. The fisheye images are undistorted before use, both as policy inputs and as inputs to theXiaomi-Robotics-U0augmentation. During real-robot evaluation, the policy is running on a separate inference server and communicates with the robot control process over a WebSocket connection on the local-area network. At 1 Hz, the robot transmits the current arm state and the three RGB views to the server, which returns an action chunk that is executed until the next query. The observation stream, 21 020406080100 Task Completion Progress (%) Average Pack Box Fold Towel Store Earphones 81.0 82.1 66.7 64.4 100.0 100.0 76.4 81.9 36.9 63.2 28.9 62.2 51.1 73.3 30.6 54.2 -44.1 -18.9 -37.8 -2.2 -48.9 -26.7 -45.8 -27.7 Original: InterferenceAug: Interference Original: Drop from BaseAug: Drop from Base Figure 17 Comparison of task completion progress between Original andXiaomi-Robotics-U0-Aug, split by evaluation group.Xiaomi-Robotics-U0-Aug is on par with Original under in-distribution conditions (base group) but achieves substantially higher progress under held-out backgrounds and lighting (interference group), leading to a clear overall advantage. communication stack, and action-chunk execution interface are identical across all compared policies. Metric. We use task completion progress as our sole evaluation metric. Rather than scoring only full success, progress assigns partial credit based on the ordered milestones a rollout completes, making it better suited to long-horizon manipulation, where a policy may reliably solve early subgoals yet fail at a later stage. Specifically, if a task hasKordered milestones and the rollout completes the firstℓof them, its progress isℓ/K ∈[0,1], where 1 denotes full success. For tasktand groupg ∈base, interference, the reported progress of the tested policy is defined as: Prog(t,g) = 1 N N X i=1 ℓ i K t , whereNis the number of rollouts for each task and group (three layouts×three trials in our setting). A higher progress score forXiaomi-Robotics-U0-Aug indicates that the world model provides useful post-training variation for improving real-world robustness across both fully and partially completed executions. Experiment results. Figure 17 reports the task completion progress of both policies. In the base group, where test conditions closely resemble the demonstration distribution,Xiaomi-Robotics-U0-Aug performs comparably to Original, with a slight decrease on some tasks—an expected consequence of sharing training capacity with augmented data that does not add new information for in-distribution scenes. In the interference group, however,Xiaomi-Robotics-U0-Aug consistently outperforms Original across all tasks, indicating that style-transferred augmentation induces visual invariance to novel backgrounds and lighting unseen in the clean demonstrations. Overall, the robustness gains under distribution shift outweigh the marginal in-distribution trade-off, which is especially valuable for real-world deployment, where scene conditions cannot be tightly controlled. To illustrate this gap qualitatively, Figure 18 presents representativeXiaomi-Robotics-U0-Aug rollouts under interference conditions for all three tasks. Despite challenging perturbations—held-out tablecloths, 22 SėüČÈ E®ČsÚüoÈt FünÄ TüĭÈn P®¾ë BüIJ Figure 18Xiaomi-Robotics-U0-Aug rollouts under interference scenarios for Store Earphones, Fold Towel, and Pack Box (top to bottom). Within each task, the three rows are the head view, left wrist view, and right wrist view, with frames sampled across the trajectory. colored lighting, and projected patterns—theXiaomi-Robotics-U0-Aug policy still executes the intended manipulation sequences: it stores both earbuds and closes the case in Store Earphones, completes the full folding sequence in Fold Towel, and packs all four objects before closing the lid in Pack Box. Figure 19 further shows the typical failure modes that arise in our evaluation: (1) failing to grasp the earbud under saturated lighting; (2) misidentifying the orientation of the earphone case; (3) failing to locate the towel edges against a visually similar tablecloth; (4) unfolding an already-folded towel due to an incomplete grasp of the cloth edge; (5)(6) attempting to grasp a projected light spot instead of the target object. Under heavily corrupted visual observations, both policies occasionally exhibit these unexpected behaviors; however, Xiaomi-Robotics-U0-Aug is more likely to act correctly despite the perturbations or self-correct after a brief deviation. The Original policy, by contrast, tends to become permanently stuck on visual artifacts or commit unrecoverable errors at a higher rate. 23 瀘溺瀙瀘溻瀙瀘溼瀙 瀘溸瀙瀘溹瀙瀘溷瀙 Figure 19 Typical failure cases across tasks: (1)–(2): Store Earphones, (3)–(4): Fold Towel, (5)–(6): Pack Box. 3.4 Video Generation WorldArena. A key capability of an Embodied World Model is serving as a data engine that generates physically plausible and visually high-quality videos from a single observation and a corresponding instruction. To evaluate this video generation capability for embodiment tasks, we benchmarkXiaomi-Robotics-U0on WorldArena [49]. Given an initial frame along with a language instruction and robot action, WorldArena assesses the generated video across 15 metrics spanning six dimensions: visual quality, motion quality, content consistency, physics adherence, 3D accuracy, and controllability. We formulate this task as processing interleaved vision-language inputs and producing multi-frame interleaved outputs with empty language tokens. The initial frame paired with the language instruction serves as input, along with a sequence of robot mask images rendered from the robot action, camera parameters, and URDF as action guidance. Quantitative results are reported in Table 3 and Figure 21, qualitative visualizations are shown in Figure 22 (rows 1–5). Xiaomi-Robotics-U0achieves the best overall EWMScore (73.64), ranking first among over 100 submitted models. In terms of controllability,Xiaomi-Robotics-U0attains the highest Instruction Following (93.86) and ranks competitively on Semantic Alignment, demonstrating its strong ability to translate language instructions into faithful visual outcomes. For physics adherence,Xiaomi-Robotics-U0leads on Interaction Quality (87.30), indicating that it captures physically plausible object interactions.Xiaomi-Robotics-U0 also achieves the best Perspectivity (98.84) and Motion Smoothness (95.51), reflecting accurate 3D viewpoint geometry and temporally stable motion. Overall,Xiaomi-Robotics-U0strikes the best balance across all evaluation dimensions, confirming its effectiveness as a data engine for embodiment tasks. Embodiment Video Generation. Beyond the simulated WorldArena benchmark, we further leverage the scene generation capability ofXiaomi-Robotics-U0to produce novel initial frames and subsequently generate videos conditioned on language instructions. Representative examples are shown in Figure 20. In the first row, the robot arm correctly interacts with the cup—grasping it, lifting it, and placing it stably on the table, demonstrating accurate contact modeling. In the second row, beyond the correct tabletop manipulation, the model also renders dynamic background changes by simulating a person walking toward the table to observe the robot’s operation, indicating awareness of scene-level dynamics beyond the immediate task. In the third row, the model faithfully simulates the physics of a deformable object (a sheet of paper): the paper bends 24 Figure 20 Qualitative video generation results of Xiaomi-Robotics-U0. realistically when picked up and set down, and exhibits a slight lag during lateral movement due to its low mass, reflecting an understanding of material-dependent physical properties. More visualizations are provided in Figure 22. Figure 21 WorldArena leaderboard (July 13, 2026). UNIS is the anonymous evaluation codename for Xiaomi-Robotics-U0. 3.5 Text2Image & Image Edit Benchmarks We evaluate the autoregressiveXiaomi-Robotics-U0image model on general text-to-image generation and image editing benchmarks. The goal of this evaluation is not to claim thatXiaomi-Robotics-U0is a dedicated open-domain image generator, but to quantify how much general image generation and editing ability is retained after adapting the model toward embodied scene synthesis. We use GenEval [20] for compositional T2I evaluation and ImgEdit [67] Basic for single-turn image editing. As shown in Table 4,Xiaomi-Robotics-U0preserves strong general T2I and image-editing capabilities after embodied post-training. It performs competitively on most appearance-oriented generation and local editing tasks. Performance drops are largely confined to highly complex scenarios that require multi-constraint 25 Table 3 WorldArena dimension-level scores and overall EWMScore. The best score in each column is bolded. Xiaomi-Robotics-U0 SisyphusWorld BWM-Fast SACWM DexWorldEngine Visual Quality Image Quality53.9445.5751.2245.8351.44 Aesthetic Quality40.7938.3840.1537.3040.96 JEPA Similarity90.6096.5897.8796.2686.71 Motion Quality Dynamic Degree73.7076.2769.5870.3569.16 Flow Score86.0299.8275.1195.7974.31 Motion Smoothness95.5191.4294.3390.6094.27 Content Consistency Subject Consistency79.0582.1181.4282.1781.46 Background Consistency86.4486.5290.1787.9088.85 Photometric Consistency2.135.693.086.504.42 Physics Adherence Interaction Quality87.3071.9879.8872.2881.04 Trajectory Accuracy41.8944.5844.8946.9551.41 3D Accuracy Depth Accuracy85.2594.8586.4194.9986.19 Perspectivity98.8490.3497.1492.3497.58 Controll- ability Instruction Following93.8682.6890.2281.6692.62 Semantic Alignment89.3589.1889.1589.1689.45 EWMScore (↑)73.6473.0672.7172.6772.66 Table 4 Text-to-image and image-editing benchmark comparison with category-level breakdowns. GenEval scores are reported on a 0–1 scale, and ImgEdit scores use a 1–5 judge scale. ModelOverall Single object Two object Counting Colors Position Attribute binding Qwen-Image0.870.990.920.890.880.760.77 Xiaomi-Robotics-U0-AR 0.740.990.920.490.880.600.53 ModelOverall Add Adjust Extract Replace Remove Background Style Hybrid Action Qwen-Image4.27 4.38 4.163.434.664.144.384.81 3.82 4.69 Xiaomi-Robotics-U0-AR 4.08 4.21 4.153.284.354.214.034.72 3.27 4.53 composition, precise spatial grounding, or symbolic binding. Table 4 demonstrates thatXiaomi-Robotics-U0retains robust general T2I and image-editing capabilities after embodied post-training. While it slightly lags behind Qwen-Image in highly complex cases—such as multi-constraint composition or precise spatial alignment—Xiaomi-Robotics-U0maintains highly comparable performance on common image generation and local editing instructions without catastrophic forgetting. 4 Related Work 4.1 Foundation Generative Models and Embodied World Models Recent years have witnessed rapid advances in foundation generative models, evolving from image synthesis to unified multimodal world modeling. Large-scale image generation and multimodal generation models, including Stable Diffusion [4], FLUX [27], GPT-Image-2.0 [42], Show-o [64], Janus [60], Janus-Pro [13], Chameleon [55], BAGEL [17], and EMU3.5 [16], have demonstrated remarkable capabilities in semantic understanding, instruction following, compositional reasoning, and controllable visual generation. More recently, foundation video and world models, such as Genie [7], Genie2 [44], Sora [41], MovieGen [46], Wan [59], and Cosmos [1], further extend generation from static images to long-horizon videos and interactive world simulation, enabling coherent prediction of future observations under a unified autoregressive or diffusion framework. Despite their impressive generalization ability, these models are designed for generic visual content and do not explicitly model the geometric constraints, robot embodiments, calibrated camera systems, or physical interaction dynamics required by embodied manipulation. To bridge this gap, recent embodied world models introduce robot-centric world modeling by adapting foundation models to manipulation trajectories and robot videos. Representative approaches include Dreamer Series [22], World Action Models (WAMs) [28,56,65,70], DreamZero [66], and Qwen-RobotWorld [72], 26 which learn predictive models of future observations for planning, simulation, or policy learning. While these methods significantly advance embodied prediction, they are typically optimized using robot-only datasets whose scale and diversity remain substantially smaller than those used for foundation generative models. Consequently, continual adaptation often sacrifices part of the semantic richness, controllability, and visual diversity inherited from foundation pre-training. In contrast, our work treats embodied generation as a natural extension of foundation image and video generation. Rather than specializing a foundation model solely on robot trajectories, we jointly optimize general-domain generation and embodied synthesis within a unified autoregressive framework, preserving the original generation capability while acquiring robot-centric geometric reasoning and interaction modeling. 4.2 Controllable and Multi-view Embodied Generation Recent advances in controllable image generation and image editing have substantially improved semantic ma- nipulation through instruction-guided synthesis and reference-conditioned editing. Representative approaches, including InstructPix2Pix [6], AnyDoor [12], OmniGen [63], GPT-Image-2.0 [42], and FLUX Kontext [27], enable high-quality editing with strong instruction following and semantic consistency. However, these methods are primarily designed for natural images and generally assume single-view observations, making them difficult to directly apply to embodied manipulation, where multiple calibrated cameras, robot embodiments, and interaction states must remain geometrically consistent. Meanwhile, multi-view generation has attracted increasing attention in image synthesis and 3D content creation. Representative methods, including MVDream [53], SyncDreamer [33], Wonder3D [35], Era3D [29], Zero123++ [52], and related multi-view diffusion models [19,54], explicitly enforce cross-view consistency to facilitate high-quality 3D reconstruction and asset generation. Nevertheless, these methods mainly focus on object-centric generation and do not explicitly consider robot embodiments, calibrated manipulation cameras, or interaction-aware scene synthesis. Our work bridges these two research directions by unifying controllable image editing, multi-view embodied scene generation, and embodied transfer within a single autoregressive framework. Furthermore, we introduce a structured control formulation that disentangles workspace layout, background appearance, foreground irrelevant objects, target objects, and lighting into independent control dimensions, enabling significantly richer embodied scene variations while preserving robot geometry, embodiment constraints, and multi-view consistency. 4.3 Embodied Video Generation and Synthetic Data Engines Video generation and embodied world simulation have recently become increasingly important for scalable robot learning. Foundation video models such as MovieGen [46], Wan [59], Cosmos [1], Genie [7], and Qwen- RobotWorld [72] demonstrate impressive capabilities in long-horizon video synthesis and future observation prediction. Meanwhile, large-scale robot datasets and simulation platforms, including Open X-Embodiment [43], RoboCasa [40], ManiSkill [39], BridgeData V2 [58], and DROID [25], have substantially accelerated the development of data-driven robot learning. More recently, embodied video datasets and synthetic data pipelines, such as RoboMIND [61], RoboCoin [62], GenieSim [69], InternA1 [57], and AgiBotWorld [8], further explore large-scale embodied video generation and robot data augmentation. Despite these advances, existing approaches typically synthesize isolated images, simulation assets, or short robot videos independently, without jointly modeling controllable scene generation, embodied transfer, and long-horizon trajectory rollout. In contrast, our work unifies embodied scene generation, embodied transfer, interleaved subtask-subgoal prediction, and embodied video generation within a single autoregressive model. By jointly learning embodied videos at multiple temporal resolutions, our model captures both long-horizon task progression and fine-grained interaction dynamics, transforming foundation world models into scalable embodied data engines for downstream policy learning and future agentic embodied systems. 27 5 Conclusions In this report, we presentedXiaomi-Robotics-U0, a unified embodied synthesis model built upon a world foundation model through lightweight continual training on general-domain and embodied datasets. Xiaomi-Robotics-U0unifies text-to-image generation, image editing, multi-view embodied scene generation, embodied transfer, and embodied video generation within a single autoregressive framework, successfully transferring the rich visual knowledge of foundation models into embodied intelligence. To the best of our knowledge,Xiaomi-Robotics-U0is the first unified embodied synthesis model supporting both high-quality multi-view embodied generation and sequential embodied video generation, as well as the first to introduce a structured control formulation for scalable embodied data generation. Extensive experiments demonstrate state-of-the-art performance across both single-step and sequential embodied generation tasks, outperforming GPT-Image-2.0 in human evaluation, achieving state-of-the-art results on the World Arena benchmark, and substantially improving the robustness of downstream robot policies through generated data. We hope this work provides a practical step toward scalable agentic data engines for embodied intelligence. Despite the promising results,Xiaomi-Robotics-U0has several limitations. First, embodied transfer still relies on depth estimation as an intermediate representation, which may introduce occasional artifacts and limit fine-grained control over object textures and local appearances. Future work will investigate embodied multi-view image editing directly on raw observations. Second, embodied scene generation and video generation are performed separately, making long-horizon rollout susceptible to accumulated errors; jointly generating scenes and subsequent videos is a promising direction. Finally, the current 32K context window limits long-duration video modeling, and extending the model to longer contexts will enable minute-level embodied interaction generation. 28 Contributions ∗ Core contributors, alphabetically listed by first name ∗ External contributor Project Lead • Xinghang Li Data: • Jun Guo ∗ • Qiwei Li ∗ • Hang Lai ∗ • Long Qian ∗ • Yueze Wang ∗ • Nan Sun ∗ • Hongyu Yan ∗ • Jingen Qu • Jiaxi Song • Jiahang Cao • Hanye Zhao • Heyun Wang • Yunhong Wang • Jiaxi Song • Caoyu Xia • Jack Zhao Training: • Jun Guo ∗ • Qiwei Li ∗ • Long Qian ∗ • Yueze Wang ∗ • Jingen Qu • Jiaxi Song • Nan Sun • Hongyu Yan Deployment & Experiments: • Jiahang Cao ∗ • Jun Guo ∗ • Hang Lai ∗ • Qiwei Li ∗ • Long Qian ∗ • Jiaxi Song ∗ • Hongyu Yan ∗ • Hanye Zhao ∗ • Futeng Liu • Wanli Peng Inference Speedup: • Xi Chen ∗ • Jingen Qu ∗ Infrastructure: • Xinghang Li ∗ • Jingen Qu • Xi Chen Senior Lead: • Jason Li ∗ • Diyun Xiang • Hangjun Ye • Heng Qu • Huaping Liu ∗ Acknowledgment We would like to express our sincere appreciation to the broader team for their tremendous support, including those not listed above: Li Jiang, Zhiyin Lai, Xiaoke Xilinjueluo, Qingyi Li, Qi Liu, Yayun Liu, Jun Xia, Feng Qiu, Donghao Wang, Yan Hou, Dong Wang, Quanyun Zhou, Liangliang He, Jiaxin Liu, Kang Zhou, Rui Cai, Shuoxue Bi, Yingchao Zhou, Kun Ma, Yiwei Zhou 29 GČ®tsėÚȾĜsĭÝėÚėÚÈ ČÝÔÚė®Čó®oľ®ČÈhĜnnij nÝhėÝėĭÝėÚüĜėÝnėÝoÔ澗 UtÈėÚÈnÈhė®Čóėüsݾë ĜsėÚȾĜs澗 LüĭÈČėÚÈnÈhė®Čóėüsݾë ĜsėÚȾ®o®oľ®ČijÝėü ėÚÈnÈhėtÝÄÈühėÚÈė®dnÈ澗 RÈ®¾ÚhüČĭ®ČÄĭÝėÚėÚÈ ČÝÔÚė®ČóėüÔČ®dėÚÈ düėnÈ澗 Pݾë ėÚÈ tóüėÚ üČ®oÔÈ ®oÄ ĭÚÝėÈ Ú®ódĜČÔ澘 ėÚÈ ČÈÄ düIJ ĭÝėÚ ÔünÄÈo tėݾët澘 ®oÄ sn®¾È ėÚÈó üo ėÚÈ tünÝÄ ÔČÈo s®nó瀨 tÝĸÈÄ ėČ®ij ĜtÝoÔ düėÚ ®Čót澗 Sė®Čė ĭÝėÚ ČÈÄ dnü¾ë üo ėÚÈ nÈhė澘 hünüĭÈÄ dij ÔČÈo dnü¾ë oÈIJė ėü ČÈÄ dnü¾ë澘 ®oÄ ÈoÄ ĭÝėÚ dnĜÈ dnü¾ë üo ėÚÈ h®Č ČÝÔÚė澗 T®ëÈ ėÚÈ ¾ünüČhĜn ¾®Čėüo óĜÔ ĭÝėÚ üoÈ ®Čó澘 Čüė®ėÈ Ýė澘 ®oÄ Ú®oÔ Ýė üoėü ėÚÈ Č®¾ë ĭÝėÚ ėĭü ČüÄt tėݾëÝoÔ üĜė澗 Pn®¾È ėÚČÈ ėÚÈ Ä®Čë ÝotÝÄÈ düĭn nÝÔÚė dnĜÈ üĜėtÝÄÈ üo ėüs üh È®¾Ú üėÚÈČ澗 GČ®d ®oÄ sn®¾È ėĭü ėÚÈ tsüČėt tÚüÈ ĭÚÝėÈ ®oÄ dn®¾ë Ýo ėÚÈ üČ®oÔÈ ¾®ČÄdü®ČÄ tÚüÈ瀨düIJ澘 ėÝst nÈhė澗 UtÈėÚÈČÝÔÚė®ČóėüÔČÝs ėÚÈdüĭnČÝó®oÄnÝhėÝė ®düyÈėÚÈė®dnÈtĜČh®¾È澗 Figure 22 Qualitative video generation results ofXiaomi-Robotics-U0. Rows 1–5: WorldArena benchmark samples con- ditioned on language instructions and initial observations. Rows 6–10: videos generated fromXiaomi-Robotics-U0pro- duced scene images. Each row shows the initial frame and language instructions followed by predicted future frames. 30 WüČëts®¾È澙 As®nÈÔČ®ijhünÄÈÄhÈnės®ÄÝtyÝtÝdnÈüo ėÚÈČÝÔÚėtÝÄÈühėÚÈĭüČëts®¾È澘s®ČėÝ®nnijüĜėühČ®óÈ澗 Iė®sÈ®ČtėüdÈó®ÄÈühÄÈotÈĭünó®ėÈČÝ®nĭÝėÚ® hĜĸijó®ėÈėÈIJėĜČÈ澗 WüČëts®¾È澙 AdČÝÔÚėČÈÄsn®tėݾė®dnÈĭÝėÚh®Ýoė óünÄÝoÔnÝoÈt®oÄ®tÈóÝ瀨ÔnütĜČh®¾È澗 LÝÔÚėÝoÔ澙 BČÝÔÚėhnĜüČÈt¾ÈoėtėČÝsthČüó®düyÈ ĭ®tÚÝoÔėÚÈĭüČëts®¾È澗 B®¾ëÔČüĜoÄ澙 AėČ®otnĜ¾ÈoėݾÈĭ®nĭÝėÚhČüĸÈo dĜddnÈt®oÄs®nÈnÝÔÚė澗AĭÚÝėÈhČütėÈÄhnüČ澗 T®ČÔÈė澙 AtÝnyÈČėÚÈČó®nhüÝndn®oëÈėĭÝėÚ®¾ČÝoënÈÄ óÈė®nݾtĜČh®¾ÈÝtsütÝėÝüoÈÄoÈ®ČėÚȾÈoėÈČühėÚÈ ĭüČëts®¾È澗 WüČëts®¾È澙 DÈsdnĜȾüo¾ČÈėÈĭ®nĭÝėÚh®Ýoė yÈČėݾ®ntÈ®ót®oÄ®o®ČüĭtėČÝsühdČĜtÚÈÄtėÈn ėČÝóČĜoÝoÔÚüČÝĸüoė®nnij®nüoÔėÚÈd®tÈ澗 WüČëts®¾È澙 A¾üd®nėdnĜÈČĜdÈČó®ėĭÝėÚ® t®ėĜČ®ėÈÄhÝoÝtÚ®oÄnÝÔÚėnijsÈddnÈÄėÈIJėĜČÈ澗 T®ČÔÈė澙 Adn®¾ëh®dČݾtÚüÈd®ÔĭÝėÚ®ó®ėÈėÈIJėĜČÈ ®oÄĭÚÝėÈsČÝoėÈÄėÈIJėÝtsütÝėÝüoÈÄoÈ®ČėÚÈČÝÔÚė ÈÄÔÈühėÚÈĭüČëts®¾È澗 T®ČÔÈė澙 As®nÈd®ódüsn®¾Èó®ėĭÝėÚÄ®ČëdČüĭo ¾ČütÚ®ė¾ÚÈÄtėČÝsÝoÔ®oÄ®hüČÈtė瀨ÔČÈotėÝė¾ÚÈÄ düČÄÈČÝt¾ÈoėÈČÈÄüoėÚÈĭüČëts®¾È澗 B®¾ëÔČüĜoÄ澙 At®oÄijÄÈtÈČėsn®ÝoĭÝėÚĭÝoÄ瀨 tóüėÚÈÄČÝssnÈt®oÄóĜėÈÄdÈÝÔÈėüoÈtÈIJėÈoÄt ÝoėüėÚÈÄÝtė®o¾ÈdÈÚÝoÄėÚÈĭüČëts®¾È澗 LÝÔÚėÝoÔ澙 DÝódnĜÈ®ódÝÈoėnÝÔÚėÝoÔĭÝėÚh®Ýoė ČÈhnȾėÝüot®oÄ®nóütėoüyÝtÝdnÈtÚ®Äüĭt澗 B®¾ëÔČüĜoÄ澙 Aè®ÔÈÄĭ®nühdn®¾ëyün¾®oݾČü¾ë ĭÝėÚÔnüĭÝoÔtÈ®ót®Ô®ÝotėÚÈh®Čĭ®n澗A¾Č®¾ëÈÄ üdtÝÄÝ®ohnüČdÈoÈ®ėÚėÚÈĭüČëts®¾È澗Asünüh üČ®oÔÈn®y®tÚÝóÈČÝoÔÝoėÚÈÄÝtė®o¾È澗 T®ČÔÈė澙 Rüdüė®ČóÚünÄt®dnĜÈÄÈoÝó¾nüėÚĭÝėÚ ČĜÔÈÄėĭÝnėÈIJėĜČÈ澗 IČÈnÈy®oė澙 OsÈoČĜtėÈľüsÈČdÝoó®ÄÈühüIJÝÄÝĸÈÄ óÈė®nĭÈ®yÈ澗IotÝÄÈėÚÈdÝo®ČÈünÝyÈh®dČݾÝėÈót澗A hünÄÈÄó®ČüoėüĭÈnĭÝėÚ¾ČÈ®óĸÝÔĸ®ÔsČÝoėt®oÄ tėÝė¾ÚÈÄÈÄÔÈt澗 B®¾ëÔČüĜoÄ澙 PünÝtÚÈÄdn®¾ëüdtÝÄÝ®ohnüČÝoÔ澗A ¾Ú®ÝČĜsÚüntėÈČÈÄÝoÄÈsĜČsnÈt®ėÝo澗 T®ČÔÈė澙 AhČütėÈÄÔn®tĜ¾ĜnÈoėt¾ĜnsėĜČÈĭÝėÚ® ėČ®otnĜ¾Èoė¾ij®o澗AdČĜtÚÈľüsÈČėÈ®süėĭÝėÚ® t®ėÝohÝoÝtÚ®oľĜČyÈÄtsüĜė澗 B®¾ëÔČüĜoÄ澙 Atė®ÝoÈÄ瀨Ôn®t¾®ėÚÈÄČ®nĭ®nndÈÚÝoÄ ėÚÈĭüČëts®¾È澗Tĭü¾®ČyÈÄü®ë¾ÚüÝČtÈ®ėtĭÝėÚ ėÈIJėĜČÈÄh®dČݾĜtÚÝüotsütÝėÝüoÈÄdÈÚÝoÄėÚÈ ĭüČëts®¾È澗 WüČëts®¾È澙 PünÝtÚÈÄdn®¾ëÔČ®oÝėÈtn®dĭÝėÚtĜdėnÈ tsȾënÈÄs®ėÈČo®oÄÔnüttijėÈIJėĜČÈ澘tÈČyÝoÔ®tėÚÈ sČÝó®ČijtĜČh®¾ÈhüČó®oÝsĜn®ėÝüo澗 B®¾ëÔČüĜoÄ澙 DÈotÈyÈnyÈėtė®ÔȾĜČė®ÝoÝoČÝ¾Ú ¾ČÝótüo澘Ú®oÔÝoÔÝodČü®ÄyÈČėݾ®nhünÄt®oľ®ė¾ÚÝoÔ ĭ®ČóėÚÈ®ėČݾ®nÚÝÔÚnÝÔÚėt澗 WüČëts®¾È澙 AdÈľüyÈČÈÄĭÝėÚdn®¾ëČĜdÈČ ó®ėÝoÔhÈ®ėĜČÝoÔÝoėÈČnü¾ëÝoÔs®oÈnt®oÄtĜdėnÈ óünÄÈÄÔČüyÈt澗 T®ČÔÈė澙 AėÝÈ瀨ÄijÈhÈtėÝy®ntÚÝČėĭÝėÚtĭÝČnÝoÔoÈüo ¾ünüČt®oÄ®tühėèÈČtÈijėÈIJėĜČÈ澗 IČÈnÈy®oė澙 AsĜČsnÈČĜdÈČüdèȾėĭÝėÚ®ó®ėÈ ėÈIJėĜČÈÝtyÝtÝdnÈoÈ®ČėÚÈČÝÔÚėÈÄÔÈühėÚÈ ĭüČëts®¾È澗 OČÝÔÝo®n S¾ÈoÈMüÄÝÓÝÈÄ DÈt¾ČÝsėÝüoGÈoÈČ®ėÈÄ S¾ÈoÈ Figure 23 Qualitative results of embodied transfer. We extract depths and modify descriptions from the original scene to generate the new scene. 31 RüdüėAČóTijsÈ澙AÔÝBüėG溷澗 WüČëts®¾È澙Adnü¾ëühhnüČ®nhü®ótü®ëÈÄÝoĭ®ėÈČ澘yÝdČ®oėÔČÈo®oÄsüČüĜt澘ĜtÈÄhüȮȮoÔÈóÈoėt澗TÚÈtĜČh®¾ÈÝtÄ®ós澘 ¾ČĜódnij澘®oÄČȾė®oÔĜn®Č澗 OdèȾėt澙AdĜoÄnÈühnüoÔ瀨tėÈóÈÄČÈÄČütÈtĭÝėÚėÚüČottėČÝsÈÄnÝÈtüoėÚÈnÈhė澗AČünühÔČÈohnüČ®nė®sÈtÝėtĜsČÝÔÚėoÈ®Č ėÚȾÈoėÈČ澗As®ÝČühsČĜoÝoÔtÚÈ®ČtĭÝėÚČÈÄÚ®oÄnÈtČÈtėtüoėÚÈČÝÔÚėÈÄÔÈ澗AÚ®oÄhĜnühd®dij灄tdČÈ®ėÚ澘ÄÈnݾ®ėÈ®oÄĭÚÝėÈ澘 t¾®ėÈČtoÈ®ČėÚÈhČüoė澗AtsünühÔünÄČÝdüo澘tÚÝoij®oÄo®Čüĭ澘Ĝotsüntėüĭ®ČÄėÚÈd®¾ë澗Asn®tėݾy®tÈ澘¾nÈ®Č®oľijnÝoÄČݾ®n澘 tė®oÄtÈósėijÝoėÚÈh®Č¾üČoÈČ澗 LÝÔÚėÝo Ô 澙BČÝÔÚė澘¾ünüČhĜntėĜÄÝünÝÔÚėtÝnĜóÝo®ėÈėÚÈhnüĭÈČt澘ÈoÚ®o¾ÝoÔėÚÈÝČyÝyÝÄÚĜÈt澗SühėtÚ®Äüĭth®nndÈÚÝoÄėÚÈtėÈót澘 ¾ČÈ®ėÝoÔÄÈsėÚÝoėÚÈ®Č®oÔÈóÈoė澗 B®¾ëÔČüĜoÄ澙W®nntnÝoÈÄĭÝėÚtÚÈnyÈtühÄČÝÈÄhnüĭÈČt®oÄČÝdüotÝoÈyÈČij¾ünüČtĜČüĜoÄėÚÈ®ČÈ®澗L®ČÔÈĭÝoÄüĭtnüëüĜė üoėü®dĜtėnÝoÔhnüĭÈČó®ČëÈėtėČÈė澗BĜ¾ëÈėtühČÈtÚdnüóthÝnėÚÈhnüČts®¾ÈÝoėÚÈd®¾ëÔČüĜoÄ澗A¾Ú®nëdü®ČÄóÈoĜnÝtėÝoÔ hnüĭÈČsČݾÈtÚ®oÔtüoėÚÈd®¾ëĭ®n澗 PČüósėGÈoÈČ®ėÈÄ MĜnėÝ瀨VÝÈĭ OdtÈČĬ®ėÝüot RüdüėAČóTijsÈ澙AÔÝBüėG溷澗 WüČëts®¾È澙AtȾėÝüoühėÚÈIoėÈČo®ėÝüo®nSs®¾ÈSė®ėÝüo灄tÈIJėÈČÝüČÚ®oÄČ®Ýn澘¾ü®ėÈÄÝoĭÚÝėÈėÚÈČó®nÝotĜn®ėÝüoó®ėÈČÝ®n®oÄ óݾČüóÈėÈüČüÝÄsÝėt澗 OdèȾėt澙AtsȾݮnÝĸÈÄėüČċĜÈĭČÈo¾ÚĭÝėÚ®ėÈėÚÈČ¾nÝsÝtó®ÔoÈėݾ®nnij®ė®¾ÚÈÄoÈ®ČėÚÈèüÝoė¾üoȾėÝüo澗At®ósnȾ®oÝtėÈČ tÈ®nÈÄĭÝėÚ®ČÈľ®shnü®ėtȾĜČÈÄdij®dĜoÔȾüČÄÝoėÚȾÈoėÈČ澗AhČ®ÔóÈoėühÔünÄÈoóĜnėÝn®ijÈČÝotĜn®ėÝüohüÝnhnĜėÈČt nütÈnijoÈ®ČėÚÈČÝÔÚė¾n®ós澗AÄ®ė®tn®ėÈĭÝėÚ®¾Č®¾ëÈÄt¾ČÈoÝtėČ®sÈÄėüėÚÈČ®ÝnoÈ®ČėÚÈnÈhėÈÄÔÈ澗AÔnüyÈÄėünsüĜ¾Ú ó®ÄÈühKÈyn®ČÚ®oÔtyÈČėݾ®nnij®ėÚÈd®¾ë澗 LÝÔÚėÝoÔ澙UohÝnėÈČÈÄtün®Č®ÄÝ®ėÝüosČüyÝÄÈtÝoėÈotÈ澘dnÝoÄÝoÔĭÚÝėÈnÝÔÚėĭÝėÚ®dtünĜėÈdn®¾ëtÚ®ÄüĭtĭÚÈČÈėÚÈtĜoÝt dnü¾ëÈÄ澗 B®¾ëÔČüĜoÄ澙TÚȾĜČy®ėĜČÈühėÚÈE®ČėÚĭÝėÚtĭÝČnÝoÔdnĜÈü¾È®ot®oÄĭÚÝėȾnüĜÄtÄüóÝo®ėÈtėÚÈyÝÈĭdÈnüĭ澗TÚÈÝohÝoÝėÈ dn®¾ëyüÝÄühts®¾ÈÝtÄüėÈÄĭÝėÚdČÝÔÚė澘ĜodnÝoëÝoÔtė®Čt澗Sün®Čs®oÈntĭÝėÚ®dnĜÈėÝoėÈIJėÈoÄÝoėüėÚÈÄÝtė®o¾ÈüoėÚÈnÈhėtÝÄÈ澗 AČüdüėݾ®ČótÈÔóÈoėÝts®ČėÝ®nnijyÝtÝdnȮԮÝotėÚÈd®¾ëÄČüsühėÚÈsn®oÈė澗 RüdüėAČóTijsÈ澙ARXRüdüė澗 WüČëts®¾È澙TÚÈÔn®tėüsüh®ÚÝÔÚ瀨®nėÝėĜÄÈóÈėÈüČünüÔݾ®ntė®ėÝüoė®dnÈ澘hČütėÈÄĭÝėÚݾȾČijtė®nt®oÄyÝdČ®ėÝoÔtnÝÔÚėnijhČüó ėÚÈĭÝoÄ澗 OdèȾėt澙AdČ®ttd®ČüóÈėÈČĭÝėÚ®¾Č®¾ëÈÄh®¾ÈtÝėtȾĜČÈnijdünėÈÄėüėÚÈoÈ®ČnÈhė¾üČoÈČ澗Atė®¾ëühĭÈė澘Ú®oÄĭČÝėÈo nüÔdüëtėÝÈÄĭÝėÚtėČÝoÔČÈtėtÝoėÚȾÈoėÈČ澗AhČüĸÈoóÈČ¾ĜČijėÚÈČóüóÈėÈČnÝÈtÚüČÝĸüoė®nnij®¾ČütėÚÈČÝÔÚėtÝÄÈ澗As®ÝČüh ėÚݾëĭünÈoÔnüyÈtĭÝėÚݾȾnÝoÔÝoÔėüėÚÈhÝdÈČtÝėtdijėÚÈÄÔÈ澗Ató®nėÝoühó®ė¾ÚÈt澘Ú®nh瀨Èósėij®oÄ®ós澘Ýtsn®¾ÈÄoÈ®Č ėÚÈnüÔdüët澗AóÈė®n¾üós®tĭÝėÚ®ÔnüĭÝoÔÄÝ®nČÈtėtüoėüsühėÚÈÔnüyÈt澗 LÝÔÚėÝoÔ澙H®ČtÚ澘s®nÈdnĜÈóÈČÔÈo¾ijLEDtėČÝstóüĜoėÈÄüoėÚȾÈÝnÝoÔ¾®tėtėÈČÝnÈ澘tÚ®ÄüĭnÈttnÝÔÚėÚ®ėČÈhnȾėtÚ®Čsnijüh ėÚÈݾijÔn®t澗 B®¾ëÔČüĜoÄ澙TÚݾëČÈÝohüČ¾ÈÄĭÝoÄüĭtČÈyÈ®n®dnÝoÄÝoÔĭÚÝėÈdnÝĸ®ČÄtĭÝČnÝoÔüĜėtÝÄÈ澗SėÈntĜsüČėdÈ®ót¾ü®ėÈÄÝohČütė ¾ČÝt¾ČütėÚÈČüó澗AČÈÄÈóÈČÔÈo¾ijhÝČÈIJėÝoÔĜÝtÚÈČÝtóüĜoėÈÄüoėÚÈh®Čĭ®n澗TÚÈtüĜoÄühÚüĭnÝoÔĭÝoÄyÝdČ®ėÈtėÚČüĜÔÚ ėÚÈoėÝČÈtėČĜ¾ėĜČÈ澗 RüdüėAČóTijsÈ 澙 ARXRüdüė澗 WüČëts®¾È澙Ao®o¾ÝÈoė澘Ýoë瀨tė®ÝoÈÄĭüÄÈosČÝoėÝoÔsČÈttdÈÄ澘¾®ČyÈÄĭÝėÚÝoėČݾ®ėÈhnüČ®ns®ėÈČot®oÄtóÈnÝoÔüh ėĜČsÈoėÝoÈ®oÄünÄs®sÈČ澗 OdèȾėt澙AÚÈ®yijnÈ®ÄėijsÈėČ®ijhÝnÈÄĭÝėÚČÈyÈČtÈÄnÈėÈČtÝėtüoėÚÈnÈhėtÝÄÈ澗AČünÈČ¾üyÈČÈÄÝoėÚݾëdn®¾ëÝoëČÈtėt ÄÝ®Ôüo®nnij®¾ČütėÚȾÈoėÈČ澗AtÚÈėühÚ®oÄó®ÄÈČݾÈs®sÈČĭÝėÚ®hČÈtÚ¾ČÝótüosČÝoėnÝÈtüoėÚÈČÝÔÚė澗Ató®nnsüČ¾Èn®Ýodüĭn ühĭ®ėÈČhüČ¾nÈ®oÝoÔdČĜtÚÈtÝtoÈ®ČėÚÈhČüoėÈÄÔÈ澗AĭüÄÈoó®nÈėĭÝėÚ®ĭüČoÚ®oÄnÈnÈ®ot®Ô®ÝotėÚÈÝoëČünÈČ澗S¾®ėÈČÈÄ ¾üsÈČtÚ®yÝoÔthČüóėijsȾ®tėÝoÔnÈ®óoÈ®ČėÚÈd®¾ë¾üČoÈČ澗 LÝÔÚėÝoÔ澙Sühė澘ÄÝČȾėÝüo®nÄ®ijnÝÔÚėtėČÈ®ótėÚČüĜÔÚ®ÚÝÔÚ¾nÈČÈtėüČijĭÝoÄüĭ澘ÝnĜóÝo®ėÝoÔÄĜtėóüėÈtÄ®o¾ÝoÔ®düyÈėÚÈÝoë瀨 tė®ÝoÈÄĭüÄ澗 B®¾ëÔČüĜoÄ澙T®nntÚÈnyÈts®¾ëÈÄĭÝėÚėÚüĜt®oÄtühĭüÄÈoėijsȾ®tÈtnÝoÈėÚÈÄÝónijnÝėĭ®nnt澗An®ČÔÈÚ®oÔÝoÔsĜnÈijtijtėÈó hüČnÝhėÝoÔs®sÈČÈ®ótÄüóÝo®ėÈtėÚȾÈÝnÝoÔts®¾È澗TÚÈhnüČÝt¾üo¾ČÈėÈ澘tė®ÝoÈÄĭÝėÚÄȾ®ÄÈtühtsÝnÈÄÝoë澗F®ÝoėtüĜoÄtüh® ¾ÝėijtėČÈėÄČÝhėÝoėÚČüĜÔÚėÚÈüsÈoĭÝoÄüĭ®düyÈ澗 RüdüėAČóTijsÈ澙AÔÝdüėG溸澗 WüČëts®¾È澙TÚȾüoėČün¾üotünÈüh®ČÈėČü瀨hĜėĜČÝtėݾts®¾ÈtÚÝs¾ü¾ësÝėhČüó®溷溿溽溶t¾Ý瀨hÝóüyÝÈtÈė澗 OdèȾėt澙BnÝoëÝoÔ®o®nüÔÄÝ®nt澘ėüÔnÈtĭÝė¾ÚÈtĭÝėÚĭüČon®dÈnt澘®oÄ®èüijtėݾëĭÝėÚ®ÔnüĭÝoÔ¾Čijtė®nėÝs澗 LÝÔÚėÝoÔ澙PĜnt®ėÝoÔoÈüonÝÔÚėtÝosÝoë®oÄėÈ®nÈódÈÄÈÄÝoėÚÈÄ®tÚdü®ČÄ澘¾®tėÝoÔ®¾ijdÈČoÈėݾÔnüĭ澗 B®¾ëÔČüĜoÄ澙An®ČÔÈyÝÈĭsüČėtÚüĭÝoÔ®tė®ČhÝÈnÄtėČÈ®ëÝoÔdij®ėĭ®ČstsÈÄ澘ĭÝėÚoÈdĜn®ÈÝotÚ®ÄÈtühsĜČsnÈ®oÄÔünÄ澗 RüdüėAČóTijsÈ 澙 AÔÝdüėG溸澗 WüČëts®¾È澙AtÚÈėühėÚݾë澘Ú®oÄó®ÄÈČݾÈs®sÈČ澘¾ČÈ®ó瀨¾ünüČÈÄ®oÄėÈIJėĜČÈÄĭÝėÚyÝtÝdnÈhÝdÈČt澘tėČÈė¾ÚÈÄė®ĜėüyÈČ® ĭüÄÈohČ®óÈ澗TÚÈtĜČh®¾ÈÝtÄÈnݾ®ėÈ®oÄó®ėÈ澗 OdèȾėt澙AėČ®ÄÝėÝüo®n¾®nÝÔČ®sÚijdČĜtÚĭÝėÚdn®¾ëÝoëÄČÝsÝoÔhČüóėÚÈėÝsnÝÈtÄÝ®Ôüo®nnij®¾ČütėÚȾÈoėÈČ澗AtünÝÄÝoëtėݾë澘 ¾®ČyÈÄĭÝėÚ®ÄČ®Ôüo澘ČÈtėtüoėÚÈnÈhėtÝÄÈ澗Ató®nnsüČ¾Èn®Ýoĭ®ėÈČÄČüsÈČtÚ®sÈÄnÝëÈ®sÈ®¾ÚtÝėtoÈ®ČėÚÈČÝÔÚėÈÄÔÈ澗AČÈÄ tÈ®ntė®ósĭÝėÚÔünÄ¾Ú®Č®¾ėÈČtÝtsn®¾ÈÄ®ėÚÈhČüoė澗AhünÄÈÄh®oĭÝėÚ®n®oÄt¾®sÈs®ÝoėÝoÔnÈ®ot®Ô®ÝotėÚÈd®¾ëhČ®óÈ澗 LÝÔÚėÝoÔ澙Sühė澘o®ėĜČ®nÝÔÚėhČüó®s®sÈČn®oėÈČoÚ®oÔÝoÔoÈ®Čdij¾ČÈ®ėÈt®ÔÈoėnÈ澘ĭ®ČóÔnüĭ澗SÚ®Äüĭt®ČÈtühė®oÄd®ČÈnij sÈČ¾ÈsėÝdnÈüoėÚÈėÈIJėĜČÈÄs®sÈČ澗 B®¾ëÔČüĜoÄ澙AėČ®ÄÝėÝüo®nJ®s®oÈtÈėÈ®ČüóĭÝėÚtnÝÄÝoÔtÚüèÝt¾ČÈot®oÄė®ė®óÝó®ėhnüČÝoÔ澗At¾Čünns®ÝoėÝoÔüh®ĭ®ėÈČh®n Ú®oÔtüoėÚÈh®Čĭ®n澗AdČüoĸÈÝo¾ÈotÈdĜČoÈČÈóÝėtėÚÝoėČ®ÝntühtóüëÈÝoėÚȾüČoÈČ澗Büot®ÝėČÈtÝotÚ®nüĭsüėtnÝoÈėÚÈ ĭÝoÄüĭtÝnÝoėÚÈd®¾ëÔČüĜoÄ澗 RüdüėAČóTijsÈ澙AÔÝnÈIJPÝsÈČ澗 WüČëts®¾È澙Atn®dühÄ®Čëtn®ėÈtėüoȾüóüonijĜtÈÄhüČtÈČyÝoÔüijtėÈČt澘ĭÈėĭÝėÚtÈ®ĭ®ėÈČ®oľČĜtÚÈÄݾÈ澗TÚÈtĜČh®¾ÈÝt dn®¾ë澘ČÈhnȾėÝyÈ澘®oÄÈIJėČÈóÈnij¾ünÄ澗 OdèȾėt澙SÝIJhČÈtÚüijtėÈČtüoėÚÈÚ®nhtÚÈn®ČÈ®Č®oÔÈÄÝo®tÈóÝ瀨¾ÝČ¾nÈüoėÚÈnÈhė澗AĭÈÄÔÈühnÈóüoĭÝėÚdČÝÔÚėijÈnüĭČÝoÄ tÝėtÝoėÚȾÈoėÈČ澗Ató®nÔn®tÄÝtÚühČÈľü¾ëė®Ýnt®Ĝ¾ÈÝtsn®¾ÈÄüoėÚÈČÝÔÚėÈÄÔÈ澗AtsČÝÔühÔČÈos®ČtnÈijÔ®ČoÝtÚÈtėÚÈ hČüoė¾üČoÈČ澗AdüėnÈüsÈoÈČĭÝėÚ®ĭüÄÈoÚ®oÄnÈČÈtėtüoėÚÈd®¾ëÈÄÔÈ澗 LÝÔÚėÝoÔ澙BČÝÔÚė澘¾ünČÈtė®ĜČ®oėtsüėnÝÔÚėÝoÔhČüó®düyÈÚÝÔÚnÝÔÚėtėÚÈÔnÝtėÈoÝoÔóüÝtėĜČÈüoėÚÈüijtėÈČt®oÄݾÈ澗RÈhnȾėÝüot ühėÚÈÄÝoÝoÔČüó¾Ú®oÄÈnÝÈČtÄ®o¾ÈüoėÚÈĭÈėtn®ėÈ澗 B®¾ëÔČüĜoÄ澙AdĜtėnÝoÔĜst¾®nÈtÈ®hüÄČÈtė®ĜČ®oėĭÝėÚdnĜČÈÄÝoÈČtÝoėÚÈd®¾ëÔČüĜoÄ澗WÚÝėÈė®dnȾnüėÚt®oÄtÝnyÈČĭ®ČÈ ®ČÈyÝtÝdnÈüo®Ä设Èoė®dnÈt澗L®ČÔÈ®ċĜ®ČÝĜóė®oëtĭÝėÚtĭÝóÝoÔhÝtÚnÝoÈėÚÈh®Čĭ®n澗W®ÝėÈČtÝodn®¾ëyÈtėtóüyÈċĜݾënij ėÚČüĜÔÚėÚÈt¾ÈoÈ澗Sühėè®ĸóĜtݾÝtÝósnÝÈÄdijėÚÈnÈÔ®oė®ėóütsÚÈČÈ澗 RüdüėAČóTijsÈ澙AÔÝnÈIJPÝsÈČ澗 WüČëts®¾È澙Aó®tÝyÈdnü¾ëüh¾nÈ®ČHÝó®n®ij®ot®nė澘sÝoë®oľČijtė®nÝoÈ澘¾®ČyÈÄÝoėü®hn®ėsČÈs®Č®ėÝüotĜČh®¾È澗TÚÈėÈIJėĜČÈÝt ÔČ®Ýoij®oÄČij澘nÈ®yÝoÔ®hÝoÈÄĜtėüo®oijėÚÝoÔÝėüĜ¾ÚÈt澗 OdèȾėt澙AhÝnÈėühČ®ĭt®nóüo澘üČ®oÔÈ®oÄÔnÝtėÈoÝoÔ澘nÝÈt¾ÈoėČ®nnijüoėÚÈt®nėdnü¾ë澗AtsČÝÔühČÈtÚÄÝn澘dČÝÔÚėÔČÈo®oÄ hÈ®ėÚÈČij澘ČÈtėtüoėÚÈhÝtÚ澗Ató®nndüĭnühdn®¾ësÈsÈČ¾üČot澘ČüĜÔÚ®oÄ®Čë澘tÝėtüoėÚÈnÈhė¾üČoÈČ澗AnÈóüoĭÈÄÔÈ澘ijÈnüĭ ®oÄèĜݾij澘Ýtsn®¾ÈÄoÈ®ČėÚÈČÝÔÚėÈÄÔÈ澗A¾ÚÈh灄tëoÝhÈĭÝėÚ®D®ó®t¾ĜtėÈndn®ÄÈnÈ®ot®Ô®ÝotėÚÈd®¾ë澗A¾nĜtėÈČüh¾ü®ČtÈ tÈ®t®nė¾Čijtė®nt澘ĭÚÝėÈ®oÄè®ÔÈÄ澘t¾®ėÈČtoÈ®ČėÚÈhČüoė澗 LÝÔÚėÝoÔ澙W®Čó澘ÔünÄÈonÝÔÚėhČüó®Ú®oÔÝoÔEÄÝtüodĜndÈoÚ®o¾ÈtėÚÈsÝoëÚĜÈtühėÚÈt®nė®oÄėÚÈüČ®oÔÈühėÚÈhÝtÚ澗 SÚ®Äüĭt®ČÈtühė®oÄÝoyÝėÝoÔ澘¾ČÈ®ėÝoÔ®¾üĸij¾ĜnÝo®Čij®ėóütsÚÈČÈ澗 B®¾ëÔČüĜoÄ澙RĜtėݾëÝė¾ÚÈoĭ®nntó®ÄÈühÈIJsütÈÄdČݾë®oÄÚ®oÔÝoÔ¾üsÈČsüėthÝnėÚÈd®¾ëÔČüĜoÄ澗WüÄÈotÚÈnyÈt tėü¾ëÈÄĭÝėÚtsݾÈè®Čt®oÄČÝÈÄÚÈČdtnÝoÈėÚÈČÈ®Č澗AtėüoÈhÝČÈsn®¾ÈÔnüĭtĭ®ČónijÝoėÚÈh®Č¾üČoÈČ澗TÚÈhnüČÝtó®ÄÈüh ĭÝÄÈ瀨sn®oëü®ë澘ĭüČotóüėÚdijijÈ®ČtühĜtÈ澗 Figure 24 Embodied scene generation results. 32 Figure 25 Text-to-image results of Xiaomi-Robotics-U0. 33 AÄ ® tó®n ĭüÄÈo dü®ė澘 ĭÝėÚ ® Ä®Čë ČÈÄ ÚĜnnCÚ®oÔÈ ėÚÈ ĭ®ė¾Ú tėČ®s ėü ® Ä®Čë dČüĭo ¾Čü¾üÄÝnÈ瀨ÈódütÈÄ nÈ®ėÚÈČ tėČ®s AÄ ® sÈČtüo ĭ®nëÝoÔ ®nüoÔ ėÚÈ ÄÝČė s®ėÚ澘 h®¾ÝoÔ ėüĭ®ČÄ ėÚÈ ü¾È®o澘 ĭÈ®ČÝoÔ ® d®¾ës®¾ë ®oÄ ¾®tĜ®n ÚÝëÝoÔ ¾nüėÚÈt澗 RÈsn®¾È ėÚÈ dÝČÄ Ýo ėÚÈ Ýó®ÔÈ ĭÝėÚ ® tó®n Č®dÝė澗 AÄ ® h®óÝnij üh ÄĜ¾ët tĭÝóÝoÔ Ýo ėÚÈ ėĜČċĜüÝtÈ ĭ®ėÈČ oÈ®Č ėÚÈ ĭ®ėÈČh®nnt澗 AÄ ĭ®Čó tĜotÈė nÝÔÚėÝoÔ澗 CÚ®oÔÈ ėÚÈ ÚüĜtÈ ėü ® ¾ü®tė®n tÈėÝoÔ ĭÝėÚ ® dÈ®¾Ú ®oÄ ü¾È®o Ýo ėÚÈ d®¾ëÔČüĜoÄ澗 AÄ ® tsČÝÔ üh ¾Ýn®oėČü ėü ėÚÈ s®ÝoėÝoÔ澗 CÚ®oÔÈ ėÚÈ d®¾ëÔČüĜoÄ ėü ó®ėÈ dn®¾ë澗 AÄ óüČoÝoÔ ÄÈĭ ÄČüst澗 EIJėČ®¾ė ėÚÈ ijÈnüĭ oȾëėÝÈ ĭüČo Ýo ėÚÈ Ýó®ÔÈ澗 M®ëÈ Ýė nüë nÝëÈ ®dtėČ®¾ė ®Čė澗 RÈóüyÈ ėÚÈ ¾ünüČhĜn Ýohn®ė®dnÈ düĜo¾È ÚüĜtÈ ĭÝėÚ tnÝÄÈt Ýo ėÚÈ hüČÈÔČüĜoÄ澗 TČ®othÈČ ėÚÈ Ýó®ÔÈ Ýoėü ® Ú®oÄ瀨t¾ĜnsėÈÄ ¾n®ijó®ėÝüotėijnÈ澗 RÈsn®¾È ėÚÈ ÄČ®Ôüo ó®ÄÈ üh sÝsÈ ¾nÈ®oÈČt Ýo ėÚÈ Ýó®ÔÈ ĭÝėÚ ® dČÝÔÚė ijÈnüĭ ČĜdÈČ ÄĜ¾ë澗 RÈóüyÈ ėÚÈ ¾®ósÈČ y®o Ýo ėÚÈ Ýó®ÔÈ ®oÄ ó®ëÈ ėÚÈ tĜČüĜoÄÝoÔ hüČÈtė óüČÈ tÈ®ónÈt澗 M®ëÈ Ýė nüë nÝëÈ ®o ®Č¾ÚÝėȾėĜČ®n tëÈė¾Ú澗 CÚ®oÔÈ ėÚÈ ó®Ýo üdèȾė ¾ünüČ ĭÚÝnÈ sČÈtÈČyÝoÔ ÄÈė®Ýnt澗 AÄ tühė üyÈČÚ騀 h®¾ėüČij nÝÔÚėÝoÔ澗CüoyÈČė ėÚÈ Ýó®ÔÈ tėijnÈ ėü Ýoë澗 RÈsn®¾È ėÚÈ ėÝoij ÚüĜtÈ üo ĭÚÈnt Ýo ėÚÈ Ýó®ÔÈ ĭÝėÚ ® yÝoė®ÔÈ ¾®Č Figure 26 Image editing results of Xiaomi-Robotics-U0. 34 References [1]Niket Agarwal, Arslan Ali, Jon Allen, Martin Antolini, Adeline Aubame, Alisson Azzolini, Junjie Bai, Maciej Bala, Yogesh Balaji, Josh Bapst, et al. Cosmos 3: Omnimodal world models for physical ai.arXivpreprint arXiv:2606.02800, 2026. [2]Timur Akhtyamov, Mohamad Al Mdfaa, Javier Antonio Ramirez Benavides, Arthur Nigmatzyanov, Sergey Bakulin, German Devchich, Denis Fatykhov, Diego Ruiz Salinas, Alexander Mazurov, Kristina Zipa, et al. Egowalk: A multimodal dataset for robot navigation in the wild.arXivpreprintarXiv:2505.21282, 2025. [3]Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-VL technical report.arXivpreprintarXiv:2511.21631, 2025. [4]Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXivpreprintarXiv:2311.15127, 2023. [5]Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. RT-1: Robotics transformer for real-world control at scale.arXivpreprintarXiv:2212.06817, 2022. [6]Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. InProceedingsoftheIEEE/CVFconferenceoncomputervisionandpatternrecognition, pages 18392–18402, 2023. [7] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. InForty-first InternationalConferenceonMachineLearning, 2024. [8] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xuan Hu, Xu Huang, et al. AgiBot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems.arXivpreprintarXiv:2503.06669, 2025. [9]Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. InEuropeanConferenceonComputerVision, pages 370–387. Springer, 2024. [10]Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zilong Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. InProceedingsoftheComputerVisionandPattern RecognitionConference, pages 22831–22840, 2025. [11] Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et al. Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation.arXivpreprintarXiv:2506.18088, 2025. [12]Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. InProceedingsoftheIEEE/CVFconferenceoncomputervisionandpatternrecognition, pages 6593–6602, 2024. [13] Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus-pro: Unified multimodal understanding and generation with data and model scaling.arXivpreprint arXiv:2501.17811, 2025. [14]Yuzhi Chen, Ronghan Chen, Dongjie Huo, Yandan Yang, Dekang Qi, Haoyun Liu, Tong Lin, Shuang Zeng, Junjin Xiao, Xinyuan Chang, et al. Abot-physworld: Interactive world foundation model for robotic manipulation with physics alignment.arXivpreprintarXiv:2603.23376, 2026. [15]Open X-Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, Albert Tung, Alex Bewley, Alex Herzog, Alex Irpan, et al. Open X-Embodiment: Robotic learning datasets and RT-X models. https://arxiv.org/abs/2310.08864, 2023. [16] Yufeng Cui, Honghao Chen, Haoge Deng, Xu Huang, Xinghang Li, Jirong Liu, Yang Liu, Zhuoyan Luo, Jin- sheng Wang, Wenxuan Wang, et al. Emu3. 5: Native multimodal models are world learners.arXivpreprint arXiv:2510.26583, 2025. 35 [17]Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining.arXivpreprintarXiv:2505.14683, 2025. [18]Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets.AdvancesinNeuralInformationProcessingSystems, 36:27092–27112, 2023. [19]Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models.arXiv preprintarXiv:2405.10314, 2024. [20] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment.AdvancesinNeuralInformationProcessingSystems, 36:52132–52152, 2023. [21]Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. InProceedingsoftheIEEE/CVFconferenceoncomputervisionandpatternrecognition, pages 18995–19012, 2022. [22]Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination.arXivpreprintarXiv:1912.01603, 2019. [23]Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al.π 0.5 : A Vision-Language-Action Model with Open-World Generalization.arXivpreprintarXiv:2504.16054, 2025. [24]Tao Jiang, Tianyuan Yuan, Yicheng Liu, Chenhao Lu, Jianning Cui, Xiao Liu, Shuiqi Cheng, Jiyang Gao, Huazhe Xu, and Hang Zhao. Galaxea open-world dataset and g0 dual-system vla model.arXivpreprintarXiv:2509.00576, 2025. [25]Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset.arXivpreprintarXiv:2403.12945, 2024. [26]Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In ProceedingsoftheACMSIGOPS29thSymposiumonOperatingSystemsPrinciples, 2023. [27] Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas Müller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space, 2025. URLhttps: //arxiv.org/abs/2506.15742. [28]Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, et al. Causal world modeling for robot control.arXivpreprintarXiv:2601.21998, 2026. [29]Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention.Advancesin NeuralInformationProcessingSystems, 37:55975–56000, 2024. [30] Xinqing Li, Xin He, Le Zhang, Min Wu, Xiaoli Li, and Yun Liu. A comprehensive survey on world models for embodied ai.arXivpreprintarXiv:2510.16732, 2025. [31]Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. InProceedingsofthe IEEE/CVFConferenceonComputerVisionandPatternRecognition, pages 22160–22169, 2024. [32] Yang Liu, Weixing Chen, Yongjie Bai, Xiaodan Liang, Guanbin Li, Wen Gao, and Liang Lin. Aligning cyber space with physical world: A comprehensive survey on embodied ai.IEEE/ASMETransactionsonMechatronics, 2025. [33]Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. InInternationalconferenceonlearning representations, volume 2024, pages 27676–27697, 2024. 36 [34]Qian Long, Yueze Wang, Jiaxi Song, Junbo Zhang, Peiyan Li, Wenxuan Wang, Yuqi Wang, Haoyang Li, Shaoxuan Xie, Guocai Yao, et al. Scaling world model for hierarchical manipulation policies.arXivpreprintarXiv:2602.10983, 2026. [35]Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In ProceedingsoftheIEEE/CVFconferenceoncomputervisionandpatternrecognition, pages 9970–9980, 2024. [36]Xiaoxiao Long, Qingrui Zhao, Kaiwen Zhang, Zihao Zhang, Dingrui Wang, Yumeng Liu, Zhengjie Shu, Yi Lu, Shouzheng Wang, Xinzhe Wei, et al. A survey: Learning embodied intelligence from physical simulators and world models.arXivpreprintarXiv:2507.00917, 2025. [37] Leland McInnes, John Healy, and S. Astels. hdbscan: Hierarchical density based clustering.J.OpenSource Softw., 2:205, 2017. URL https://api.semanticscholar.org/CorpusID:53231359. [38]Byeon Minwoo, Park Beomhee, Kim Haecheon, Lee Sungjun, Woonhyuk Baek, and Kim Saehoon. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022. [39]Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xuanlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations.arXivpreprint arXiv:2107.14483, 2021. [40]Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots.arXivpreprint arXiv:2406.02523, 2024. [41]OpenAI.Video generation models as world simulators.https://openai.com/research/ video-generation-models-as-world-simulators, 2024. Technical report. [42]OpenAI. Gpt image api.https://platform.openai.com/docs/guides/image-generation, 2025. Accessed: 2026-07-03. [43]Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In2024IEEEInternationalConferenceonRoboticsand Automation(ICRA), pages 6892–6903. IEEE, 2024. [44]Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Christos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei Zhang, Frederic Besse, Tim Harley, Anna Mitenkova, Jane Wang, Jeff Clune, Demis Hassabis, Raia Hadsell, Adrian Bolton, Satinder Singh, and Tim Rocktäschel. Genie 2: A large-scale foundation world model. 2024. URL https://deepmind.google/discover/blog/genie-2-a-large-scale-foundation-world-model/. [45]Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. InInternational ConferenceonLearningRepresentations, volume 2024, pages 1862–1874, 2024. [46]Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models.arXivpreprint arXiv:2410.13720, 2024. [47] Xuanchi Ren, Yifan Lu, Tianshi Cao, Ruiyuan Gao, Shengyu Huang, Amirmojtaba Sabour, Tianchang Shen, Tobias Pfaff, Jay Zhangjie Wu, Runjian Chen, et al. Cosmos-drive-dreams: Scalable synthetic driving data generation with world foundation models.arXivpreprintarXiv:2506.09042, 2025. [48]Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models.Advancesinneuralinformationprocessingsystems, 35:25278–25294, 2022. [49]Yu Shang, Zhuohang Li, Yiding Ma, Weikang Su, Xin Jin, Ziyou Wang, Lei Jin, Xin Zhang, Yinzhou Tang, Haisheng Su, et al. Worldarena: A unified benchmark for evaluating perception and functional utility of embodied world models.arXivpreprintarXiv:2602.08971, 2026. 37 [50]Yu Shang, Xin Zhang, Yinzhou Tang, Lei Jin, Chen Gao, Wei Wu, and Yong Li. Roboscape: Physics-informed embodied world model.AdvancesinNeuralInformationProcessingSystems, 38:63674–63698, 2026. [51]Fengyuan Shi, Zhuoyan Luo, Yixiao Ge, Yujiu Yang, Ying Shan, and Limin Wang. Scalable image tokenization with index backpropagation quantization. InProceedingsoftheIEEE/CVFInternationalConferenceonComputer Vision, pages 16037–16046, 2025. [52] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model.arXivpreprint arXiv:2310.15110, 2023. [53]Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. InInternationalconferenceonlearningrepresentations, volume 2024, pages 39838–39859, 2024. [54]Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. InEuropeanConferenceonComputerVision, pages 175–191. Springer, 2024. [55] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models, 2024.URLhttps://arxiv. org/abs/2405.09818, 9(8), 2024. [56]MotuBrain Team, Chendong Xiang, Fan Bao, Haitian Liu, Hengkai Tan, Hongzhe Bi, James Li, Jiabao Liu, Jingrui Pang, Kiro Jing, et al. Motubrain: An advanced world action model for robot control.arXivpreprint arXiv:2604.27792, 2026. [57]Yang Tian, Yuyin Yang, Yiman Xie, Zetao Cai, Xu Shi, Ning Gao, Hangxu Liu, Xuekun Jiang, Zherui Qiu, Feng Yuan, et al. Interndata-a1: Pioneering high-fidelity synthetic data for pre-training generalist policy. In ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition, pages 976–985, 2026. [58]Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, An- dre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. InConferenceonRobotLearning, pages 1723–1736. PMLR, 2023. [59]Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, et al. Wan: Open and advanced large-scale video generative models.arXivpreprintarXiv:2503.20314, 2025. [60]Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. InProceedingsoftheComputerVisionandPatternRecognitionConference, pages 12966–12977, 2025. [61] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation.arXivpreprintarXiv:2412.13877, 2024. [62] Shihan Wu, Xuecheng Liu, Shaoxuan Xie, Pengwei Wang, Xinghang Li, Bowen Yang, Zhe Li, Kai Zhu, Hongyu Wu, Yiheng Liu, et al. Robocoin: An open-sourced bimanual robotic data collection for integrated manipulation. arXivpreprintarXiv:2511.17441, 2025. [63]Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. InProceedingsoftheIEEE/CVFConference onComputerVisionandPatternRecognition, pages 13294–13304, 2025. [64]Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. InInternationalConferenceonLearningRepresentations, volume 2025, pages 28240–28264, 2025. [65]Angen Ye, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Hao Li, Hengtao Li, Jie Li, Jindi Lv, Jingyu Liu, et al. Gigaworld-policy: An efficient action-centered world–action model.arXivpreprintarXiv:2603.17240, 2026. [66]Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, et al. World action models are zero-shot policies.arXivpreprint arXiv:2602.15922, 2026. 38 [67]Yang Ye, Xianyi He, Zongjian Li, Shenghai Yuan, Zhiyuan Yan, Bohan Hou, Li Yuan, et al. Imgedit: A unified image editing dataset and benchmark.AdvancesinNeuralInformationProcessingSystems, 38, 2026. [68]Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. InProceedingsoftheIEEE/CVFInternationalConferenceonComputerVision, pages 12–22, 2023. [69] Chenghao Yin, Da Huang, Di Yang, Jichao Wang, Nanshu Zhao, Chen Xu, Wenjun Sun, Linjie Hou, Zhijun Li, Junhui Wu, Zhaobo Liu, Zhen Xiao, Sheng Zhang, Lei Bao, Rui Feng, Zhenquan Pang, Jiayu Li, Qian Wang, and Maoqing Yao. Genie sim 3.0 : A high-fidelity comprehensive simulation platform for humanoid robot, 2026. URL https://arxiv.org/abs/2601.02078. [70] Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-wam: Do world action models need test-time future imagination?arXivpreprintarXiv:2603.16666, 2026. [71]Yuguang Yue, Irakli Salia, Samuel Hunt, Chris Green, Wenzhe Shi, and Jonathan J Hunt. Scaling behavior cloning improves causal reasoning: An open model for real-time video game playing.arXivpreprintarXiv:2601.04575, 2026. [72] Jie Zhang, Xiaoyue Chen, Anzhe Chen, Chenxu Lv, Deqing Li, Gengze Zhou, Hang Yin, Haoqi Yuan, Haoyang Li, Jiahao Li, et al. Qwen-robotworld technical report: Unifying embodied world modeling through language- conditioned video generation.arXivpreprintarXiv:2606.17030, 2026. [73]Junkang Zhou, Yefei He, Feng Chen, Weijie Wang, and Bohan Zhuang. Flashar: Efficient post-training acceleration for autoregressive image generation, 2026. URL https://arxiv.org/abs/2605.09430. 39