摘要
近期的基础图像与视频生成模型展现出强大的泛化能力和可控性，但其在具身场景中的直接应用受到多视角一致性、几何连贯性以及机器人本体约束等要求的限制。现有方法通常利用有限的机器人数据对基础模型进行适配，往往牺牲了在大规模预训练过程中获得的视觉知识。
我们提出小米机器人-U0（Xiaomi-Robotics-U0），一个拥有 380 亿参数的多模态自回归模型，用于统一的具身合成。它将具身生成视为基础图像与视频生成的延伸，并联合优化了文生图、图像编辑、具身场景生成、具身迁移以及具身视频生成。这种统一框架在将预训练的世界基础模型适配到具身场景的同时，保留了其泛化能力。
小米机器人-U0 是首个支持跨多种机器人本体进行高质量多视角场景生成的模型，并引入了结构化、可控的具身迁移，用于实现细粒度编辑，同时保持多视角一致性和交互动态。它在单步生成和序列生成任务上取得了最先进的结果，在具身场景生成与迁移的人工评估中超越了 GPT-Image-2.0，在具身视频生成的 World Arena 排行榜上排名第一，并将具有挑战性的真实世界操作任务中的分布外成功率从 36.9% 提升至 63.2%。这些结果表明，基础世界模型既可以作为具身世界模型，也可以作为具身智能的可扩展数据引擎。代码和检查点可在 https://robotics.xiaomi.com/xiaomi-robotics-u0.html 获取。
完整作者列表请参见贡献章节。请将通信发送至
mi-robotics@xiaomi.com
1 引言
近期，基础图像与视频生成模型 [podell2024sdxl, labs2025flux1kontextflowmatching, openai_gptimage_2025, blattmann2023stable, cui2025emu3, xie2025show, wu2025janus, chen2025janus, team2024chameleon] 通过互联网数据训练，在语义理解、可控生成和视觉推理方面取得了显著进展。大规模生成模型现已能够根据多种多模态输入合成高度逼真的图像和视频，展现出远超训练数据分布的惊人泛化能力。这些能力使基础生成模型成为具身智能 [yuan2026fast, ye2026world, shang2026roboscape, long2026scaling] 的理想起点——在具身智能中，机器人需要在行动前对复杂环境进行推理，并想象未来的交互过程。
然而，具身生成 [long2025survey, liu2025aligning, li2025comprehensive] 带来的挑战与传统的图像和视频合成有着本质区别。与自然图像生成不同，具身场景要求严格的多视角一致性、跨摄像头的精确几何与物理连贯性、明确的机器人本体约束，以及时间上一致的交互动态。生成的观测结果必须与机器人运动学、相机标定以及下游操作策略保持兼容，而不仅仅是视觉上看起来逼真。因此，直接将现有的基础图像或视频生成模型应用于具身场景，往往会导致几何不一致、机器人状态不合理，以及与机器人控制的兼容性差等问题。
近期涌现的具身世界模型 [agarwal2026cosmos, zhang2026qwen, shang2026roboscape] 试图通过利用机器人轨迹 [Neill2023open_x_embodiment, tian2026interndata, yin2026geniesim30, jiang2025galaxea, walke2023bridgedata, brohan2022rt] 或第一人称操作视频 [akhtyamov2025egowalk, grauman2022ego4d] 持续适配预训练基础模型，来弥合这一差距。尽管这些方法已展现出令人鼓舞的进展，但其适配范式与基础模型预训练存在根本性差异。后训练过程通常仅在机器人专用数据集上进行，而非保留那些最初赋予基础模型强大语义理解与可控生成能力的多样化图像及视频生成任务。由于具身数据集与互联网规模的视觉语料库 [gadre2023datacomp, schuhmann2022laion, kakaobrain2022coyo-700m] 相比规模显著更小、多样性更低且高度重复，这种适配不可避免地削弱了从基础模型继承而来的泛化能力，并限制了具身生成的多样性、可控性与可扩展性。
在本工作中，我们提出 Xiaomi-Robotics-U0，一种统一具身合成模型，该模型将基础生成与具身生成统一在单一训练范式内。如图 1 所示，Xiaomi-Robotics-U0 并非将基础模型特化为机器人专用生成器，而是将具身合成重新定义为基础图像与视频生成的自然延伸。从世界基础模型出发，Xiaomi-Robotics-U0 在统一自回归目标下，持续在通用领域数据集与具身数据集上联合训练，共同学习文本到图像生成、图像编辑、具身场景生成、具身迁移以及具身视频生成。图 2 展示了 Xiaomi-Robotics-U0 的上述能力。这种统一公式在保留基础模型丰富语义知识与可控生成能力的同时，有效地将其适配到以机器人为中心的多视角推理与交互建模中。
超越单步生成，小米机器人 U0 进一步将具身合成扩展到序列世界建模。通过联合学习跨多个时间分辨率的交错具身序列与操作视频，该模型既能捕捉长程任务进展，也能捕捉细粒度的交互动态。更重要的是，这使得具身生成能够从静态场景合成演变为可扩展的轨迹生成引擎：生成的具身场景可直接展开为时间上连贯的操作视频，为下游策略学习提供丰富的合成轨迹。我们相信，这一能力代表着向智能体具身数据生成迈出的重要一步——基础世界模型通过自主交互展开，持续生成日益多样化的机器人经验。
大量实验表明，小米机器人 U0 在单步与序列具身生成任务上均确立了新的最优水平。它是首个能够在单一自回归框架内实现高质量多视角具身场景生成、结构化可控具身迁移以及零样本具身视频生成的统一具身合成模型。在具身场景生成与具身迁移的人工评估中，小米机器人 U0 超越了 GPT-Image-2.0；在 World Arena 基准测试上达到最优性能；并显著提升了使用生成数据的下游机器人策略的分布外鲁棒性。这些结果表明，世界基础模型不仅可以作为具身世界模型，还可以作为具身智能的可扩展数据引擎。
-
我们推出小米机器人-U0，这是首个统一的具身合成模型，能够在单一自回归框架内无缝整合基础图像生成、图像编辑、具身场景生成、具身迁移以及具身视频生成。通过在统一持续训练范式下联合优化通用领域与具身生成任务，小米机器人-U0 有效将基础世界模型的泛化能力迁移至具身生成领域，同时保留了其原有的视觉生成能力。
-
我们提出了首个面向多视角具身生成的统一框架，涵盖具身场景生成与可控具身迁移。通过与大规模文生图及图像编辑数据的联合训练，小米机器人-U0 实现了最先进的性能，具备卓越的多视角一致性、细粒度可控性与指令遵循能力，在人工评估中超越了 GPT-Image-2.0。此外，我们引入了一种结构化控制公式，将工作空间、背景、前景无关物体、目标物体与光照解耦为独立控制维度，从而在保持几何一致性与交互动态的前提下，实现可扩展的具身视频增强。生成的数据显著提升了下游机器人策略在具有挑战性的真实世界操作任务中的鲁棒性。
-
我们将具身生成从静态场景合成扩展至序列化具身世界建模，通过交错子任务-子目标学习与多帧率具身视频生成实现。小米机器人-U0 在 World Arena 基准测试中取得了最先进的性能，并展现出强大的零样本多视角具身视频生成能力。通过从合成具身场景出发实现连贯的长程推演，小米机器人-U0 将基础世界模型转化为可扩展的具身数据引擎，为未来策略学习提供了生成多样化具身交互轨迹的智能体化路径。
2 小米机器人-U0
2.1 问题形式化
我们的目标是开发一个统一的具身世界模型，在单一自回归框架下支持多样化的生成与预测任务。给定一个多模态上下文
| (1) |
其中每个元素可以是文本 token、图像 token 或机器人相关的控制 token，该模型学习后续多模态序列的联合分布
| (2) |
使用标准的下一 token 预测（NTP）。
为支持具身智能，小米机器人 U0 在多种具有不同时间结构的互补数据格式上进行训练。具体而言，我们将所有训练样本分为两类。
单步样本描述输入与输出之间的一次性映射，不涉及显式的时间演化。这些样本包括：(i) 文生图（T2I），模型根据文本描述生成图像；(ii) 任意到图像（X2I），使用一到三张参考图像结合文本进行图像编辑与生成；(iii) 具身场景生成，模型根据机器人形态和场景描述生成初始的多视角机器人观测；(iv) 具身迁移，模型根据当前具身观测和目标场景描述预测迁移后的多视角观测。
序列样本建模长程交互与未来演化。我们考虑两种代表性格式：(i) 图文交错的子任务-子目标序列，其中每个子任务描述后跟随完成后的对应多视角观测；(ii) 以多种帧率采集的具身操作视频，使模型能够同时捕捉稀疏的任务级转换和密集的交互动态。
尽管语义不同，所有任务均被表示为统一的多模态 token 序列，并在相同的自回归目标下进行优化，从而使知识能够在图像生成、场景理解、具身推理和未来预测之间自然共享。
2.2 模型架构
图 3 展示了小米机器人 U0 的整体架构。我们采用开源的 EMU3.5 [cui2025emu3] 作为初始化模型，该模型基于 Qwen-3-32B 仅解码器 Transformer 架构构建，在图像生成和图文交错建模方面展现出强大能力。图像首先由 IBQ 分词器 [shi2025scalable] 进行分词处理，其空间压缩比为 。通过将 IBQ 码本扩展至原始 Qwen 词表，所有模态均在一个统一的离散词表中表示，并采用下一 token 预测的方式进行建模。
与引入任务特定预测头的传统具身世界模型相比，小米机器人 U0 将所有生成任务——包括图像生成、具身场景合成、场景迁移、顺序子任务预测以及视频生成——均建模为单一的自回归序列建模问题。这种统一框架能够实现高效的多任务联合训练，同时保留从大规模预训练模型中继承的强视觉生成能力。
为提升推理效率，我们进一步采用了受 FlashAR [zhou2026flasharefficientposttrainingacceleration] 启发的推理加速适配方案。具体而言，引入了一个额外的垂直预测头，以反对角线顺序解码图像 token，从而允许同时生成多个视觉 token。结合优化的 KV 缓存管理以及 vLLM [kwon2023efficient] 提供的连续批处理能力，小米机器人 U0 在 1024×1024 分辨率下的图像生成速度相比原始 NTP 模式最高提升 ，同时保持了大部分生成质量。
2.3 数据集
2.3.1 数据集整理
数据收集
为使小米机器人-U0具备广泛的视觉理解与生成能力，我们整合了一个涵盖六个不同领域的多样化训练语料库，每个领域都贡献了互补的视觉和语义先验知识。
-
通用图文数据提供了基础的视觉-语义对齐。我们纳入了ShareGPT4V [chen2024sharegpt4v] 及类似的大规模视觉语言基准数据集，为基于文本条件的生成建立强基线。
-
具身操作数据弥合了被动视觉理解与主动物理交互之间的鸿沟。该领域结合了真实世界和模拟两种来源。在真实数据方面，我们利用了AgiBotWorld-Beta [bu2025agibot]、Open X-Embodiment [Neill2023open_x_embodiment] 以及内部采集的专有多视角机器人操作数据集（我们称之为MiBot）。在模拟数据方面，我们整合了来自RoboTwin2.0 [chen2025robotwin]、GenieSim [yin2026geniesim30] 和InternData-A1 [tian2026interndata] 的数据，从而覆盖了物理采集困难或危险的场景。
-
自动驾驶数据捕捉了具有复杂空间布局和动态物体的结构化室外场景。我们通过Cosmos-Drive-Dreams [ren2025cosmosdrive] 获取这些数据，该数据集提供了包含多种天气、光照和交通条件的高质量驾驶场景。
-
源自 EgoWalk [akhtyamov2025egowalk] 的自我中心数据提供了第一人称视角的录制内容，这些内容反映了人类自然的视角和运动模式，丰富了模型对人类中心视觉体验的理解。
-
3D 重建数据提供了密集的多视角对应关系和精确的几何结构。我们使用了 ScanNet++ v2 [yeshwanth2023scannet++] 和 DL3DV-10K [ling2024dl3dv]，两者都提供了高质量的室内外 3D 场景捕捉数据。
-
源自 Open-P2P [yue2026openp2p] 的游戏数据引入了风格化和程序化生成的环境，将视觉分布的多样性扩展到了真实世界捕捉之外。
数据过滤
大规模收集的原始数据不可避免地包含会损害生成质量的伪影。受 Abot Physical World [chen2026abot] 启发，我们实现了一个自主数据过滤流程，包括运动模糊检测、图像质量检测和夹爪状态对齐。
总而言之，我们收集了 950 万个样本用于单步生成，对应 564 亿个模型 token。对于序列生成，我们选取了 260 万个视频片段，对应 496 亿个模型 token。
2.3.2 数据集标注
统一标注流程
大多数原始样本缺乏细粒度的文本标注，仅带有粗略的任务级描述，这无法支持精确的文本驱动具身生成。因此，我们构建了一个统一的标注流程，以标准化异构数据上的监督信号。
统一标注流程
我们利用 Qwen3-VL-235B [bai2025qwen3] 这一强大的视觉语言模型（VLM），为通用帧、多源视频序列和机器人操作轨迹生成密集、结构化的标注。该流程由四个具有不同标注目标的解耦分支组成：
-
通用场景标注 所有涵盖通用视觉语料库和具身机器人序列的视频数据均无领域限制地进行处理。每个视频片段都经过时间下采样以选取高清晰度的关键帧。VLM 为每个采样帧生成整体描述，说明整体场景布局、物体交互以及物体间的空间关系。
-
结构化具身场景标注 标注仅针对具身操作数据执行。轨迹处理输出的分解子任务引导场景划分为五个正交语义维度：工作空间、任务对象、无关对象、光照和背景。根据任务相关性为每个维度生成独立描述，支持推理过程中的组合式推理和细粒度可控生成。
-
几何模态提取 通过 Video Depth Anything [chen2025videodepthanything] 为所有帧和视频提取时序一致的反向深度图，为跨域具身迁移任务提供几何条件信号。在该深度表示中，较亮像素对应较近表面，较暗像素对应较远区域，使模型能够在几何先验条件下合成逼真的 RGB 输出。
-
轨迹级子任务分解 如图 5 所示，HDBSCAN 聚类 [McInnes2017hdbscanHD] 利用末端执行器姿态和夹爪信号对机器人轨迹进行分割，以识别运动和抓取过渡边界。在这些边界处采样具有密集物理交互的关键帧，并发送至 VLM 生成带帧边界的时序对齐子任务计划。对具有丰富物理交互的过渡帧进行场景标注处理，为具身场景生成和具身迁移任务生成训练样本。
面向任务的数据组织
为支持小米机器人 U0 的三个训练目标，我们将所有标注数据重新组织为特定任务格式：
-
具身迁移。每个训练样本包含一条系统指令、一个场景描述和多视角深度图作为输入，以多视角 RGB 图像作为监督目标。该任务训练模型从几何线索中幻觉出逼真的外观。
-
场景生成。每个样本将一条系统指令与一个场景描述配对，以多视角 RGB 图像作为监督。该任务学习无需几何先验的文本到多视角场景合成。
-
视频生成。每个样本包含一条系统指令、一条任务指令以及一帧初始观测画面，并由 15–25 帧 RGB 图像序列进行监督。对于提供机器人动作标签的数据集，在初始观测帧之后还会额外附加一组根据机器人动作和 URDF 渲染出的机器人掩码图像，作为动作引导。该任务学习根据任务语义和初始上下文预测未来的视觉状态。
数据平衡
我们的训练语料包含异构的多领域数据，这容易导致领域偏差以及对主导数据分布的过拟合。我们采用基于数据领域和生成任务类型（多视角生成、场景编辑、单视角生成）的样本级重加权，核心目标是保持对机器人操作场景的全面覆盖。
具体来说，我们首先剔除无效的具身片段，包括视觉损坏、时间戳错位、缺少本体感知或动作信号、执行记录失败以及缺乏机器人-物体交互的静态轨迹。所有剩余样本均通过子任务和结构化场景标注产生的层级化标注进行分类。分组标准涵盖任务目标、场景上下文、目标物体、空间关系、子任务阶段以及原子化操作技能。
这种细粒度的分组方式比简单的数据集层级加权能够实现更精确的平衡：共享相同具身-任务-物体组合的冗余样本会被降采样，而包含新颖视觉场景、物体布局、操作技能或时间阶段的片段则会被优先处理。更高的采样权重被分配给长尾类别，包括稀有物体、不常用的操作技能以及核心交互阶段（抓取、释放、放置和物体状态转换）。由此产生的训练分布消除了冗余的重复演示，同时保留了多样化的视觉场景和全面的操作技能覆盖。
2.4 训练
为了在保留预训练图像生成模型原有视觉生成能力的同时，将其高效适配为统一的具身世界模型，我们采用了两种持续训练方案：单步训练和序列训练。
单步训练
该方案侧重于在保持原始图像生成能力的同时，使模型与多样化的单步生成任务对齐。我们联合训练了四个互补的数据集，包括文本到图像（T2I）、任意到图像（X2I）、具身场景生成和具身迁移。
T2I 和 X2I 数据集持续强化从预训练模型继承而来的通用视觉生成和图像编辑能力，防止在具身适配过程中出现灾难性遗忘。同时，具身场景生成教会模型根据机器人具身形态和场景描述，合成物理上合理的初始机器人观测结果；而具身迁移则实现了跨场景的多视角生成，能够保留机器人交互状态和相机几何结构。联合训练这些任务，可以在统一的回归目标下，有效地将视觉生成先验与具身场景理解对齐。
序列训练
序列训练将模型扩展到长程具身推理和交互建模。
我们首先引入了图文交错的子任务-子目标序列，其中每个由语言描述的子任务都与执行后对应的多视角观测结果配对。这种数据格式教会模型在预测未来具身观测结果的同时，对任务分解进行推理，从而进一步对齐从基础模型继承而来的序列建模能力。
为了增强动态交互建模，我们还额外纳入了从 InternData-A1 [tian2026interndata]、GenieSim [yin2026geniesim30]、AgiBotWorld-Beta [bu2025agibot]、RoboCOIN [wu2025robocoin] 和 RoboMIND [wu2024robomind] 收集的大规模具身操作视频。与传统的低帧率轨迹数据集不同，我们构建了帧率（FPS）分别为 1、3 和 5 的多帧率视频序列。稀疏序列有助于长期规划与未来预测，而更密集的序列则使模型能够捕捉细粒度的操作动态和连续的交互过程。在多个时间分辨率上进行训练，显著提升了模型在统一的自回归框架内，同时对高层任务进展和低层具身动态进行建模的能力。
2.5 推理
为了进一步加速推理，我们为小米机器人-U0 引入了一个 FlashAR+ 扩展。该扩展保持了对参考图像及其他输入信号的条件建模不变，同时仅对最终的目标图像区域应用并行自回归生成。此设计旨在保留小米机器人-U0 的多源条件建模能力，并降低 X2I 推理过程中的串行解码成本。
2.5.1 FlashAR+ 扩展
FlashAR 基线
FlashAR [zhou2026flasharefficientposttrainingacceleration] 提出了一种面向自回归文生图的后训练与推理加速范式。其核心思想是在保留原始模型生成能力的同时，引入额外的垂直预测，使得图像 token 能够沿反对角线分组并行生成。这减少了图像区域中串行解码步骤的数量。受 FlashAR 启发，我们在小米机器人-U0 之上进行了 FlashAR+ 后训练，将其从单图 T2I 生成扩展到更广泛的 X2I 场景，并将得到的模型用于小米机器人-U0 中的 X2I 推理加速。图 6(a) 展示了 FlashAR+ 扩展的解码步骤。
在小米机器人-U0 中，X2I 涵盖了 T2I、I2I 以及多参考图像生成等任务。输入条件可能包括文本指令以及一张或多张参考图像。
FlashAR+ 掩码与注意力
与标准自回归生成相比，小米机器人-U0 FlashAR+ 扩展引入了表 1 总结的关键变化。
| 方面 | 标准自回归 | 小米机器人-U0 FlashAR+ 扩展 | 效果 |
| 目标图像生成 | 按光栅顺序逐 token 串行生成 | 在目标图像中按反对角线步骤并行生成 | 减少串行解码步骤 |
| 目标图像注意力 | 标准一维因果注意力 | 条件前缀可见；目标图像内部使用二维步骤因果注意力 | 防止信息泄露 |
| 预测头 | 单个自回归预测头 | 水平头、垂直头与门控融合 | 支持反对角线生成 |
| 方向辅助损失 | 标准监督 | 引入 H/V 损失约束 | 约束两个方向头 |
在训练过程中，FlashAR+ 扩展在数据解析阶段识别图像块并定义目标图像。所有前置内容，包括文本指令、参考图像、控制条件以及目标图像头部，均被视为条件前缀。并行自回归损失仅针对目标图像区域计算，而参考图像仅作为条件上下文。这防止了参考图像被错误地作为生成目标进行监督，并确保模型学会从多源条件生成目标图像。
注意力约束也仅限于目标图像区域，如图 6(b) 所示。条件前缀保持标准的自回归可见性。目标图像 token 可以关注整个条件前缀，但在目标图像内部，它们只能关注来自更早反斜线步骤的 token。对于目标网格位置，其并行步骤定义为；同一步骤内的 token 彼此不可见，这防止了并行生成过程中的信息泄露。参考图像、文本 token 和控制 token 不参与目标图像内部的反斜线掩码构建，从而保留了 X2I 条件的语义完整性。
此约束可表示为目标图像内的一个可见性函数。对于目标图像的查询 token 和键 token，
FlashAR+ 损失
小米机器人 U0 FlashAR+ 扩展引入了对方向性辅助监督的关键调整。我们将水平辅助损失权重与垂直辅助损失权重进行匹配：
此处， 是经过 H/V 门控融合后的目标图像预测损失， 和 分别监督水平头和垂直头， 是 AR 蒸馏项。对于目标位置 ，令 和 表示来自水平头和垂直头的 logits，令 表示门控值。融合后的 logits 为
对应的融合损失为
方向性辅助损失函数使用相同的目标 token 分别监督 和 。H 头也需要适应目标区域的分布。如果 ，H 头可能无法充分收敛，H/V 门控在后续训练阶段可能向 H 分支坍缩。当门控坍缩到训练不足的 H 头时，融合预测退化为不稳定的单方向预测，继续训练会导致生成质量下降。显式添加 可以监督 H 头并稳定 H/V 融合。
小米机器人-U0 FlashAR+ 推理
在推理过程中，小米机器人-U0 首先将文本指令和参考图像编码为条件前缀，后接目标图像头部。然后，模型仅在目标图像网格上进行反对角线生成：位置按照 的递增顺序逐步生成，同一 步内的所有位置并行采样。边界位置使用可用的单方向预测，而内部位置使用 H/V 门控融合。
该过程保留了多源条件施加的约束，同时将图像区域并行解码扩展到小米机器人-U0 FlashAR 扩展推理。因此，模型在保持任意到图像条件生成能力的同时，降低了目标图像区域的串行生成成本。
2.5.2 vLLM 集成
为了进一步提升小米机器人-U0 的图像生成效率，我们在 FlashAR+ 基础上集成了 vLLM。该实现保持 FlashAR+ 的解码规则不变：vLLM 处理条件前缀、批量执行和分页 KV 缓存管理，而对角线调度器发出每个反对角线步骤的视觉 token，并同步用于无分类器引导的成对条件请求和无条件请求。
如图7所示，加速路径在具身任务上保持了高生成质量。在单块H20 GPU上进行指定分辨率的文生图（T2I）生成时，标准自回归（AR）模型每张图像耗时450.77秒。FlashAR+将该延迟降低至每张图像16.56秒，相比AR实现了加速；而FlashAR+结合vLLM，在max_num_seq=28的设置下，进一步将延迟降低至每张图像5.44秒，相比FlashAR+实现了额外的加速。
3 实验
我们的模型统一了具身迁移、具身场景生成和具身视频生成，同时保留了通用的图像编辑与生成能力。本节展示了我们在具身迁移和场景生成任务上的基准测试结果与定性结果，这些结果在人工评估中优于GPT-Image-2.0。具体而言，我们的具身迁移方法在保持几何一致性的同时，显著增强了机器人轨迹的多样性，将成功率从36.9%提升至63.2%。我们进一步通过在Geneval和ImageEdit基准上的评估，验证了模型通用的图像生成能力，确认我们的方法保持了领先的通用图像生成性能。在具身视频生成方面，我们的模型在World Arena基准上取得了最先进的结果。该模型还展现出对自合成初始场景的强大泛化能力，从而构建了一个闭环数据引擎，能够基于任意场景描述和机器人指令生成无限视频。
3.1 具身迁移
基准与协议
为了评估具身迁移能力，我们构建了一个包含 300 个样本的基准测试集，每个样本由一段文本场景描述和一组多视角深度图组成；模型需要根据描述，合成出多视角一致的、关于所描述机器人操作场景的 RGB 图像。此外，每个样本还提供了来自原始数据集的多视角 RGB 参考图像。这些参考图像与（编辑后的）场景描述并不完全对应，但它们捕捉了深度信息之外的结构信息，并用于计算基准测试指标。由于深度图来自与训练集相同的领域，我们特意编辑了部分场景描述，使其要求生成训练数据中不存在的内容，从而测试模型的通用生成能力，而非仅仅是重建能力。根据编辑后的描述与原始场景的偏离程度，我们将基准测试集分为 150 个简单样本（描述与底层场景接近）和 150 个困难样本（描述需要大量新颖内容）。我们与强大的图像生成模型 GPT-Image-2 进行了对比。
指标
我们从三个互补维度评估生成质量，所有指标均基于参考视图计算。在深度一致性方面，我们使用共享估计器从生成图像和参考图像中提取单目深度，并报告 SI-RMSE（尺度不变 RMSE）、（相对深度误差低于 的像素比例）和 AbsRel（绝对相对误差）；这些指标衡量生成图像在多大程度上忠实于输入深度所施加的 3D 几何结构。在结构保真度方面，我们计算 Canny 边缘 F1、精确率和召回率（），以捕捉生成布局是否保留了场景的精细轮廓。在语义对齐方面，我们运行开放词汇检测与分割，并报告分割 mIoU 和召回率（），反映所描述的物体是否被正确合成并定位。如表 2 所示，Xiaomi-Robotics-U0 在 Easy 和 Hard 两个子集上的所有指标均大幅优于 GPT-Image-2，表明其对输入几何结构的遵循程度显著更高，与场景描述的语义对齐也更紧密。
| 指标 | GPT-Image-2 | Xiaomi-Robotics-U0（我们的方法） | ||
| Easy | Hard | Easy | Hard | |
| 深度 SI-RMSE () | 0.4007 | 0.3444 | 0.1407 (0.260) | 0.1435 (0.201) |
| 深度 () | 0.5307 | 0.5340 | 0.8017 (0.271) | 0.7846 (0.251) |
| 深度 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 精确率 () | 0.0560 | 0.0536 | 0.2037 (0.148) | 0.1770 (0.123) |
| Canny 召回率 () | 0.0572 | 0.0582 | 0.2290 (0.172) | 0.2176 (0.159) |
| 分割 mIoU () | 0.4105 | 0.4098 | 0.7621 (0.352) | 0.7436 (0.334) |
| 分割召回率 () | 0.6208 | 0.6524 | 0.8136 (0.193) | 0.8032 (0.151) |
定性结果
图 8 展示了具身迁移的定性示例。对于每个示例，我们展示了场景描述、来自源数据集的原始多视角 RGB 图像，以及由小米机器人-U0 生成的图像。我们的流程首先从原始图像估计多视角深度图，然后编辑场景描述——编辑后的部分以高亮显示——以要求生成与原始场景不同的内容；提取的深度图和编辑后的描述被共同输入模型进行生成。在多种场景下，无论场景描述是自然语言形式还是结构化形式，小米机器人-U0 都能忠实遵循编辑后的文本描述和多视角深度条件，正确渲染所描述的物体、光照和背景，同时尊重每个视角的空间几何结构。关键在于，生成的各个视角在物体身份、外观和布局上保持相互一致，这表明我们的统一模型能够生成连贯的多视角场景，而非各自合理但相互冲突的图像。
我们进一步在图9中将Xiaomi-Robotics-U0与GPT-Image-2进行了对比，使用了源数据集中的三个样本，每个样本展示了场景描述、输入的多视图深度图以及两个模型的输出。GPT-Image-2生成的图像在视觉上具有吸引力，且文本与图像的对齐效果很强：描述中的物体、材质、光照和背景都呈现得较为合理，单独看每个视角都令人信服。然而，它对输入几何结构和跨视角关系的理解较差。其输出经常违反深度条件——物体被放置在不一致的距离上，表面发生扭曲，整体空间布局偏离了给定的几何结构——这反映在其较高的SI-RMSE值和较低的表格2指标上。此外，为同一场景生成的各个视角在很大程度上是相互独立的：物体身份、姿态和排列无法在不同视角间保持一致，因此三张图像并不对应一个连贯的单一3D场景。相比之下，Xiaomi-Robotics-U0忠实遵循了每个视角的深度信息，同时保持了多视图输出之间的相互一致性，这表明它不仅具备GPT-Image-2那样强大的文本跟随能力，而且真正将生成过程植根于输入的几何结构之中。
零样本数据增强
超越基准评估，我们利用小米机器人-U0的结构化文本到图像能力，来增强为四个真实世界任务（整理盒子、折叠毛巾、整理手机、整理耳机）收集的示范轨迹，旨在赋予下游策略更好的泛化能力以应对新场景。对于每条轨迹，我们提示一个视觉语言模型（Qwen3-VL），基于当前的深度图和原始场景描述，为除任务对象外的每个维度生成1000个多样化的替代方案，从而在背景、工作空间和光照方面鼓励最大程度的多样性。然后，我们随机组合这些按维度划分的描述，并将其输入小米机器人-U0进行生成。如图10和图11所示，生成的场景展现出丰富的多样性，并且这些组合均未出现在训练集中，这表明小米机器人-U0能够以零样本方式显著扩展所收集轨迹的视觉覆盖范围。
3.2 具身场景生成
基准与协议
为了评估具身场景生成能力，我们利用大语言模型生成的场景提示词构建了一个专用基准测试。该基准测试包含 400 个样本，平均分为两个难度级别：200 个简单样本用于常规室内操作场景，200 个困难样本用于无约束的开放域场景。每个提示词指定了一种机械臂类型和结构化的场景描述，涵盖工作空间布局、物体配置、光照和背景。我们将小米机器人-U0 与 GPT-Image-2 作为基线进行比较。为确保公平比较，我们向 GPT-Image-2 提供了与每种目标机械臂类型对应的多视角参考图像作为视觉上下文。
基于人工的成对评估
量化跨视角几何一致性仍是一个未解决的挑战，因为目前没有自动化指标能够可靠地评估跨视角的全局三维连贯性。因此，我们采用人工成对偏好协议进行定量评估。标注者会看到小米机器人-U0 和 GPT-Image-2 的成对输出，这些输出以随机打乱的左右顺序呈现，以减轻位置偏好偏差。标注者根据两个标准选择更优结果：（1）多视角几何一致性，衡量具有不变物体位置、尺度和跨视角遮挡的统一三维布局；（2）文本指令遵循度，衡量与提示词中工作空间、物体、光照和背景的对齐程度。
图 14 报告了总体胜率。小米机器人-U0 在大多数比较中优于 GPT-Image-2，在多视角一致性方面取得了明显优势，同时保持了相当的指令遵循性能。图 13 的定性比较显示，GPT-Image-2 生成了高质量图像，但存在严重的跨视角几何不一致问题，不同视角下的物体位置、尺度和布局相互冲突。相比之下，小米机器人-U0 生成了几何一致的多视角观测结果，空间布局连贯，同时保持了照片级真实感。这验证了小米机器人-U0 能够同时实现高视觉保真度和稳健的跨视角一致性。
定性结果
图 12 展示了涵盖四种机械臂类型的定性结果。小米机器人 U0 模型能够针对多样化的场景提示词，生成具有照片级真实感且几何一致的多视角初始场景，展现出对任意文本描述的强泛化能力。这表明，我们的模型通过学习机械臂构型与相机投影之间的内在几何关系，将传统的单视角文本到图像合成扩展到了多视角机器人观测生成。值得注意的是，在评估的四种机械臂类型中，AGIBOT G2 仅使用模拟数据进行训练，但小米机器人 U0 模型仍能针对该特定机械臂类型生成与真实世界对齐的照片级真实感观测。这验证了该模型通过习得的真实世界视觉先验，在从模拟到现实的泛化方面具有良好潜力。
3.3 真实世界实验
我们开展真实世界评估，以探究所提出的世界模型是否能够改善下游策略学习。具体而言，我们使用小米机器人 U0 模型来增强专家演示数据，并在一系列真实世界任务中，对比基于原始数据和增强数据训练得到的 VLA 策略。
任务套件
该基准测试包含三个真实世界的桌面操作任务：收纳耳机、叠毛巾和装箱，涵盖了灵巧操作、可变形物体处理以及多物体交互等多种挑战。每个任务由一条语言指令和一组有序的进度里程碑定义。图 15 可视化了每个任务的里程碑状态，为后文定义的进度指标提供了具体参考。
策略训练与数据
对于上述套件中的每个任务，我们均从 pi05_base 检查点出发，使用其官方 PyTorch 实现，独立地对一个 [intelligence2025pi_0_5] 策略进行后训练。所有对比策略的训练设置保持一致；唯一的变量是监督微调的数据混合方式。对于每个任务，我们在下文描述的机器人平台上收集约 40 小时的真实世界演示数据，构成特定于任务的干净数据。为了构建增强数据，我们将具身场景生成流程（第 3.2 节）应用于干净演示数据，为每个任务生成约 40 小时的风格迁移片段，这些片段在视觉外观（背景、光照和纹理）上有所变化，同时保留原始的机器人状态和动作标签。我们比较了由此产生的两个每任务策略：原始策略仅在特定于任务的干净数据上进行后训练；小米机器人 U0 增强策略则在干净数据及其对应的风格迁移增强数据的混合数据上进行后训练。在机器人执行过程中未使用任何推理时的世界模型模块；世界模型仅通过后训练数据混合方式影响策略。
评估计划
为了在保持对比受控的前提下，评估超出原始演示分布之外的鲁棒性，我们在两组真实世界条件下评估每种策略：基础组和干扰组。每组包含每个任务的三种物体布局变体，两种策略均在相同的条件安排下进行评估。在基础组中，三种布局仅在物体放置位置上有所不同；背景使用演示数据中见过的桌布，场景在标准实验室光照条件下进行评估。在干扰组中，布局同样改变物体放置位置，但背景替换为演示数据中未出现过的、留出的桌布——通常具有更复杂的图案以及不同的纹理或表面反射率。干扰组还引入了留出的光照条件，包括低光、彩色光以及动态图案光照（如迪斯科球效果）。每个任务和组的完整评估布局集如图 16 所示。对于每种策略、每个任务、每个测试组和每种布局，我们运行三次独立的真实机器人试验。这样每个策略-任务对共产生 18 次试验（2 组 × 3 种布局 × 3 次试验）。
机器人设置
我们将所有训练好的策略部署在一个双臂 ARX 机器人平台上，该平台配备两条 6 自由度机械臂，每条机械臂均装有一个 1 自由度平行爪夹持器。每次策略查询时，观测信息包括当前本体感知的机械臂状态以及三个 RGB 视角：一个来自头部摄像头的头部视角，以及两个分别来自左腕和右腕摄像头的腕部视角。对于“收纳耳机”和“折叠毛巾”这两个任务，这些视角通过 Intel RealSense D405 摄像头采集；对于“装箱”任务，我们使用相同视角配置的鱼眼摄像头，以覆盖更大的装箱工作空间。鱼眼图像在使用前会进行去畸变处理，既作为策略输入，也作为 Xiaomi-Robotics-U0 数据增强的输入。在实际机器人评估过程中，策略运行在独立的推理服务器上，并通过局域网内的 WebSocket 连接与机器人控制进程通信。机器人以 1 Hz 的频率将当前机械臂状态和三个 RGB 视角传输给服务器，服务器返回一个动作片段，该片段被执行直至下一次查询。所有被比较的策略在观测流、通信栈和动作片段执行接口上均保持一致。
评估指标
我们仅使用任务完成进度作为评估指标。与仅评判完全成功不同，进度指标根据一次运行所完成的有序里程碑给予部分得分，这使得它更适合长时域操作任务——在这种任务中，策略可能可靠地完成早期子目标，却在后续阶段失败。具体来说，如果一个任务包含 个有序里程碑，且一次运行完成了前 个里程碑，则其进度为 ，其中 表示完全成功。对于任务 和组 ，被测试策略的报告进度定义为：
其中 是每个任务和组的运行次数（在我们的设置中为三种布局 × 三次试验）。Xiaomi-Robotics-U0-Aug 的进度得分更高，表明该世界模型提供了有用的训练后变化，从而在完全和部分完成的执行中均提升了实际环境中的鲁棒性。
实验结果
图 17 报告了两种策略的任务完成进度。在基础组中，测试条件与演示分布高度相似，Xiaomi-Robotics-U0-Aug 的表现与 Original 相当，仅在部分任务上略有下降——这是将训练容量与增强数据共享的预期结果，因为这些数据并未为分布内场景提供新信息。然而，在干扰组中，Xiaomi-Robotics-U0-Aug 在所有任务上均持续优于 Original，这表明风格迁移增强使模型对干净演示中未见的新背景和光照产生了视觉不变性。总体而言，分布偏移下的鲁棒性提升超过了分布内性能的轻微折损，这对于场景条件无法严格控制的现实世界部署尤为宝贵。
为定性说明这一差距，图 18 展示了 Xiaomi-Robotics-U0-Aug 在干扰条件下针对所有三项任务的代表性 rollout。尽管存在具有挑战性的扰动——未见过的桌布、彩色灯光和投影图案——Xiaomi-Robotics-U0-Aug 策略仍能执行预期的操作序列：在“收纳耳机”任务中同时放入两只耳机并合上充电盒，在“折叠毛巾”任务中完成完整的折叠序列，以及在“装箱”任务中在合上盖子前将所有四个物体装入箱中。
图 19 进一步展示了我们评估中出现的典型失败模式：(1) 在饱和光照下未能抓取耳机；(2) 错误识别耳机盒的方向；(3) 在视觉相似的桌布上未能定位毛巾边缘；(4) 因未能完全抓住布边而展开已折叠的毛巾；(5)(6) 试图抓取投影光斑而非目标物体。在严重受损的视觉观测下，两种策略偶尔都会表现出这些意外行为；然而，Xiaomi-Robotics-U0-Aug 更有可能在扰动下正确行动，或在短暂偏离后自我纠正。相比之下，Original 策略则更倾向于永久性地陷入视觉伪影，或以更高频率犯下不可恢复的错误。
3.4 视频生成
WorldArena
具身世界模型的一个关键能力是作为数据引擎，能够从单张观测图像和对应的指令生成物理上合理且视觉质量高的视频。为了评估这种用于具身任务的视频生成能力，我们在 WorldArena [shang2026worldarena] 上对小米机器人-U0 进行了基准测试。给定初始帧以及语言指令和机器人动作，WorldArena 会从六个维度（视觉质量、运动质量、内容一致性、物理遵循度、3D 准确性和可控性）的 15 个指标对生成的视频进行评估。我们将此任务定义为处理交错的视觉-语言输入，并生成带有空语言 token 的多帧交错输出。初始帧与语言指令配对作为输入，同时输入一系列从机器人动作、相机参数和 URDF 渲染出的机器人掩码图像作为动作引导。定量结果报告于表 3 和图 21，定性可视化结果如图 22（第 1-5 行）所示。小米机器人-U0 取得了最佳的整体 EWMScore（73.64），在超过 100 个提交的模型中排名第一。在可控性方面，小米机器人-U0 在指令遵循（93.86）上得分最高，并在语义对齐上排名靠前，展示了其将语言指令转化为忠实视觉结果的强大能力。在物理遵循度方面，小米机器人-U0 在交互质量（87.30）上领先，表明它能捕捉到物理上合理的物体交互。小米机器人-U0 还在透视性（98.84）和运动平滑度（95.51）上取得了最佳成绩，反映了准确的 3D 视角几何和时序稳定的运动。总体而言，小米机器人-U0 在所有评估维度上取得了最佳平衡，证实了其作为具身任务数据引擎的有效性。
具身视频生成
在模拟的 WorldArena 基准测试之外，我们进一步利用小米机器人 U0 的场景生成能力来生成新颖的初始帧，并随后根据语言指令生成视频。代表性示例如图 20 所示。在第一行中，机械臂正确与杯子交互——抓取它、提起它，并将其稳定放置在桌面上，展示了精确的接触建模。在第二行中，除了正确的桌面操作外，模型还通过模拟一个人走向桌子观察机器人操作来渲染动态背景变化，表明其意识到了超出当前任务的场景级动态。在第三行中，模型忠实地模拟了可变形物体（一张纸）的物理特性：纸张在被拿起和放下时真实地弯曲，并且在横向移动时由于质量较轻而表现出轻微滞后，反映了对材料相关物理属性的理解。更多可视化结果见图 22。
| 小米机器人 U0 | SisyphusWorld | BWM-Fast | SACWM | DexWorldEngine | ||
| 视觉质量 | 图像质量 | 53.94 | 45.57 | 51.22 | 45.83 | 51.44 |
| 美学质量 | 40.79 | 38.38 | 40.15 | 37.30 | 40.96 | |
| JEPA 相似度 | 90.60 | 96.58 | 97.87 | 96.26 | 86.71 | |
| 运动质量 | 动态程度 | 73.70 | 76.27 | 69.58 | 70.35 | 69.16 |
| 光流得分 | 86.02 | 99.82 | 75.11 | 95.79 | 74.31 | |
| 运动平滑度 | 95.51 | 91.42 | 94.33 | 90.60 | 94.27 | |
| 内容一致性 | 主体一致性 | 79.05 | 82.11 | 81.42 | 82.17 | 81.46 |
| 背景一致性 | 86.44 | 86.52 | 90.17 | 87.90 | 88.85 | |
| 光度一致性 | 2.13 | 5.69 | 3.08 | 6.50 | 4.42 | |
| 物理遵循度 | 交互质量 | 87.30 | 71.98 | 79.88 | 72.28 | 81.04 |
| 轨迹精度 | 41.89 | 44.58 | 44.89 | 46.95 | 51.41 | |
| 3D 精度 | 深度精度 | 85.25 | 94.85 | 86.41 | 94.99 | 86.19 |
| 透视性 | 98.84 | 90.34 | 97.14 | 92.34 | 97.58 | |
| 可控性 | 指令遵循 | 93.86 | 82.68 | 90.22 | 81.66 | 92.62 |
| 语义对齐 | 89.35 | 89.18 | 89.15 | 89.16 | 89.45 | |
| EWMScore () | 73.64 | 73.06 | 72.71 | 72.67 | 72.66 | |
3.5 文生图与图像编辑基准测试
我们在通用文生图和图像编辑基准上评估了自回归模型 Xiaomi-Robotics-U0。本次评估的目的并非宣称 Xiaomi-Robotics-U0 是一款专用的开放域图像生成器，而是量化该模型在向具身场景合成适配后，保留了多少通用图像生成与编辑能力。我们使用 GenEval [ghosh2023geneval] 进行组合式文生图评估，并使用 ImgEdit [ye2026imgedit] Basic 进行单轮图像编辑评估。
| 模型 | 总体 | 单物体 | 双物体 | 计数 | 颜色 | 位置 | 属性绑定 |
| Qwen-Image | 0.87 | 0.99 | 0.92 | 0.89 | 0.88 | 0.76 | 0.77 |
| Xiaomi-Robotics-U0-AR | 0.74 | 0.99 | 0.92 | 0.49 | 0.88 | 0.60 | 0.53 |
| 模型 | 总体 | 添加 | 调整 | 提取 | 替换 | 移除 | 背景 | 风格 | 混合 | 动作 |
| Qwen-Image | 4.27 | 4.38 | 4.16 | 3.43 | 4.66 | 4.14 | 4.38 | 4.81 | 3.82 | 4.69 |
| Xiaomi-Robotics-U0-AR | 4.08 | 4.21 | 4.15 | 3.28 | 4.35 | 4.21 | 4.03 | 4.72 | 3.27 | 4.53 |
如表 4 所示，Xiaomi-Robotics-U0 在经历具身后训练后，仍保留了强大的通用文生图和图像编辑能力。它在大多数面向外观的生成任务和局部编辑任务上表现具有竞争力。性能下降主要局限于需要多约束组合、精确空间定位或符号绑定的高度复杂场景。
表 4 表明，Xiaomi-Robotics-U0 在具身后训练后仍保留了稳健的通用文生图和图像编辑能力。虽然在高度复杂的案例中（例如多约束组合或精确空间对齐）略逊于 Qwen-Image，但 Xiaomi-Robotics-U0 在常见图像生成和局部编辑指令上保持了高度可比的表现，未出现灾难性遗忘。
4 相关工作
4.1 基础生成模型与具身世界模型
近年来，基础生成模型取得了飞速发展，从图像合成演进到统一的多模态世界建模。大规模图像生成与多模态生成模型，包括 Stable Diffusion [blattmann2023stable]、FLUX [labs2025flux1kontextflowmatching]、GPT-Image-2.0 [openai_gptimage_2025]、Show-o [xie2025show]、Janus [wu2025janus]、Janus-Pro [chen2025janus]、Chameleon [team2024chameleon]、BAGEL [deng2025emerging] 以及 EMU3.5 [cui2025emu3]，已在语义理解、指令遵循、组合推理和可控视觉生成方面展现出卓越能力。更近期，基础视频与世界模型，例如 Genie [bruce2024genie]、Genie2 [parkerholder2024genie2]、Sora [openai2024sora]、MovieGen [polyak2024movie]、Wan [wan2025wan] 和 Cosmos [agarwal2026cosmos]，进一步将生成能力从静态图像拓展至长时视频与交互式世界模拟，能够在统一的自回归或扩散框架下实现对未来观测的连贯预测。尽管这些模型具备令人印象深刻的泛化能力，但它们是为通用视觉内容而设计的，并未显式地对具身操作所需的几何约束、机器人本体、标定相机系统或物理交互动力学进行建模。
为弥合这一差距，近期具身世界模型通过将基础模型适配至操作轨迹与机器人视频，引入了以机器人为中心的世界建模方法。代表性方法包括 Dreamer 系列 [hafner2019dream]、世界动作模型（WAMs）[yuan2026fast, ye2026gigaworld, li2026causal, team2026motubrain]、DreamZero [ye2026world] 以及 Qwen-RobotWorld [zhang2026qwen]，这些方法通过学习未来观测的预测模型，用于规划、仿真或策略学习。尽管这些方法显著推进了具身预测，但它们通常仅使用机器人数据集进行优化，而此类数据集的规模与多样性仍远小于基础生成模型所使用的数据。因此，持续适配往往牺牲了从基础预训练中继承的部分语义丰富性、可控性与视觉多样性。相比之下，我们的工作将具身生成视为基础图像与视频生成的自然延伸。我们并非将基础模型专门用于机器人轨迹，而是在统一的自回归框架内联合优化通用域生成与具身合成，在保留原有生成能力的同时，获取以机器人为中心的几何推理与交互建模能力。
4.2 可控多视角具身生成
近期可控图像生成与图像编辑领域的进展，通过指令引导合成与参考条件编辑，显著提升了语义操控能力。代表性方法包括 InstructPix2Pix [brooks2023instructpix2pix]、AnyDoor [chen2024anydoor]、OmniGen [xiao2025omnigen]、GPT-Image-2.0 [openai_gptimage_2025] 以及 FLUX Kontext [labs2025flux1kontextflowmatching]，这些方法在保持强指令遵循与语义一致性的同时，实现了高质量编辑。然而，这些方法主要针对自然图像设计，且通常假设为单视角观测，因此难以直接应用于具身操作场景——在该场景中，多个标定相机、机器人本体及交互状态必须保持几何一致性。
与此同时，多视角生成在图像合成与 3D 内容创作领域日益受到关注。代表性方法包括 MVDream [shi2024mvdream]、SyncDreamer [liu2024syncdreamer]、Wonder3D [long2024wonder3d]、Era3D [li2024era3d]、Zero123++ [shi2023zero123++] 以及相关的多视角扩散模型 [gao2024cat3d, tang2024mvdiffusion++]，这些方法通过显式强制跨视角一致性来促进高质量 3D 重建与资产生成。然而，这些方法主要聚焦于以物体为中心的生成，并未明确考虑机器人本体、标定过的操作相机或交互感知的场景合成。
我们的工作通过将可控图像编辑、多视角具身场景生成以及具身迁移统一到一个自回归框架中，从而桥接了这两个研究方向。此外，我们引入了一种结构化控制公式，将工作空间布局、背景外观、前景无关物体、目标物体以及光照解耦为独立的控制维度，从而在保持机器人几何结构、本体约束和多视角一致性的同时，实现显著更丰富的具身场景变化。
4.3 具身视频生成与合成数据引擎
视频生成与具身世界模拟近年来在可扩展机器人学习领域日益重要。诸如 MovieGen [polyak2024movie]、Wan [wan2025wan]、Cosmos [agarwal2026cosmos]、Genie [bruce2024genie] 以及 Qwen-RobotWorld [zhang2026qwen] 等基础视频模型，在长程视频合成与未来观测预测方面展现了令人印象深刻的能力。与此同时，包括 Open X-Embodiment [o2024open]、RoboCasa [nasiriany2024robocasa]、ManiSkill [mu2021maniskill]、BridgeData V2 [walke2023bridgedata] 和 DROID [khazatsky2024droid] 在内的大规模机器人数据集与仿真平台，极大地加速了数据驱动的机器人学习发展。最近，诸如 RoboMIND [wu2024robomind]、RoboCoin [wu2025robocoin]、GenieSim [yin2026geniesim30]、InternA1 [tian2026interndata] 和 AgiBotWorld [bu2025agibot] 等具身视频数据集与合成数据管线，进一步探索了大规模具身视频生成与机器人数据增强。
尽管取得了这些进展，现有方法通常独立地合成孤立的图像、仿真资产或短时机器人视频，未能联合建模可控场景生成、具身迁移以及长程轨迹展开。相比之下，我们的工作将具身场景生成、具身迁移、交错子任务-子目标预测以及具身视频生成统一在单个自回归模型之中。通过在多个时间分辨率上联合学习具身视频，我们的模型既能捕捉长程任务进展，也能捕捉细粒度的交互动态，从而将基础世界模型转化为可扩展的具身数据引擎，用于下游策略学习与未来的智能体具身系统。
5 结论
在本报告中，我们介绍了小米机器人-U0，这是一个统一的具身合成模型，基于世界基础模型，通过对通用领域和具身数据集进行轻量级持续训练构建而成。小米机器人-U0 将文本到图像生成、图像编辑、多视角具身场景生成、具身迁移以及具身视频生成统一在单个自回归框架内，成功将基础模型的丰富视觉知识迁移至具身智能。据我们所知，小米机器人-U0 是首个同时支持高质量多视角具身生成和序列化具身视频生成的统一具身合成模型，也是首个引入结构化控制公式以实现可扩展具身数据生成的模型。大量实验表明，该模型在单步和序列化具身生成任务中均达到了最先进的性能，在人工评估中优于 GPT-Image-2.0，在 World Arena 基准测试上取得了最先进的结果，并通过生成的数据显著提升了下游机器人策略的鲁棒性。我们希望这项工作能为具身智能的可扩展智能体数据引擎提供一条切实可行的路径。
尽管结果令人鼓舞，但小米机器人-U0 仍存在若干局限性。首先，具身迁移仍然依赖深度估计作为中间表示，这可能会引入偶发的伪影，并限制对物体纹理和局部外观的精细控制。未来的工作将探索直接在原始观测上进行具身多视角图像编辑。其次，具身场景生成和视频生成是分开进行的，这使得长程推演容易受到累积误差的影响；联合生成场景及后续视频是一个有前景的方向。最后，当前 32K 的上下文窗口限制了长时长视频的建模能力，将模型扩展到更长的上下文将能够实现分钟级的具身交互生成。
贡献者
∗核心贡献者，按名字字母顺序排列 ∗∗外部贡献者
-
李星航
-
郭军∗
-
李奇伟∗
-
赖航∗
-
钱龙∗
-
王悦泽∗
-
孙楠∗
-
闫鸿宇∗
-
曲金根
-
贾松
-
曹佳航
-
赵涵野
-
王鹤云
-
王蕴红
-
贾松
-
夏草雨
-
赵杰克
训练：
-
郭军*
-
李奇伟*
-
钱龙*
-
王悦泽*
-
曲金根
-
贾松
-
孙楠
-
闫宏宇
部署与实验：
-
曹佳航*
-
郭军*
-
赖航*
-
李奇伟*
-
钱龙*
-
贾松*
-
闫宏宇*
-
赵涵野*
-
刘富腾
-
彭万里
推理加速：
-
陈曦*
-
曲金根*
基础设施：
-
李星航*
-
曲金根
-
陈曦
高级负责人：
-
李杰森*
-
向迪云
-
叶杭军
-
曲恒
-
刘华平**
致谢
我们衷心感谢更广泛团队的大力支持，包括未在上方列出的成员：李江、赖志颖、西克新觉罗、李清怡、刘琦、刘雅云、夏军、邱峰、王东浩、侯燕、王东、周全云、何亮亮、刘佳欣、周康、蔡锐、毕硕雪、周颖超、马坤、周逸伟
参考文献
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 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 https://robotics.xiaomi.com/xiaomi-robotics-u0.html.
See Contributions section for full author list. Please send correspondence to
mi-robotics@xiaomi.com
1 Introduction
Recent foundation image and video generation models [podell2024sdxl, labs2025flux1kontextflowmatching, openai_gptimage_2025, blattmann2023stable, cui2025emu3, xie2025show, wu2025janus, chen2025janus, team2024chameleon] 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 [yuan2026fast, ye2026world, shang2026roboscape, long2026scaling], where robots are required to reason about complex environments and imagine future interactions before acting.
However, embodied generation [long2025survey, liu2025aligning, li2025comprehensive] 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, 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 [agarwal2026cosmos, zhang2026qwen, shang2026roboscape] attempt to bridge this gap by continually adapting pre-trained foundation models using robot trajectories [Neill2023open_x_embodiment, tian2026interndata, yin2026geniesim30, jiang2025galaxea, walke2023bridgedata, brohan2022rt] or egocentric manipulation videos [akhtyamov2025egowalk, grauman2022ego4d]. 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 [gadre2023datacomp, schuhmann2022laion, kakaobrain2022coyo-700m], 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 present Xiaomi-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-U0 reformulates embodied synthesis as a natural extension of foundation image and video generation. Starting from a world foundation model, Xiaomi-Robotics-U0 is 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 of Xiaomi-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-U0 further 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 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 that Xiaomi-Robotics-U0 establishes 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-U0 outperforms 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.
-
We introduce Xiaomi-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-U0 effectively 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-U0 achieves 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-U0 transforms 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
| (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
| (2) |
using standard next-token prediction (NTP).
To support embodied intelligence, Xiaomi-Robotics-U0 is 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; (ii) Any-to-Image (X2I), where one to three reference images together with text are used for image editing and generation; (iii) Embodied Scene Generation, where the model generates the initial multi-view robot observations conditioned on robot embodiment and scene descriptions; and (iv) Embodied Transfer, 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 (ii) 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.
2.2 Model Architecture
Figure 3 illustrates the overall architecture of Xiaomi-Robotics-U0. We adopt the open-sourced EMU3.5 [cui2025emu3] 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 [shi2025scalable] with a spatial compression ratio of . 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 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 [zhou2026flasharefficientposttrainingacceleration]. 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 [kwon2023efficient], Xiaomi-Robotics-U0 achieves up to faster image generation at 10241024 resolution compared with the original NTP pattern, while preserving most generation quality.
2.3 Dataset
2.3.1 Dataset Curation
Data Collection
To endow Xiaomi-Robotics-U0 with 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 [chen2024sharegpt4v] 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 [bu2025agibot], Open X-Embodiment [Neill2023open_x_embodiment], 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 [chen2025robotwin], GenieSim [yin2026geniesim30], and InternData-A1 [tian2026interndata], 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 [ren2025cosmosdrive], which provides high-quality driving scenarios with diverse weather, lighting, and traffic conditions.
-
Egocentric data, drawn from EgoWalk [akhtyamov2025egowalk], 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 [yeshwanth2023scannet++] and DL3DV-10K [ling2024dl3dv], both of which provide high-quality indoor and outdoor 3D scene captures.
-
Game data, sourced from Open-P2P [yue2026openp2p], 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 [chen2026abot], 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 [bai2025qwen3], 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:
-
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 [chen2025videodepthanything] 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 [McInnes2017hdbscanHD] 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 of Xiaomi-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.
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 and sequential training.
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 capabilities 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.
To enhance dynamic interaction modeling, we additionally incorporate large-scale embodied manipulation videos collected from InternData-A1 [tian2026interndata], GenieSim [yin2026geniesim30], AgiBotWorld-Beta [bu2025agibot], RoboCOIN [wu2025robocoin], and RoboMIND [wu2024robomind]. 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 to Xiaomi-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 of Xiaomi-Robotics-U0 and reduce the serial decoding cost during X2I inference.
2.5.1 FlashAR+ Extension
FlashAR baseline
FlashAR [zhou2026flasharefficientposttrainingacceleration] 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 of Xiaomi-Robotics-U0, extending it from single-image T2I generation to broader X2I scenarios, and use the resulting model for X2I inference acceleration in Xiaomi-Robotics-U0. Figure 6(a) demonstrates the decoding steps of the FlashAR+ extension.
In Xiaomi-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.
FlashAR+ mask and attention
Compared with standard autoregressive generation, Xiaomi-Robotics-U0 FlashAR+ Extension introduces the key changes summarized in Table 1.
| Aspect | Standard 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 heads | A single AR prediction head | Horizontal head, vertical head, and gated fusion | Supports anti-diagonal generation |
| Directional auxiliary loss | Standard supervision | Introduces H/V loss constraints | Constrains both directional heads |
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 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 6(b). 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 , its parallel step is defined as ; 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 and key token ,
FlashAR+ loss
Xiaomi-Robotics-U0 FlashAR+ Extension introduces a key adjustment to directional auxiliary supervision. We match the horizontal auxiliary loss weight to the vertical auxiliary loss weight:
Here, is the target-image prediction loss after H/V gated fusion, and supervise the horizontal and vertical heads respectively, and is the AR distillation term. For a target position , let and denote the logits from the horizontal and vertical heads, and let denote the gate value. The fused logits are
The corresponding fused loss is
The directional auxiliary losses and use the same target tokens to supervise and , respectively. The H head also needs to adapt to the target-region distribution. If , the H head may fail to converge 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 explicitly supervises the H head and stabilizes H/V fusion.
Xiaomi-Robotics-U0 FlashAR+ inference
During inference, Xiaomi-Robotics-U0 first encodes the text instruction 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 of , 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 to Xiaomi-Robotics-U0 FlashAR 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 of Xiaomi-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 classifier-free guidance.
As shown in Figure 7, the accelerated path preserves high generation quality on embodied tasks. On T2I generation at 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 speedup over AR, and FlashAR+ with vLLM further reduces the latency to 5.44 seconds per image with max_num_seq=28, achieving an additional 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 of 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 150 Easy samples (descriptions close to the underlying scene) and 150 Hard 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, ), (the fraction of pixels with relative depth error below , ), 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-U0 outperforms 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.
| Metric | GPT-Image-2 | Xiaomi-Robotics-U0 (Ours) | ||
| Easy | Hard | Easy | Hard | |
| Depth SI-RMSE () | 0.4007 | 0.3444 | 0.1407 (0.260) | 0.1435 (0.201) |
| Depth () | 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) |
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 by Xiaomi-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-U0 faithfully follows both the edited textual description and the multi-view depth 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.
We further compare Xiaomi-Robotics-U0 with 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 low 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-U0 faithfully 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 of Xiaomi-Robotics-U0 to 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 to Xiaomi-Robotics-U0 for 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 that Xiaomi-Robotics-U0 can substantially expand the visual coverage of the collected trajectories in a zero-shot manner.
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 compare Xiaomi-Robotics-U0 against 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 from Xiaomi-Robotics-U0 and 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-U0 outperforms 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-U0 generates geometrically coherent multi-view observations with consistent spatial layouts, while retaining photorealistic quality. This verifies that Xiaomi-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 generates photorealistic and geometrically consistent multi-view initial scenes for diverse scene prompts, demonstrating 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, yet Xiaomi-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.
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 a [intelligence2025pi_0_5] policy starting from the pi05_base checkpoint, 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 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 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 the Xiaomi-Robotics-U0 augmentation. 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, 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 has ordered milestones and the rollout completes the first of them, its progress is , where denotes full success. For task and group , the reported progress of the tested policy is defined as:
where is the number of rollouts for each task and group (three layouts three trials in our setting). A higher progress score for Xiaomi-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 representative Xiaomi-Robotics-U0-Aug rollouts under interference conditions for all three tasks. Despite challenging perturbations—held-out tablecloths, colored lighting, and projected patterns—the Xiaomi-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.
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 benchmark Xiaomi-Robotics-U0 on WorldArena [shang2026worldarena]. 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-U0 achieves the best overall EWMScore (73.64), ranking first among over 100 submitted models. In terms of controllability, Xiaomi-Robotics-U0 attains 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-U0 leads 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-U0 strikes 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 of Xiaomi-Robotics-U0 to 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 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.
| Xiaomi-Robotics-U0 | SisyphusWorld | BWM-Fast | SACWM | DexWorldEngine | ||
| Visual Quality | Image Quality | 53.94 | 45.57 | 51.22 | 45.83 | 51.44 |
| Aesthetic Quality | 40.79 | 38.38 | 40.15 | 37.30 | 40.96 | |
| JEPA Similarity | 90.60 | 96.58 | 97.87 | 96.26 | 86.71 | |
| Motion Quality | Dynamic Degree | 73.70 | 76.27 | 69.58 | 70.35 | 69.16 |
| Flow Score | 86.02 | 99.82 | 75.11 | 95.79 | 74.31 | |
| Motion Smoothness | 95.51 | 91.42 | 94.33 | 90.60 | 94.27 | |
| Content Consistency | Subject Consistency | 79.05 | 82.11 | 81.42 | 82.17 | 81.46 |
| Background Consistency | 86.44 | 86.52 | 90.17 | 87.90 | 88.85 | |
| Photometric Consistency | 2.13 | 5.69 | 3.08 | 6.50 | 4.42 | |
| Physics Adherence | Interaction Quality | 87.30 | 71.98 | 79.88 | 72.28 | 81.04 |
| Trajectory Accuracy | 41.89 | 44.58 | 44.89 | 46.95 | 51.41 | |
| 3D Accuracy | Depth Accuracy | 85.25 | 94.85 | 86.41 | 94.99 | 86.19 |
| Perspectivity | 98.84 | 90.34 | 97.14 | 92.34 | 97.58 | |
| Controll- ability | Instruction Following | 93.86 | 82.68 | 90.22 | 81.66 | 92.62 |
| Semantic Alignment | 89.35 | 89.18 | 89.15 | 89.16 | 89.45 | |
| EWMScore () | 73.64 | 73.06 | 72.71 | 72.67 | 72.66 | |
3.5 Text2Image & Image Edit Benchmarks
We evaluate the autoregressive Xiaomi-Robotics-U0 image model on general text-to-image generation and image editing benchmarks. The goal of this evaluation is not to claim that Xiaomi-Robotics-U0 is 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 [ghosh2023geneval] for compositional T2I evaluation and ImgEdit [ye2026imgedit] Basic for single-turn image editing.
| Model | Overall | Single object | Two object | Counting | Colors | Position | Attribute binding |
| Qwen-Image | 0.87 | 0.99 | 0.92 | 0.89 | 0.88 | 0.76 | 0.77 |
| Xiaomi-Robotics-U0-AR | 0.74 | 0.99 | 0.92 | 0.49 | 0.88 | 0.60 | 0.53 |
| Model | Overall | Add | Adjust | Extract | Replace | Remove | Background | Style | Hybrid | Action |
| Qwen-Image | 4.27 | 4.38 | 4.16 | 3.43 | 4.66 | 4.14 | 4.38 | 4.81 | 3.82 | 4.69 |
| Xiaomi-Robotics-U0-AR | 4.08 | 4.21 | 4.15 | 3.28 | 4.35 | 4.21 | 4.03 | 4.72 | 3.27 | 4.53 |
As shown in Table 4, Xiaomi-Robotics-U0 preserves 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 composition, precise spatial grounding, or symbolic binding.
Table 4 demonstrates that Xiaomi-Robotics-U0 retains 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-U0 maintains 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 [blattmann2023stable], FLUX [labs2025flux1kontextflowmatching], GPT-Image-2.0 [openai_gptimage_2025], Show-o [xie2025show], Janus [wu2025janus], Janus-Pro [chen2025janus], Chameleon [team2024chameleon], BAGEL [deng2025emerging], and EMU3.5 [cui2025emu3], 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 [bruce2024genie], Genie2 [parkerholder2024genie2], Sora [openai2024sora], MovieGen [polyak2024movie], Wan [wan2025wan], and Cosmos [agarwal2026cosmos], 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 [hafner2019dream], World Action Models (WAMs) [yuan2026fast, ye2026gigaworld, li2026causal, team2026motubrain], DreamZero [ye2026world], and Qwen-RobotWorld [zhang2026qwen], 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 manipulation through instruction-guided synthesis and reference-conditioned editing. Representative approaches, including InstructPix2Pix [brooks2023instructpix2pix], AnyDoor [chen2024anydoor], OmniGen [xiao2025omnigen], GPT-Image-2.0 [openai_gptimage_2025], and FLUX Kontext [labs2025flux1kontextflowmatching], 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 [shi2024mvdream], SyncDreamer [liu2024syncdreamer], Wonder3D [long2024wonder3d], Era3D [li2024era3d], Zero123++ [shi2023zero123++], and related multi-view diffusion models [gao2024cat3d, tang2024mvdiffusion++], 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 [polyak2024movie], Wan [wan2025wan], Cosmos [agarwal2026cosmos], Genie [bruce2024genie], and Qwen-RobotWorld [zhang2026qwen] demonstrate impressive capabilities in long-horizon video synthesis and future observation prediction. Meanwhile, large-scale robot datasets and simulation platforms, including Open X-Embodiment [o2024open], RoboCasa [nasiriany2024robocasa], ManiSkill [mu2021maniskill], BridgeData V2 [walke2023bridgedata], and DROID [khazatsky2024droid], have substantially accelerated the development of data-driven robot learning. More recently, embodied video datasets and synthetic data pipelines, such as RoboMIND [wu2024robomind], RoboCoin [wu2025robocoin], GenieSim [yin2026geniesim30], InternA1 [tian2026interndata], and AgiBotWorld [bu2025agibot], 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.
5 Conclusions
In this report, we presented Xiaomi-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-U0 unifies 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-U0 is 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-U0 has 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.
Contributions
∗Core contributors, alphabetically listed by first name ∗∗External contributor
-
Xinghang Li
-
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