Xiaomi-Robotics-U0：基于 World Foundation Model 的统一具身合成

近期 foundation image 和 video generation models 展现了很强的泛化与可控性，但直接应用到 embodied scenarios 时，会受到 multi-view consistency、几何一致性和 robot embodiment constraints 的限制。现有方法通常用规模有限的机器人数据适配 foundation model，容易牺牲大规模预训练获得的视觉知识。

本文提出 Xiaomi-Robotics-U0，一个 380 亿参数的 multimodal autoregressive model，用于统一 embodied synthesis。它把具身生成视为 foundation image/video generation 的延伸，并联合优化 text-to-image generation、image editing、embodied scene generation、embodied transfer 和 embodied video generation。统一框架在适应具身场景的同时，保留预训练 world foundation model 的泛化能力。

Xiaomi-Robotics-U0 支持跨多种 robot embodiment 的高质量 multi-view scene generation，并引入结构化、可控的 embodied transfer，在保持多视角一致性和交互动力学的同时实现细粒度编辑。实验在 single-step 和 sequential generation tasks 上达到 state-of-the-art；在 embodied scene generation 和 transfer 的人工评测中超过 GPT-Image-2.0，在 World Arena 的 embodied video generation 中排名第一，并将 pi_0.5 在困难真实操作任务上的 out-of-distribution success rate 从 36.9% 提高到 63.2%。

这些结果表明，foundation world model 既可以作为 embodied world model，也可以作为 embodied intelligence 的可扩展数据引擎。论文给出的代码与 checkpoint 页面为 robotics.xiaomi.com/xiaomi-robotics-u0.html。

Foundation image 和 video generation models 通过 Internet-scale data 的训练，在 semantic understanding、controllable generation 和 visual reasoning 方面取得巨大进展。它们可以根据多模态输入合成逼真的图像和视频，泛化到远离训练分布的内容，因此成为 embodied intelligence 的有吸引力的起点：机器人需要在执行动作前理解复杂环境并想象未来交互。

但具身生成与普通图像、视频 synthesis 有根本区别。具身场景要求多摄像机之间严格一致，保持跨视角的几何和物理 coherence，满足明确的 robot embodiment constraints，并维持时间一致的 interaction dynamics。生成的 observation 不仅要看起来逼真，还必须与机器人运动学、camera calibration 和下游 manipulation policy 兼容。直接套用通用 foundation model，往往会出现几何冲突、不合理的机器人状态以及与控制系统不兼容的问题。

近期 embodied world models 通过机器人轨迹或 egocentric manipulation videos 适配预训练 foundation model。它们推进了具身预测，但适配通常只使用 robot-specific datasets；这些数据相较 Internet-scale visual corpus 更小、更单一、更重复，因而会削弱 foundation model 原有的泛化、可控性和生成多样性。

本文提出 Xiaomi-Robotics-U0，把 foundation generation 和 embodied generation 放进同一个 continual training paradigm。模型同时使用 general-domain 和 embodied datasets，在统一 autoregressive objective 下学习 T2I、image editing、embodied scene generation、embodied transfer 和 embodied video generation。这样，模型可以保留丰富的 semantic knowledge 与 controllable generation，同时学习机器人中心的 multi-view reasoning 和 interaction modeling。

模型还把 embodied synthesis 从静态场景扩展到 sequential world modeling：通过 interleaved embodied sequences 和多时间分辨率的 manipulation videos，同时捕获长时域任务进展与细粒度交互动力学。更重要的是，生成的 embodied scene 可以继续 rollout 成时间连贯的 manipulation video，成为下游 policy learning 的可扩展 synthetic trajectory engine。

论文贡献

第一，论文提出 Xiaomi-Robotics-U0，声称这是首个在单一 autoregressive framework 中无缝整合 foundation image generation、image editing、embodied scene generation、embodied transfer 与 embodied video generation 的 unified embodied synthesis model。通过统一 continual training 同时优化通用域与具身任务，模型把 world foundation model 的泛化能力迁移到具身生成，同时保留原有视觉生成能力。

第二，论文提出统一的 multi-view embodied generation framework，覆盖 scene generation 与 controllable transfer。借助大规模 T2I/X2I 联合训练，模型强调跨视角一致性、细粒度可控性与 instruction following；结构化控制把 workspace、background、foreground irrelevant objects、target objects 和 lighting 解耦成独立维度，用于在保持几何与交互动力学的情况下规模化扩充具身视频。

第三，论文通过 interleaved subtask-subgoal learning 与 multi-FPS video generation，把静态具身合成扩展到 sequential embodied world modeling。模型能够从合成初始场景继续 rollout 长时域视频，从而把 foundation world model 转换为可扩展的 embodied data engine，为后续 policy learning 生成多样交互轨迹。

Problem Formulation

目标是构建一个统一的 embodied world model，在单一 autoregressive framework 下支持多种生成和预测任务。模型接收由文本、图像和机器人控制信号共同组成的多模态上下文，并预测其后的多模态序列。

\mathcal{C}=\{x_1,x_2,\cdots,x_n\}

P(\mathcal{Y}|\mathcal{C})=\prod_{t=1}^{T}P(y_t|y_{<t},\mathcal{C})

训练样本被分为两组。Single-step samples 描述没有显式时间演化的一次性输入输出映射：Text-to-Image（T2I）根据文本生成图像；Any-to-Image（X2I）根据文本和 1 至 3 张参考图完成编辑或生成；Embodied Scene Generation 根据 robot embodiment 与场景描述生成初始 multi-view observation；Embodied Transfer 则根据当前具身 observation 和目标场景描述预测迁移后的 multi-view observation。

Sequential samples 用于长时域交互与未来演化。第一类是 image-text interleaved subtask-subgoal sequences，每段子任务描述后跟随执行完成后的 multi-view observations；第二类是以多个 frame rate 采样的 embodied manipulation videos，稀疏序列刻画任务级转移，稠密序列刻画连续交互动力学。尽管语义不同，所有任务最终都表示成统一的多模态 token sequence，并在同一 autoregressive objective 下优化。

Model Architecture

模型使用开源 EMU3.5 作为初始化模型，其基础是 Qwen-3-32B decoder-only Transformer。图像首先通过 IBQ Tokenizer 编码，空间压缩比为 16×16；将 IBQ codebook 加入原始 Qwen vocabulary 后，文本和图像都可以在统一的离散 vocabulary 中通过 next-token prediction 建模。模型最终规模为 38B 参数。

与使用 task-specific prediction heads 的传统 embodied world model 不同，Xiaomi-Robotics-U0 把 image generation、embodied scene synthesis、scene transfer、sequential subtask prediction 和 video generation 都表述为单一 autoregressive sequence modeling problem。这样可以进行 multi-task co-training，并保留大规模预训练模型的视觉生成能力。

为提高推理效率，模型进一步采用受 FlashAR 启发的 inference-accelerating adaptation：额外加入 vertical prediction head，按 anti-diagonal 顺序解码图像 token，使同一对角步内的多个视觉 token 可以并行生成。结合优化的 KV-cache management 与 vLLM continuous batching，1024×1024 图像生成相对原始 NTP pattern 最多加速 82.9×，同时保留大部分生成质量。

Dataset and Unified Annotation

数据覆盖六个互补领域。General image-text 数据（如 ShareGPT4V）建立文本条件生成所需的视觉语义对齐；embodied manipulation 数据同时包含 AgiBotWorld-Beta、Open X-Embodiment、内部 MiBot 真实数据，以及 RoboTwin2.0、GenieSim、InternData-A1 仿真数据，使模型接触真实与难以实体采集的交互场景。

Autonomous driving 数据来自 Cosmos-Drive-Dreams，提供包含复杂布局、动态物体、天气、照明和交通条件的结构化室外场景；EgoWalk 的 egocentric 数据补充自然第一人称视角与运动模式；ScanNet++ v2 和 DL3DV-10K 的 3D reconstruction 数据提供密集 multi-view correspondence 与精确几何；Open-P2P 的 game 数据则把视觉分布扩展到风格化和程序生成环境。

大规模 raw data 经过 motion blur detection、image quality detection 和 gripper-state alignment 等过滤。最终 single-step generation 使用 950 万 samples、约 564 亿 tokens；sequential generation 使用 260 万 video clips、约 496 亿 tokens。

Dataset Labeling

多数原始样本只有粗粒度 task-level 描述，无法支持精确的文本驱动具身生成。作者因此用 Qwen3-VL-235B 构建统一标注管线，为通用帧、多源视频和机器人操作轨迹生成 dense structured annotations，并把流程拆成四个目标相互独立的分支。

General Scene Annotation 不区分通用视频还是机器人视频：先在时间维下采样清晰关键帧，再由 VLM 为每帧描述整体布局、物体交互和跨物体空间关系。Structured Embodied Scene Annotation 只处理具身操作数据，并依据轨迹的子任务分解，把场景拆为 workspace、task objects、irrelevant objects、lighting 和 background 五个正交语义维度，分别生成与任务相关的描述。

Geometric Modality Extraction 使用 Video Depth Anything 为全部帧和视频提取时间一致的 inverse depth；亮像素代表近表面，暗像素代表远表面，作为跨域 Embodied Transfer 的几何条件。Trajectory-level Subtask Decomposition 使用 HDBSCAN，根据 end-effector pose 与 gripper signal 切分机器人轨迹，定位运动或抓取状态切换边界；在物理交互密集的边界处采样关键帧，并由 VLM 生成与 frame boundaries 对齐的 subtask plan。

Task-Oriented Data Organization

标注后的数据按三个训练目标重新组织。Embodied Transfer 的输入由 system instruction、scene description 和 multi-view depth maps 构成，监督目标是 multi-view RGB，训练模型从几何先验恢复 photorealistic appearance。Scene Generation 用 system instruction 与 scene description 直接监督 multi-view RGB，学习不依赖几何输入的 text-to-multi-view synthesis。

Video Generation 的输入包括 system instruction、task instruction 和 initial observation frame，监督目标是 15 至 25 帧 RGB 序列。若数据集提供 robot action label，还会把由 action 和 URDF 渲染的 robot mask sequence 接在初始帧后作为动作引导，使模型能根据任务语义、初始状态和显式 action guidance 预测未来视觉状态。

Data Balancing

多域语料容易让模型偏向样本占比更高的分布。作者按 data domain 与 generation task type（multi-view generation、scene editing、single-view generation）做 sample-level reweighting，核心目标是完整覆盖机器人操作场景。首先剔除视觉损坏、时间戳错位、缺少 proprioception/action signal、执行失败以及没有机器人-物体交互的静态轨迹。

剩余样本按 task goal、scene context、target object、spatial relation、subtask phase 和 atomic manipulation skill 分层分组。重复的 embodiment-task-object 组合被下采样；具有新视觉、新物体布局、新操作技能或新时间阶段的 clip 被优先保留；罕见物体、低频技能以及 grasping、releasing、placing、object-state transition 等关键交互阶段得到更高采样权重。这比简单 dataset-level weighting 更细，既减少重复演示，又保留视觉与技能覆盖。

Training

为了把预训练图像生成模型高效适配为 unified embodied world model，同时保留原始视觉生成能力，论文采用两种 continued training setup：single-step training 与 sequential training。

Single-step training 联合训练 T2I、X2I、Embodied Scene Generation 和 Embodied Transfer。T2I/X2I 持续强化预训练模型已有的通用生成与编辑能力，抑制 embodied adaptation 中的 catastrophic forgetting；Scene Generation 教模型根据 robot embodiment 与场景描述生成物理合理的初始 observation；Transfer 则学习在保持 robot interaction state 与 camera geometry 的同时做跨场景 multi-view generation。

Sequential training 把能力扩展到长时域推理和交互建模。image-text interleaved subtask-subgoal sequence 将语言子任务与执行后的 multi-view observations 配对，用来学习任务分解和未来 observation 预测。随后加入 InternData-A1、GenieSim、AgiBotWorld-Beta、RoboCOIN 与 RoboMIND 的大规模操作视频，并以 1、3、5 FPS 组织多时间分辨率序列：低 FPS 强化长时域规划与未来预测，高 FPS 捕获细粒度操作动力学和连续交互。

Inference

推理加速只作用于最终 target image region，reference images、文本与其他输入信号仍按原条件建模。所有先于目标图像的内容，包括 text instruction、reference images、control conditions 和 target-image header，都被视为 conditional prefix；并行 autoregressive loss 只在 target image 上计算，避免把参考图误当成生成目标。

FlashAR+ Extension

标准 AR 在 raster order 中逐 token 串行生成；FlashAR+ 把 target image 按 anti-diagonal steps 分组并行生成。Conditional prefix 保持标准自回归可见性，target token 可以看完整前缀，但在目标图内部只能看更早的对角步；同一步 token 彼此不可见，因而不会在训练时发生 future-information leakage。

M(q,k)={1}\left[s(r_k,c_k)<s(r_q,c_q)\right]

\mathcal{L}=\mathcal{L}_{\mathrm{fuse}}+0.05\mathcal{L}_h+0.05\mathcal{L}_v+0.2\mathcal{L}_{\mathrm{distill}}

z_{\mathrm{fuse}}(r,c)=g(r,c)z_h(r,c)+\left(1-g(r,c)\right)z_v(r,c)

\mathcal{L}_{\mathrm{fuse}}=-\sum_{(r,c)\in\mathbf{Y}}\log\mathrm{softmax}\left(z_{\mathrm{fuse}}(r,c)\right)_{y_{r,c}}

论文特别指出，若 L_h=0，horizontal head 可能无法充分收敛，而 H/V gate 在训练后期又可能塌缩到这个欠训练分支，使融合预测退化为不稳定的单方向预测。加入 0.05L_h 后，两个方向受到对称的显式监督，融合过程更稳定。

FlashAR+ Inference and vLLM

推理时先把 text instruction 和 reference images 编码成 conditional prefix，再接 target-image header。模型仅在目标网格上按 r+c 递增逐步生成，同一步内所有位置并行采样；边界位置使用可用的单方向预测，内部位置使用 H/V gated fusion。这样既保留 X2I 的多源条件约束，又减少目标图区域的串行成本。

vLLM integration 不改变 FlashAR+ 的解码规则：vLLM 负责 conditional prefix、batched execution 与 paged KV-cache，diagonal scheduler 负责发射每个 anti-diagonal step 的视觉 token，并同步 classifier-free guidance 所需的 conditional/unconditional request。单张 H20 上生成 1024×1024 T2I 图像时，标准 AR 需 450.77 秒；FlashAR+ 降到 16.56 秒，较 AR 加速 27.22×；加入 vLLM 且 max_num_seq=28 后降到 5.44 秒，较 FlashAR+ 再加速 3.04×。

实验围绕 Embodied Transfer、Embodied Scene Generation、Real World Experiments、Video Generation 以及通用 T2I/Image Edit 五个方面展开。总体结果显示，模型在保留 general image capability 的同时，提高了多视角几何一致性、具身迁移可控性和真实机器人策略的分布外鲁棒性。

Embodied Transfer

作者构建了 300 个样本的 benchmark，每个样本包含文本场景描述和一组 multi-view depth maps，模型需要生成与深度条件一致的 multi-view RGB images。150 个 Easy 样本的描述接近原场景，150 个 Hard 样本要求生成较多训练分布中没有的新内容。评测从 depth consistency、structural fidelity 和 semantic grounding 三方面进行。

每个样本还提供原数据中的 multi-view RGB reference。参考图并不完全对应编辑后的描述，但包含 depth 之外的结构信息，供指标计算使用。Depth consistency 使用共享单目深度估计器比较生成图与参考图，报告 SI-RMSE（越低越好）、δ1（相对深度误差小于 1.25 的像素比例，越高越好）和 AbsRel（越低越好）；structural fidelity 使用 Canny edge 的 F1、Precision、Recall；semantic grounding 使用 open-vocabulary detection/segmentation 的 mIoU 与 Recall。

GPT-Image-2 在单张图像中通常具有不错的文本对齐和视觉质量，但会违反 depth conditions，并在同一场景的不同视角中产生互相冲突的对象位置、尺度和布局。U0 则遵守每个视角的几何约束，并保持跨视角一致。

作者还用 Embodied Transfer 做 zero-shot data augmentation：对每条真实轨迹，让 Qwen3-VL 为 background、workspace 和 lighting 等维度生成 1,000 个候选描述，随机组合后交给 U0 生成新场景。这样可以扩展机器人轨迹的视觉覆盖范围，同时保持原有 robot states 和 action labels。

该扩充应用于 Pack Box、Fold Towel、Pack Phone 和 Pack Earphone 四项真实任务。Task objects 保持不变，其余维度尽量多样化；论文强调这些组合没有出现在训练集中，因此这项实验检验的是 zero-shot visual coverage expansion，而非重放训练场景。

Embodied Scene Generation

Scene generation benchmark 包含 400 个样本：200 个 Easy indoor manipulation scenes 和 200 个 Hard open-domain scenes。每个 prompt 指定 robot arm type，以及 workspace layout、object configuration、lighting 和 background。由于没有可靠的自动指标评估全局跨视角 3D coherence，作者使用随机左右顺序的人工 pairwise evaluation，比较 multi-view geometric consistency 和 textual instruction following。U0 在大多数比较中胜过 GPT-Image-2，同时保持 photorealistic quality。

为公平比较，GPT-Image-2 还会获得与目标 robot arm type 对应的 multi-view reference images 作为视觉上下文。标注者看到随机打乱左右顺序的两组输出，以降低位置偏好；multi-view consistency 要求跨视角的对象位置、尺度和遮挡对应同一个 3D layout，instruction following 则核对 workspace、objects、lighting 与 background 是否符合 prompt。

值得注意的是，四种 robot arm 中 AGIBOT G2 只在 simulation data 上训练，U0 仍能为它生成符合真实世界视觉先验的 observation，显示出潜在的 sim-to-real generalization。

Real World Experiments

真实世界实验研究生成数据是否能改善下游 policy learning。三项任务是 Store Earphones、Fold Towel 和 Pack Box，分别代表灵巧操作、可变形物体处理和多物体交互。每个任务使用约 40 小时真实 demonstration；U0-Aug 再生成约 40 小时 style-transferred episodes，改变背景、lighting 和 textures，但保留 robot states 与 action labels。Original policy 只使用 clean data，U0-Aug 使用 clean data 与 augmentation 的混合。

每项任务都从 pi05_base checkpoint 出发，使用官方 PyTorch 实现独立 post-train 一个 π0.5 policy；除 supervised fine-tuning data mixture 外，所有训练设置完全一致。机器人执行时不挂载 world model，U0 只通过训练数据影响 policy，因而实验测到的是数据增强价值，而不是推理时额外模型的收益。

每个 group 含 3 种 object-layout variant，两种 policy 按相同条件日程评测。Base group 只改变物体位置，背景 tablecloth 来自 demonstration，使用标准实验室照明；interference group 同样随机布局，但换成训练中未见、纹理和反射更复杂的 tablecloth，并加入低照度、彩色光与 disco-ball 式动态投影。每个 policy-task pair 共 2 groups × 3 layouts × 3 trials = 18 次真实机器人 rollout。

平台是双臂 ARX robot：两条 6-DoF arm，各带 1-DoF parallel-jaw gripper。Observation 包括 proprioceptive arm state、1 个 head RGB view 与左右 wrist RGB view；耳机和毛巾任务使用 Intel RealSense D405，装箱任务为覆盖更大工作区改用鱼眼相机并在输入前去畸变。Policy server 通过局域网 WebSocket 与控制进程通信，机器人以 1 Hz 上传状态和三路图像，服务器返回执行到下一次 query 的 action chunk。

\mathrm{Prog}(t,g)=\frac{1}{N}\sum_{i=1}^{N}\frac{\ell_i}{K_t}

每个 task、evaluation group 和 layout 运行三次真实机器人试验。Base group 只改变物体摆放，使用训练中见过的 tablecloth 和正常灯光；interference group 使用未见过的 tablecloth、低光、彩色光和 disco-ball 式动态图案。结果显示 U0-Aug 在 base group 与 Original 大致相当，但在 interference group 的所有任务上都更好，说明 style-transferred augmentation 学到了对新背景和 lighting 的视觉不变性。

Base group 上 U0-Aug 偶尔略低于 Original，这是增强数据占用训练容量、却没有为 in-distribution 场景增加新信息的预期代价；但 distribution shift 下的收益明显超过这点损失。典型成功 rollout 显示，U0-Aug 在新 tablecloth、彩色照明和投影图案下仍能完成收纳两只耳塞并合盖、完整折叠毛巾，以及放入四件物体后关箱。

失败分析列出六类典型错误：饱和光照下抓不到耳塞；误判耳机盒方向；毛巾与桌布视觉相近时找不到边缘；因未完整抓住布边而把已折好的毛巾重新展开；以及两例把投影光斑当作目标去抓。两种 policy 在严重视觉污染下都可能异常，但 U0-Aug 更常继续正确动作或短暂偏离后自我修正；Original 更容易卡在视觉伪影上或进入不可恢复状态。

Embodied Video Generation and WorldArena

WorldArena 从 visual quality、motion quality、content consistency、physics adherence、3D accuracy 和 controllability 六个维度共 15 个指标评估视频。模型处理初始 frame、language instruction 和 robot action，并通过由 action、camera parameters 与 URDF 渲染的 robot mask sequence 提供动作引导；输出形式是带空 language token 的多帧 interleaved sequence。U0 的 EWMScore 为 73.64，在 100 多个提交模型中排名第一。

分项上，U0 的 Instruction Following 为 93.86、Interaction Quality 为 87.30、Perspectivity 为 98.84、Motion Smoothness 为 95.51，均为对比中的最高值。它并非每个指标都第一，例如 Trajectory Accuracy 和 Depth Accuracy 仍落后于个别模型；论文的结论是 U0 在各维度之间取得最佳综合平衡，而不是全面统治所有子指标。

在 WorldArena 之外，作者还先用 U0 生成全新的 initial scene，再按语言指令生成后续视频。示例中，robot arm 能抓起杯子、抬升并稳定放回桌面；另一个序列同时模拟背景人物走近桌面；纸张示例则表现出弯曲以及横向移动时因低质量产生的轻微滞后，用来说明模型对 scene-level dynamics、接触和可变形物体物理的建模能力。

通用 T2I 与 Image Edit

这组评测的目的不是宣称 U0 是专用 open-domain image generator，而是量化模型向具身场景适配后保留了多少通用生成和编辑能力。GenEval 用于 compositional T2I，分数范围为 0–1；ImgEdit Basic 用于 single-turn image editing，judge 分数范围为 1–5。

U0-AR 的 GenEval overall 为 0.74，Qwen-Image 为 0.87；ImgEdit overall 为 4.08，Qwen-Image 为 4.27。U0 在 single/two object、colors、Adjust 和 Remove 等常见生成或局部编辑上接近甚至局部超过基线，下降主要集中于 Counting、Position、Attribute binding、Hybrid 等需要 multi-constraint composition、精确 spatial grounding 或 symbolic binding 的复杂场景。结果支持‘没有 catastrophic forgetting’，但不等于通用能力毫无代价。

Foundation Generative Models and Embodied World Models

Foundation generative models 从图像 synthesis 逐渐发展到统一 multimodal world modeling。Stable Diffusion、FLUX、GPT-Image、Show-o、Janus、BAGEL 和 EMU3.5 等模型展示了 semantic understanding、instruction following、compositional reasoning 和 controllable visual generation；Genie、Sora、MovieGen、Wan 和 Cosmos 则把能力扩展到长时域视频和交互式 world simulation。它们虽泛化强，但通常没有显式建模 embodied manipulation 所需的几何约束、机器人 embodiment、calibrated camera 和物理交互动力学。

Embodied world models 通过 manipulation trajectories 与 robot videos 学习未来 observation，用于 planning、simulation 或 policy learning，代表工作包括 Dreamer 系列、World Action Models、DreamZero 和 Qwen-RobotWorld。与这些主要依赖 robot-only dataset 的 adaptation 不同，U0 将具身生成视为 foundation image/video generation 的自然延伸，同时优化 general-domain generation 与 embodied synthesis，以保留视觉多样性、语义丰富度和通用知识。

Controllable and Multi-view Embodied Generation

Controllable image editing 和 multi-view generation 分别解决文本/参考图控制和跨视角一致性，但许多方法面向 natural image 或 object-centric 3D generation，不处理机器人 embodiment、calibrated manipulation cameras 和 interaction states。U0 把 workspace、background、foreground irrelevant objects、target objects 和 lighting 解耦为独立控制维度，把 controllable editing 与 multi-view embodied generation 结合起来。

前一方向的代表方法包括 InstructPix2Pix、AnyDoor、OmniGen、GPT-Image-2.0 和 FLUX Kontext；后一方向包括 MVDream、SyncDreamer、Wonder3D、Era3D、Zero123++ 等。论文认为，前者通常假设 single-view natural image，后者主要服务 3D reconstruction 或 asset generation，二者都没有完整覆盖机器人几何、标定相机和交互状态。

Embodied Video Generation and Synthetic Data Engines

Embodied video generation 和 synthetic data engines 近年用于规模化机器人学习，但现有方法常把 isolated images、simulation assets 或 short robot videos 分开合成，缺少对 controllable scene generation、embodied transfer 和 long-horizon trajectory rollout 的联合建模。U0 通过 multi-resolution embodied video 和 interleaved subtask-subgoal learning，把 foundation world model 转成可服务下游 policy learning 的 embodied data engine。

相关生态一侧是 MovieGen、Wan、Cosmos、Genie、Qwen-RobotWorld 等 video/world model，另一侧是 Open X-Embodiment、RoboCasa、ManiSkill、BridgeData V2、DROID、RoboMIND、RoboCOIN、GenieSim、InternData-A1 和 AgiBotWorld 等数据集或平台。U0 的差异化主张是把 scene generation、transfer、subtask-subgoal prediction 和 video generation 放到同一个 AR model 中联合学习。

本文将模型定位为统一的 embodied world model 和可扩展的数据引擎。关键机制是 continual multi-task training，而不是把 foundation model 完全改造成 robot-only model。模型通过统一 token sequence 共享视觉知识，再用 depth、robot mask、URDF、camera information 和 temporal video 数据补充具身约束。

本文提出 Xiaomi-Robotics-U0，一个在 world foundation model 上通过 general-domain 与 embodied datasets 的 continual training 得到的统一 embodied synthesis model。它在单一 autoregressive framework 内统一 text-to-image、image editing、multi-view embodied scene generation、embodied transfer 和 embodied video generation，并将 foundation model 的视觉知识迁移到 embodied intelligence。

实验表明，U0 在 single-step 和 sequential embodied generation 上取得 state-of-the-art，在人工评测中超过 GPT-Image-2.0，在 WorldArena 上取得领先结果，并通过生成数据提升下游 robot policy 的鲁棒性。论文希望这成为构建可扩展 agentic data engine 的一步：模型不仅预测世界，还能持续生成更多样的机器人经验。

论文按角色列出完整贡献者。星号表示核心贡献者，按名字字母顺序排列；双星号表示外部贡献者。以下姓名与分工保持论文原文，不翻译人名。

作者衷心感谢更广泛团队给予的大力支持，包括未在上方贡献列表中列出的 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 和 Yiwei Zhou。

参考文献条目保留英文题名和原始 bibliographic 信息，不翻译文献内容。下列条目从 arXiv TeX Source 的 main.bib 中筛选，仅保留正文实际引用的文献。

1. Agarwal, Niket; Ali, Arslan; Allen, Jon; Antolini, Martin; Aubame, Adeline; Azzolini, Alisson; Bai, Junjie; Bala, Maciej; Balaji, Yogesh; Bapst, Josh; others (2026). Cosmos 3: Omnimodal world models for physical ai. arXiv preprint arXiv:2606.02800.
2. Akhtyamov, Timur; Mdfaa, Mohamad Al; Benavides, Javier Antonio Ramirez; Nigmatzyanov, Arthur; Bakulin, Sergey; Devchich, German; Fatykhov, Denis; Salinas, Diego Ruiz; Mazurov, Alexander; Zipa, Kristina; others (2025). Egowalk: A multimodal dataset for robot navigation in the wild. arXiv preprint arXiv:2505.21282.
3. 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; Luke Smith (2025). FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space.
4. Blattmann, Andreas; Dockhorn, Tim; Kulal, Sumith; Mendelevitch, Daniel; Kilian, Maciej; Lorenz, Dominik; Levi, Yam; English, Zion; Voleti, Vikram; Letts, Adam; others (2023). Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127.
5. Brohan, Anthony; Brown, Noah; Carbajal, Justice; Chebotar, Yevgen; Dabis, Joseph; Finn, Chelsea; Gopalakrishnan, Keerthana; Hausman, Karol; Herzog, Alex; Hsu, Jasmine; others (2022). RT-1: Robotics Transformer for Real-World Control at Scale. arXiv preprint arXiv:2212.06817.
6. Brooks, Tim; Holynski, Aleksander; Efros, Alexei A (2023). Instructpix2pix: Learning to follow image editing instructions. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition.
7. Bruce, Jake; Dennis, Michael D; Edwards, Ashley; Parker-Holder, Jack; Shi, Yuge; Hughes, Edward; Lai, Matthew; Mavalankar, Aditi; Steigerwald, Richie; Apps, Chris; others (2024). Genie: Generative interactive environments. Forty-first International Conference on Machine Learning.
8. Bu, Qingwen; Cai, Jisong; Chen, Li; Cui, Xiuqi; Ding, Yan; Feng, Siyuan; Gao, Shenyuan; He, Xindong; Hu, Xuan; Huang, Xu; others (2025). AgiBot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669.
9. Chen, Lin; Li, Jinsong; Dong, Xiaoyi; Zhang, Pan; He, Conghui; Wang, Jiaqi; Zhao, Feng; Lin, Dahua (2024). Sharegpt4v: Improving large multi-modal models with better captions. European Conference on Computer Vision.
10. Chen, Sili; Guo, Hengkai; Zhu, Shengnan; Zhang, Feihu; Huang, Zilong; Feng, Jiashi; Kang, Bingyi (2025). Video depth anything: Consistent depth estimation for super-long videos. Proceedings of the Computer Vision and Pattern Recognition Conference.
11. Chen, Tianxing; Chen, Zanxin; Chen, Baijun; Cai, Zijian; Liu, Yibin; Li, Zixuan; Liang, Qiwei; Lin, Xianliang; Ge, Yiheng; Gu, Zhenyu; others (2025). Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation. arXiv preprint arXiv:2506.18088.
12. Chen, Xi; Huang, Lianghua; Liu, Yu; Shen, Yujun; Zhao, Deli; Zhao, Hengshuang (2024). Anydoor: Zero-shot object-level image customization. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition.
13. Chen, Xiaokang; Wu, Zhiyu; Liu, Xingchao; Pan, Zizheng; Liu, Wen; Xie, Zhenda; Yu, Xingkai; Ruan, Chong (2025). Janus-pro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811.
14. Chen, Yuzhi; Chen, Ronghan; Huo, Dongjie; Yang, Yandan; Qi, Dekang; Liu, Haoyun; Lin, Tong; Zeng, Shuang; Xiao, Junjin; Chang, Xinyuan; others (2026). Abot-physworld: Interactive world foundation model for robotic manipulation with physics alignment. arXiv preprint arXiv:2603.23376.
15. 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; Maoqing Yao (2026). Genie Sim 3.0 : A High-Fidelity Comprehensive Simulation Platform for Humanoid Robot.
16. Cui, Yufeng; Chen, Honghao; Deng, Haoge; Huang, Xu; Li, Xinghang; Liu, Jirong; Liu, Yang; Luo, Zhuoyan; Wang, Jinsheng; Wang, Wenxuan; others (2025). Emu3. 5: Native multimodal models are world learners. arXiv preprint arXiv:2510.26583.
17. Deng, Chaorui; Zhu, Deyao; Li, Kunchang; Gou, Chenhui; Li, Feng; Wang, Zeyu; Zhong, Shu; Yu, Weihao; Nie, Xiaonan; Song, Ziang; others (2025). Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683.
18. Gadre, Samir Yitzhak; Ilharco, Gabriel; Fang, Alex; Hayase, Jonathan; Smyrnis, Georgios; Nguyen, Thao; Marten, Ryan; Wortsman, Mitchell; Ghosh, Dhruba; Zhang, Jieyu; others (2023). Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems.
19. Gao, Ruiqi; Holynski, Aleksander; Henzler, Philipp; Brussee, Arthur; Martin-Brualla, Ricardo; Srinivasan, Pratul; Barron, Jonathan T; Poole, Ben (2024). Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314.
20. Ghosh, Dhruba; Hajishirzi, Hannaneh; Schmidt, Ludwig (2023). Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems.
21. Grauman, Kristen; Westbury, Andrew; Byrne, Eugene; Chavis, Zachary; Furnari, Antonino; Girdhar, Rohit; Hamburger, Jackson; Jiang, Hao; Liu, Miao; Liu, Xingyu; others (2022). Ego4d: Around the world in 3,000 hours of egocentric video. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition.
22. Hafner, Danijar; Lillicrap, Timothy; Ba, Jimmy; Norouzi, Mohammad (2019). Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603.
23. Intelligence, Physical; Black, Kevin; Brown, Noah; Darpinian, James; Dhabalia, Karan; Driess, Danny; Esmail, Adnan; Equi, Michael; Finn, Chelsea; Fusai, Niccolo; others (2025). $\pi_0.5$: A Vision-Language-Action Model with Open-World Generalization. arXiv preprint arXiv:2504.16054.
24. 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; Tim Rocktäschel (2024). Genie 2: A Large-Scale Foundation World Model.
25. Jiang, Tao; Yuan, Tianyuan; Liu, Yicheng; Lu, Chenhao; Cui, Jianning; Liu, Xiao; Cheng, Shuiqi; Gao, Jiyang; Xu, Huazhe; Zhao, Hang (2025). Galaxea open-world dataset and g0 dual-system vla model. arXiv preprint arXiv:2509.00576.
26. Junkang Zhou; Yefei He; Feng Chen; Weijie Wang; Bohan Zhuang (2026). FlashAR: Efficient Post-Training Acceleration for Autoregressive Image Generation.
27. Khazatsky, Alexander; Pertsch, Karl; Nair, Suraj; Balakrishna, Ashwin; Dasari, Sudeep; Karamcheti, Siddharth; Nasiriany, Soroush; Srirama, Mohan Kumar; Chen, Lawrence Yunliang; Ellis, Kirsty; others (2024). Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.12945.
28. Leland McInnes; John Healy; S. Astels (2017). hdbscan: Hierarchical density based clustering. J. Open Source Softw..
29. Li, Lin; Zhang, Qihang; Luo, Yiming; Yang, Shuai; Wang, Ruilin; Han, Fei; Yu, Mingrui; Gao, Zelin; Xue, Nan; Zhu, Xing; others (2026). Causal World Modeling for Robot Control. arXiv preprint arXiv:2601.21998.
30. Li, Peng; Liu, Yuan; Long, Xiaoxiao; Zhang, Feihu; Lin, Cheng; Li, Mengfei; Qi, Xingqun; Zhang, Shanghang; Luo, Wenhan; Tan, Ping; others (2024). Era3d: High-resolution multiview diffusion using efficient row-wise attention. Advances in Neural Information Processing Systems.
31. Li, Xinqing; He, Xin; Zhang, Le; Wu, Min; Li, Xiaoli; Liu, Yun (2025). A comprehensive survey on world models for embodied ai. arXiv preprint arXiv:2510.16732.
32. Ling, Lu; Sheng, Yichen; Tu, Zhi; Zhao, Wentian; Xin, Cheng; Wan, Kun; Yu, Lantao; Guo, Qianyu; Yu, Zixun; Lu, Yawen; others (2024). Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.
33. Liu, Yang; Chen, Weixing; Bai, Yongjie; Liang, Xiaodan; Li, Guanbin; Gao, Wen; Lin, Liang (2025). Aligning cyber space with physical world: A comprehensive survey on embodied ai. IEEE/ASME Transactions on Mechatronics.
34. Liu, Yuan; Lin, Cheng; Zeng, Zijiao; Long, Xiaoxiao; Liu, Lingjie; Komura, Taku; Wang, Wenping (2024). Syncdreamer: Generating multiview-consistent images from a single-view image. International conference on learning representations.
35. Long, Qian; Wang, Yueze; Song, Jiaxi; Zhang, Junbo; Li, Peiyan; Wang, Wenxuan; Wang, Yuqi; Li, Haoyang; Xie, Shaoxuan; Yao, Guocai; others (2026). Scaling world model for hierarchical manipulation policies. arXiv preprint arXiv:2602.10983.
36. Long, Xiaoxiao; Guo, Yuan-Chen; Lin, Cheng; Liu, Yuan; Dou, Zhiyang; Liu, Lingjie; Ma, Yuexin; Zhang, Song-Hai; Habermann, Marc; Theobalt, Christian; others (2024). Wonder3d: Single image to 3d using cross-domain diffusion. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition.
37. Long, Xiaoxiao; Zhao, Qingrui; Zhang, Kaiwen; Zhang, Zihao; Wang, Dingrui; Liu, Yumeng; Shu, Zhengjie; Lu, Yi; Wang, Shouzheng; Wei, Xinzhe; others (2025). A survey: Learning embodied intelligence from physical simulators and world models. arXiv preprint arXiv:2507.00917.
38. Minwoo, Byeon; Beomhee, Park; Haecheon, Kim; Sungjun, Lee; Woonhyuk Baek; Saehoon, Kim (2022). COYO-700M: Image-Text Pair Dataset. https://github.com/kakaobrain/coyo-dataset.
39. Mu, Tongzhou; Ling, Zhan; Xiang, Fanbo; Yang, Derek; Li, Xuanlin; Tao, Stone; Huang, Zhiao; Jia, Zhiwei; Su, Hao (2021). Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483.
40. Nasiriany, Soroush; Maddukuri, Abhiram; Zhang, Lance; Parikh, Adeet; Lo, Aaron; Joshi, Abhishek; Mandlekar, Ajay; Zhu, Yuke (2024). Robocasa: Large-scale simulation of everyday tasks for generalist robots. arXiv preprint arXiv:2406.02523.
41. 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; others (2023). Open X-Embodiment: Robotic Learning Datasets and RT-X Models. https://arxiv.org/abs/2310.08864.
42. OpenAI (2024). Video Generation Models as World Simulators. https://openai.com/research/video-generation-models-as-world-simulators.
43. OpenAI (2025). GPT Image API. https://platform.openai.com/docs/guides/image-generation.
44. O’Neill, Abby; Rehman, Abdul; Maddukuri, Abhiram; Gupta, Abhishek; Padalkar, Abhishek; Lee, Abraham; Pooley, Acorn; Gupta, Agrim; Mandlekar, Ajay; Jain, Ajinkya; others (2024). Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. 2024 IEEE International Conference on Robotics and Automation (ICRA).
45. Podell, Dustin; English, Zion; Lacey, Kyle; Blattmann, Andreas; Dockhorn, Tim; Müller, Jonas; Penna, Joe; Rombach, Robin (2024). Sdxl: Improving latent diffusion models for high-resolution image synthesis. International Conference on Learning Representations.
46. Polyak, Adam; Zohar, Amit; Brown, Andrew; Tjandra, Andros; Sinha, Animesh; Lee, Ann; Vyas, Apoorv; Shi, Bowen; Ma, Chih-Yao; Chuang, Ching-Yao; others (2024). Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720.
47. Ren, Xuanchi; Lu, Yifan; Cao, Tianshi; Gao, Ruiyuan; Huang, Shengyu; Sabour, Amirmojtaba; Shen, Tianchang; Pfaff, Tobias; Wu, Jay Zhangjie; Chen, Runjian; others (2025). Cosmos-drive-dreams: Scalable synthetic driving data generation with world foundation models. arXiv preprint arXiv:2506.09042.
48. Schuhmann, Christoph; Beaumont, Romain; Vencu, Richard; Gordon, Cade; Wightman, Ross; Cherti, Mehdi; Coombes, Theo; Katta, Aarush; Mullis, Clayton; Wortsman, Mitchell; others (2022). Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems.
49. Shang, Yu; Li, Zhuohang; Ma, Yiding; Su, Weikang; Jin, Xin; Wang, Ziyou; Jin, Lei; Zhang, Xin; Tang, Yinzhou; Su, Haisheng; others (2026). Worldarena: A unified benchmark for evaluating perception and functional utility of embodied world models. arXiv preprint arXiv:2602.08971.
50. Shang, Yu; Zhang, Xin; Tang, Yinzhou; Jin, Lei; Gao, Chen; Wu, Wei; Li, Yong (2026). Roboscape: Physics-informed embodied world model. Advances in Neural Information Processing Systems.
51. Shi, Fengyuan; Luo, Zhuoyan; Ge, Yixiao; Yang, Yujiu; Shan, Ying; Wang, Limin (2025). Scalable image tokenization with index backpropagation quantization. Proceedings of the IEEE/CVF International Conference on Computer Vision.
52. Shi, Ruoxi; Chen, Hansheng; Zhang, Zhuoyang; Liu, Minghua; Xu, Chao; Wei, Xinyue; Chen, Linghao; Zeng, Chong; Su, Hao (2023). Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110.
53. Shi, Yichun; Wang, Peng; Ye, Jianglong; Mai, Long; Li, Kejie; Yang, Xiao (2024). Mvdream: Multi-view diffusion for 3d generation. International conference on learning representations.
54. Shuai Bai; Yuxuan Cai; Ruizhe Chen; Keqin Chen; Xionghui Chen; Zesen Cheng; Lianghao Deng; Wei Ding; Chang Gao; Chunjiang Ge; others (2025). Qwen3-VL Technical Report. arXiv preprint arXiv:2511.21631.
55. Tang, Shitao; Chen, Jiacheng; Wang, Dilin; Tang, Chengzhou; Zhang, Fuyang; Fan, Yuchen; Chandra, Vikas; Furukawa, Yasutaka; Ranjan, Rakesh (2024). Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. European Conference on Computer Vision.
56. Team, Chameleon (2024). Chameleon: Mixed-modal early-fusion foundation models, 2024. URL https://arxiv. org/abs/2405.09818.
57. Team, MotuBrain; Xiang, Chendong; Bao, Fan; Liu, Haitian; Tan, Hengkai; Bi, Hongzhe; Li, James; Liu, Jiabao; Pang, Jingrui; Jing, Kiro; others (2026). Motubrain: An advanced world action model for robot control. arXiv preprint arXiv:2604.27792.
58. Tian, Yang; Yang, Yuyin; Xie, Yiman; Cai, Zetao; Shi, Xu; Gao, Ning; Liu, Hangxu; Jiang, Xuekun; Qiu, Zherui; Yuan, Feng; others (2026). Interndata-a1: Pioneering high-fidelity synthetic data for pre-training generalist policy. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.
59. Walke, Homer Rich; Black, Kevin; Zhao, Tony Z; Vuong, Quan; Zheng, Chongyi; Hansen-Estruch, Philippe; He, Andre Wang; Myers, Vivek; Kim, Moo Jin; Du, Max; others (2023). Bridgedata v2: A dataset for robot learning at scale. Conference on Robot Learning.
60. Wan, Team; Wang, Ang; Ai, Baole; Wen, Bin; Mao, Chaojie; Xie, Chen-Wei; Chen, Di; Yu, Feiwu; Zhao, Haiming; Yang, Jianxiao; others (2025). Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314.
61. Woosuk Kwon; Zhuohan Li; Siyuan Zhuang; Ying Sheng; Lianmin Zheng; Cody Hao Yu; Joseph E. Gonzalez; Hao Zhang; Ion Stoica (2023). Efficient Memory Management for Large Language Model Serving with PagedAttention. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles.
62. Wu, Chengyue; Chen, Xiaokang; Wu, Zhiyu; Ma, Yiyang; Liu, Xingchao; Pan, Zizheng; Liu, Wen; Xie, Zhenda; Yu, Xingkai; Ruan, Chong; others (2025). Janus: Decoupling visual encoding for unified multimodal understanding and generation. Proceedings of the Computer Vision and Pattern Recognition Conference.
63. Wu, Kun; Hou, Chengkai; Liu, Jiaming; Che, Zhengping; Ju, Xiaozhu; Yang, Zhuqin; Li, Meng; Zhao, Yinuo; Xu, Zhiyuan; Yang, Guang; others (2024). Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation. arXiv preprint arXiv:2412.13877.
64. Wu, Shihan; Liu, Xuecheng; Xie, Shaoxuan; Wang, Pengwei; Li, Xinghang; Yang, Bowen; Li, Zhe; Zhu, Kai; Wu, Hongyu; Liu, Yiheng; others (2025). Robocoin: An open-sourced bimanual robotic data collection for integrated manipulation. arXiv preprint arXiv:2511.17441.
65. Xiao, Shitao; Wang, Yueze; Zhou, Junjie; Yuan, Huaying; Xing, Xingrun; Yan, Ruiran; Li, Chaofan; Wang, Shuting; Huang, Tiejun; Liu, Zheng (2025). Omnigen: Unified image generation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.
66. Xie, Jinheng; Mao, Weijia; Bai, Zechen; Zhang, David Junhao; Wang, Weihao; Lin, Kevin Qinghong; Gu, Yuchao; Chen, Zhijie; Yang, Zhenheng; Shou, Mike Zheng (2025). Show-o: One single transformer to unify multimodal understanding and generation. International Conference on Learning Representations.
67. Ye, Angen; Wang, Boyuan; Ni, Chaojun; Huang, Guan; Zhao, Guosheng; Li, Hao; Li, Hengtao; Li, Jie; Lv, Jindi; Liu, Jingyu; others (2026). GigaWorld-Policy: An Efficient Action-Centered World--Action Model. arXiv preprint arXiv:2603.17240.
68. Ye, Seonghyeon; Ge, Yunhao; Zheng, Kaiyuan; Gao, Shenyuan; Yu, Sihyun; Kurian, George; Indupuru, Suneel; Tan, You Liang; Zhu, Chuning; Xiang, Jiannan; others (2026). World action models are zero-shot policies. arXiv preprint arXiv:2602.15922.
69. Ye, Yang; He, Xianyi; Li, Zongjian; Yuan, Shenghai; Yan, Zhiyuan; Hou, Bohan; Yuan, Li; others (2026). Imgedit: A unified image editing dataset and benchmark. Advances in Neural Information Processing Systems.
70. Yeshwanth, Chandan; Liu, Yueh-Cheng; Nießner, Matthias; Dai, Angela (2023). Scannet++: A high-fidelity dataset of 3d indoor scenes. Proceedings of the IEEE/CVF International Conference on Computer Vision.
71. Yuan, Tianyuan; Dong, Zibin; Liu, Yicheng; Zhao, Hang (2026). Fast-wam: Do world action models need test-time future imagination?. arXiv preprint arXiv:2603.16666.
72. Yue, Yuguang; Salia, Irakli; Hunt, Samuel; Green, Chris; Shi, Wenzhe; Hunt, Jonathan J (2026). Scaling Behavior Cloning Improves Causal Reasoning: An Open Model for Real-Time Video Game Playing. arXiv preprint arXiv:2601.04575.
73. Zhang, Jie; Chen, Xiaoyue; Chen, Anzhe; Lv, Chenxu; Li, Deqing; Zhou, Gengze; Yin, Hang; Yuan, Haoqi; Li, Haoyang; Li, Jiahao; others (2026). Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation. arXiv preprint arXiv:2606.17030.