大语言模型为什么能scale？一个重要基础是，海量来源各异的知识，被统一进了同一种输入输出空间——文本。输入是文本，输出也是文本，主干模型在这个统一空间中学习、组合与迁移知识。那么，具身智能要获得类似的泛化能力，通往Physical AGI，它的原生统一空间是什么？今天的具身智能，感知、动作与世界变化往往采用不同的表示方式：从输入图像到不同本体的动作表征，再到世界变化，一直在不同的特征空间里来回翻译。同一个推杯子的动作，换个相机视角，像素变化就会不同；相同的像素位移，也未必对应相同的物理距离。模型的能力，被消耗在了翻译而不是理解上。跨维智能最新论文UNITAS给出的回答是：三维物理世界本身。
Why can large language models scale? An important foundation is that massive knowledge from diverse sources is unified into the same input-output space—text. Input is text, output is text, and the backbone model learns, combines, and transfers knowledge in this unified space. So what is the "native unified space" for embodied intelligence to gain similar generalization capabilities on the path to Physical AGI? Today's embodied intelligence often uses different representations for perception, action, and world change: from input images to action representations of different bodies, to world changes, constantly translating back and forth between different feature spaces. The same cup-pushing action looks different in pixel changes from a different camera angle; the same pixel displacement does not necessarily correspond to the same physical distance. The model's capability is consumed in translation, not understanding. DexForce's latest paper UNITAS answers: the 3D physical world itself.
UNITAS将机器人看见什么、自己怎么动、世界因此怎么变化，统一放进同一个三维物理表示空间，用点的位置与随时间变化的轨迹在统一的输入输出三维空间里描述观察、动作和环境响应。就像文本为大模型提供了共同的输入输出空间，UNITAS尝试让Physical AI的输入输出，都生长在统一的三维物理空间里。这不仅是一个技术思路，UNITAS已在多项机器人操作评测中展现出屠榜实力：仅凭1.7B参数，在LIBERO上取得99.8%的平均成功率、在LIBERO-Plus上达到89.1%，并在RoboTwin 2.0双臂操作的Clean与Randomized设置中分别取得92.94%和93.04%的成功率，超越参数规模达6B的Fast-WAM、8B的Motus，以及π0.5、LingBot-VA、PointWorld等模型，在多项基准测试中全面夺冠，进一步在真实机器人操作执行上同样表现突出，任务平均成功率达到85%。
UNITAS unifies what the robot sees, how it moves, and how the world changes as a result into the same 3D physical representation space, using point positions and time-varying trajectories to describe observations, actions, and environmental responses in a unified input-output 3D space. Just as text provides a common input-output space for large models, UNITAS attempts to let Physical AI's inputs and outputs grow in a unified 3D physical space. This is not just a technical idea—UNITAS has demonstrated dominant strength in multiple robot manipulation benchmarks: with only 1.7B parameters, it achieves 99.8% average success on LIBERO, 89.1% on LIBERO-Plus, and 92.94% and 93.04% success on RoboTwin 2.0 dual-arm manipulation in Clean and Randomized settings respectively, surpassing 6B Fast-WAM, 8B Motus, as well as π0.5, LingBot-VA, PointWorld, and others, sweeping multiple benchmarks and also performing prominently in real robot manipulation, with average task success reaching 85%.
看到3D，也许你会好奇，机器人领域已经有不少方法引入了三维信息。UNITAS的3D原生究竟意味着什么？关键区别在于：3D是作为辅助信息存在，还是作为模型理解世界的基础。过去的一些方法可以称为3D感知世界动作模型，它们会引入深度、点云或几何约束，帮助模型获得额外的空间信息，但模型的核心表示仍然建立在二维视觉或其他潜在特征之上。换句话说，模型能够感知三维信息，但三维并不是它理解世界的原生语言。例如，视频世界模型主要通过连续图像预测未来状态，它可以判断画面中的物体如何变化，但物体之间真实的距离、空间关系以及运动规律，需要模型从二维视觉变化中进一步推断。
Seeing "3D," you might wonder—the robotics field already has many methods introducing 3D information. What exactly does UNITAS's 3D-native mean? The key difference: is 3D present as auxiliary information, or as the foundation for the model's understanding of the world? Some past methods can be called 3D-aware World Action Models—they introduce depth, point clouds, or geometric constraints to help the model gain extra spatial information, but the model's core representation remains built on 2D vision or other latent features. In other words, the model can "perceive" 3D information, but 3D is not its native language for understanding the world. For example, video world models mainly predict future states through consecutive images. They can judge how objects in the frame change, but real distances, spatial relationships, and motion laws between objects require the model to further infer from 2D visual changes.
部分3D方法进一步增强了空间理解能力。比如PointWorld能够预测三维场景变化，但动作轨迹由外部输入，模型更接近一个预测环境变化的模拟器，而不是同时理解为什么行动以及行动后世界如何变化的统一模型。包括蚂蚁灵波的LingBot-VA、英伟达的DreamZero，主要沿着二维视觉建模的路线展开，将未来视频预测与机器人动作生成结合。而在UNITAS探索的3D原生思路中，核心变化在于：它不再让模型从二维视觉中推断三维世界，而是直接让模型在三维物理空间中学习世界。简而言之，世界模型直接作用在统一坐标系下的三维Token上，而不通过图片Token作为桥梁。UNITAS的主干模型直接处理三维Token，将三维空间作为机器人观察、行动和理解环境变化的共同表示。
Some 3D methods further enhance spatial understanding. For example, PointWorld can predict 3D scene changes, but action trajectories are externally input, making the model closer to a simulator predicting environmental changes than a unified model that simultaneously understands "why act" and "how the world changes after action." LingBot-VA from Ant Lingbo and NVIDIA's DreamZero mainly follow the 2D visual modeling route, combining future video prediction with robot action generation. In the 3D-native approach UNITAS explores, the core change is: it no longer lets the model infer the 3D world from 2D vision, but directly lets the model learn the world in 3D physical space. Simply put, the world model acts directly on 3D tokens in a unified coordinate system, without using image tokens as a bridge. UNITAS's backbone model directly processes 3D tokens, using 3D space as the common representation for robot observation, action, and understanding environmental changes.
观察对应视觉信息到三维位置，动作对应操作本体上三维点随时间变化的轨迹，环境变化对应观测点在相同时间空间里的运动轨迹。拿起杯子时，夹爪怎样靠近、杯子如何离开桌面、最终移动多远，都可以在同一个具有真实距离尺度的坐标系中表达。模型既能根据任务生成可执行动作，也能预测给定动作带来的场景变化，将机器人控制与世界预测连接起来。据论文作者调研，UNITAS是全球首个机器人任务原生3D世界模型。
Observation corresponds to visual information mapped to 3D positions; action corresponds to the time-varying trajectories of 3D points on the operating body; environmental change corresponds to the motion trajectories of observation points in the same time-space. When picking up a cup, how the gripper approaches, how the cup leaves the table, and how far it ultimately moves can all be expressed in the same coordinate system with real distance scale. The model can both generate executable actions based on tasks and predict scene changes brought by given actions, connecting robot control with world prediction. According to the paper authors' survey, UNITAS is the world's first robot-task-native 3D world model.
常见的策略模型学习的是我应该怎么动；传统世界模型学习的是世界接下来会怎样。这被当作两个异构的问题，用两套目标分别优化。但从物理世界的角度看，机器人伸手拿杯子与杯子被移动，本来就不是两个问题，是整体且因果关系的事件。UNITAS用三维点的轨迹描述这个过程：动作流是人手或机器人夹爪上的点如何随时间移动；场景流是以这些点为条件，杯子等场景中的点如何移动。于是在模型眼里，在任务指令条件下，预测这个物理世界里所有相关的点接下来会怎样运动。论文的消融实验发现：Scene Flow监督之后，动作策略本身的性能还在继续提升——理解世界会发生什么，帮助机器人更好地知道自己该怎么做。
Common policy models learn "how should I move"; traditional world models learn "what will the world do next." These are treated as two heterogeneous problems, optimized with two separate objectives. But from the physical world's perspective, the robot reaching for a cup and the cup being moved are not two problems—they are an integrated, causally related event. UNITAS uses 3D point trajectories to describe this process: Action Flow is how points on the human hand or robot gripper move over time; Scene Flow is, conditioned on these points, how points in the scene like the cup move. So in the model's view, under task instruction conditions, it predicts how all relevant points in this physical world will move next. The paper's ablation experiments found: after Scene Flow supervision, the action policy's own performance continues to improve—"understanding what will happen in the world" helps the robot "better know what it should do."
要实现这样的联合建模，并在各种各样的本体、数据集上进行联合学习，首先需要解决深度缺失和轨迹时间、频率不对齐的问题。UNITAS提出世界对齐的3D位置嵌入模块，为观测、动作流和场景流引入统一三维坐标编码；针对深度缺失数据，模型可在视锥范围内学习补全深度，从而支持RGB-D与RGB数据的训练和推理。针对不同数据集存在轨迹时长和频率差异的情况，UNITAS提出了物理时间轨迹编码器模块，该模块将不同参数的轨迹按照轨迹上逐点的实际物理时间进行点时间编码，使用FiLM调制的方式将一条轨迹编码到一个统一的token里，使不同轨迹在主干模型中对齐到同一物理时间特征空间。解决了以上数据异构之后，UNITAS的主干模型Point Dynamics DiT就可以在统一的特征空间上进行学习泛化。
To achieve such joint modeling and joint learning across various bodies and datasets, the first problems to solve are depth missing and trajectory time/frequency misalignment. UNITAS proposes a World-Aligned 3D Position Embedding module, introducing unified 3D coordinate encoding for observations, action flows, and scene flows; for depth-missing data, the model can learn to complete depth within the view frustum, supporting both RGB-D and RGB data training and inference. For differences in trajectory duration and frequency across datasets, UNITAS proposes a Physical Time Trajectory Tokenizer module, which encodes trajectories with different parameters by point-time encoding based on the actual physical time of each point on the trajectory, using FiLM modulation to encode a trajectory into a unified token, aligning different trajectories to the same physical time feature space in the backbone model. After resolving the above data heterogeneity, UNITAS's backbone model Point Dynamics DiT can learn and generalize in a unified feature space.
人用手拿杯子，机器人用夹爪拿杯子，身体结构和控制方式都不同。但如果把目光放在运动本身，两者都可以用操作端上的点经过了哪些位置，来描述接近、抓取和移动的过程。仿真中的夹爪、真实机器人的夹爪，以及人类双手，虽然来自不同的数据源，却都被表达为三维点的运动轨迹。对UNITAS来说，它们是同一种语言。UNITAS以这些三维点轨迹为共同表示，让人手、不同机器人以及仿真中的操作数据能够一起参与训练。论文将九个数据来源用于混合预训练，再针对具体任务微调，观察到了操作性能的提升。在LIBERO-Plus上，混合预训练使整体成功率提高了2.6个百分点；其中，机器人初始状态扰动下的成功率从84.8%提升至94.2%。
A human picks up a cup with a hand; a robot picks up a cup with a gripper—body structures and control methods differ. But if you look at the motion itself, both can be described by which positions points on the operating end pass through, describing the process of approaching, grasping, and moving. Simulated grippers, real robot grippers, and human hands, though from different data sources, are all expressed as 3D point motion trajectories. To UNITAS, they are the same language. UNITAS uses these 3D point trajectories as a common representation, allowing human hands, different robots, and simulated manipulation data to participate in training together. The paper uses nine data sources for mixed pretraining, then fine-tunes for specific tasks, observing improved manipulation performance. On LIBERO-Plus, mixed pretraining raised overall success by 2.6 percentage points; among them, success under robot initial state perturbation rose from 84.8% to 94.2%.
UNITAS以17亿参数，在LIBERO、LIBERO-Plus、RoboTwin 2.0和VLABench等多项机器人操作基准中全面夺冠。LIBERO平均成功率达到99.8%，长程任务子集LIBERO-Long也达到99.4%。RoboTwin 2.0双臂操作成功率分别对应Clean和Randomized设置，超过Fast-WAM（6B）和Motus（8B）等参数规模更大的视频世界动作模型。LIBERO-Plus整体成功率89.1%，面对相机视角、机器人初始状态等七类扰动，模型依然保持较强的任务完成能力。场景运动预测误差降低约35%，在RoboTwin仿真测试中，运动点平均轨迹误差由PointWorld的0.656厘米降至0.427厘米，误差不超过1厘米的预测准确率达到92.31%，四项场景预测指标全部领先。
With 1.7B parameters, UNITAS sweeps multiple robot manipulation benchmarks including LIBERO, LIBERO-Plus, RoboTwin 2.0, and VLABench. LIBERO average success reaches 99.8%, and the long-horizon subset LIBERO-Long also reaches 99.4%. RoboTwin 2.0 dual-arm manipulation success corresponds to Clean and Randomized settings, surpassing larger video world action models like Fast-WAM (6B) and Motus (8B). LIBERO-Plus overall success is 89.1%; facing seven types of perturbations including camera viewpoint and robot initial state, the model maintains strong task completion. Scene motion prediction error drops about 35%—in RoboTwin simulation tests, average trajectory error for moving points fell from PointWorld's 0.656 cm to 0.427 cm, with 92.31% prediction accuracy within 1 cm error, leading all four scene prediction metrics.
真实机器人测试则进一步检验UNITAS在实际物理环境中的执行效果。在论文报告的真实任务中，UNITAS平均成功率达到85%，其中在Organize Books上达到90%成功率，在Pack the Ball上达到80%。综合来看，UNITAS的成绩并非某一项指标的突破，而是体现了标准操作、双臂协作、扰动适应、场景运动预测和真实机器人执行等多方面综合实力。更值得关注的是，UNITAS仅有1.7B参数，却在多个指标和表现上均超过参数规模更大的对比模型。
Real robot tests further examine UNITAS's execution in actual physical environments. In real tasks reported in the paper, UNITAS average success reaches 85%, including 90% on Organize Books and 80% on Pack the Ball. Overall, UNITAS's results are not a breakthrough in a single metric but reflect comprehensive strength across standard manipulation, dual-arm collaboration, perturbation adaptation, scene motion prediction, and real robot execution. More notably, UNITAS has only 1.7B parameters yet surpasses larger comparison models on multiple metrics and performances.
把视觉3D位置编码、动作流监督、动作流历史token这三项3D原生设计全部去掉，模型即退化为常规VLA基线，整体成功率从86.47%跌到67.58%——近19个百分点的差距，全部由表示方式本身贡献。论文在LIBERO-Plus上逐项验证：3D位置编码去掉后相机视角扰动下成功率从79.11%跌至66.73%，对齐的三维信息是跨视角泛化的关键。动作流监督去掉后整体成功率从85.74%降至78.89%，对齐输入观测和输出动作空间是学习可泛化机器人动作控制的核心。动作流历史token去掉后机器人初始状态扰动下成功率骤降19.87个百分点，历史动作上下文增强了模型鲁棒性。场景流监督在上述三项之上叠加后整体成功率从85.74%升至86.47%，看懂世界会怎样反过来让自己该怎么动更准。
Removing all three 3D-native designs—visual 3D position encoding, action flow supervision, and action flow history tokens—degrades the model to a conventional VLA baseline, with overall success dropping from 86.47% to 67.58%—nearly 19 percentage points, all contributed by the representation itself. The paper verifies item by item on LIBERO-Plus: removing 3D position encoding drops success under camera viewpoint perturbation from 79.11% to 66.73%, with aligned 3D information key to cross-viewpoint generalization. Removing action flow supervision drops overall success from 85.74% to 78.89%; aligning input observation and output action spaces is core to learning generalizable robot action control. Removing action flow history tokens drops success under robot initial state perturbation by 19.87 percentage points; historical action context enhances model robustness. Adding scene flow supervision on top of the above three raises overall success from 85.74% to 86.47%—"seeing what the world will do" in turn makes "knowing how to move" more accurate.
跨维智能的另一项研究TaskAnchor，关注的是任务状态混淆。它从执行历史中提取上下文，一方面通过语言输入告诉模型任务进行到了哪个阶段，另一方面为视觉特征补充历史线索，帮助视觉语言动作模型选择下一步动作。这一轻量模块可以接入不同基础模型，无需改变原有动作生成机制。接入π0.5后，TaskAnchor在RMBench上将平均成功率从14%提升至79%，在RoboMemArena上将累积成功率从38.7%提升至64.7%，增加的参数约为2700万，单次策略调用的额外延迟约为2毫秒。团队表示，未来计划将TaskAnchor的相关能力整合进UNITAS，让模型在理解空间交互的同时，也能利用历史上下文完成更长的任务。
Another DexForce study, TaskAnchor, focuses on task state confusion. It extracts context from execution history, using language input to tell the model which stage the task has reached, and supplementing visual features with historical clues to help the vision-language-action model choose the next action. This lightweight module can plug into different foundation models without changing the original action generation mechanism. After plugging into π0.5, TaskAnchor raised average success on RMBench from 14% to 79% and cumulative success on RoboMemArena from 38.7% to 64.7%, adding about 27 million parameters with about 2 milliseconds extra latency per policy call. The team says it plans to integrate TaskAnchor's capabilities into UNITAS, letting the model understand spatial interaction while also using historical context to complete longer tasks.
大模型找到了知识世界的统一语言，Physical AI还没有。这是具身智能绕不开的一道坎：没有统一表示，数据就难以汇聚；数据难以汇聚，Scaling就无从谈起。至于世界模型这个词，今天谁都能讲。更值得追问的是Physical AI的原生表示空间应该长什么样——UNITAS的回答是：物理世界本身。语言模型已经证明：找到统一表示，Scaling就会随之而来。现在，轮到物理世界了。
Large models found the unified language of the knowledge world; Physical AI has not yet. This is an unavoidable hurdle for embodied intelligence: without unified representation, data is hard to aggregate; without data aggregation, Scaling cannot begin. As for the term "world model," anyone can talk about it today. More worth asking is what Physical AI's native representation space should look like—UNITAS's answer: the physical world itself. Language models have already proven: find the unified representation, and Scaling follows. Now it is the physical world's turn.
