机器人需要自己的“文本空间”？跨维发布机器人任务原生3D世界模型UNITAS

亿邦动力 2026-10-12 16:41

亿邦动力 2026/10/12 16:41

邦小白快读

全文速览

跨维智能发布UNITAS，这是一种让机器人直接以三维物理空间为学习基础的任务原生3D世界模型，而不是从二维图像推断三维世界。

1. 核心概念：UNITAS把机器人的观察、动作和场景变化统一到同一个三维坐标系中联合建模，让机器人用空间位置关系理解世界，而不是走VLA路线那种模仿式捷径。

2. 效果亮点：仅1.7B参数就在LIBERO测试中取得99.8%平均成功率，超过参数更大的模型；真实机器人整理书籍和装球的平均成功率是85%；场景运动预测误差从0.656厘米降到0.427厘米。

3. 对普通人的价值：机器人在面对初始状态扰动和不同任务时更稳定，比如整理书籍、装球这类日常操作成功率达到80%到90%，这为家用和服务机器人变得更可靠打下基础。

跨维智能以技术路线选择构建品牌差异化，其产品研发、资本背书和生态布局可作为科技品牌建设的参考样本。

1. 产品研发与品牌定位：跨维智能从成立起就确定用AI解决三维物理世界的感知、决策与执行，并搭建仿真引擎DexVerse、数据集Aligned DexWorld和世界模型DexWorldModel的技术闭环，以UNITAS为核心产品打出原生3D世界模型标签，明确反对VLA捷径学习，形成独特技术心智。

2. 融资和市场背书：B轮融资超10亿元，投后估值突破100亿元，投资方包括深创投、前海母基金、蓝思科技、工银资本等，说明资本市场对其技术路线和商业化前景高度认可，也为品牌定价和拓展提供信用支撑。

3. 产品趋势启示：具身智能走向Physical AGI，机器人操作能力大幅提升，品牌商可关注机器人任务原生模型对智能家居、商业服务等产品智能化升级带来的潜在需求，并注意三维空间理解能力将成为下一代智能产品的重要竞争维度。

具身智能赛道正成为资本和产业关注的高增长市场，跨维智能的最新模型为卖家提供了观察新机会和评估风险的具体窗口。

1. 增长市场信号：跨维智能B轮融资超10亿元、估值突破100亿元，资本集中涌入，反映机器人操作智能化进入加速期，卖家可关注由此带动的自动化服务需求。

2. 消费需求变化：UNITAS让机器人在真实环境中完成整理书籍、装球等任务，平均成功率85%，意味着物流分拣、仓储整理、门店服务等场景可能引入更可靠机器人，降低人力成本。

3. 机会与风险：不同来源数据统一为三维点轨迹后，模型鲁棒性明显提升，面对初始状态扰动时成功率从84.8%提升到94.2%；但当前场景流监督主要来自仿真数据，真实环境迁移仍需独立复现验证，卖家在尝试引入机器人时应评估效果稳定性风险。

4. 可学习点：跨维智能用生成式仿真引擎持续生产高真实度数据，并贯通人类示教、真机遥操和仿真合成，这种多源数据协同方式对卖家构建灵活供应链和跨渠道运营也有借鉴意义。

UNITAS为工厂自动化升级提供了具体的技术参考，尤其是机械臂双臂操作精度和面对干扰时的适应性。

1. 生产任务匹配：模型支持机械臂、双臂或人形本体在三维坐标系中完成任务，RoboTwin双臂操作测试中Clean和Randomized设置成功率分别达92.94%和93.04%，可用于复杂装配、协同搬运等工序。

2. 数据获取效率：DexVerse生成式仿真引擎可以产出多样化高真实度场景，Aligned DexWorld数据集实现人类、仿真到真机的时空动作对齐，帮助工厂降低真实操作数据采集的门槛和成本，方便快速验证自动化方案。

3. 数字化生产启示：混合预训练后，面对机器人初始状态扰动时成功率从84.8%提升至94.2%，说明在柔性生产环境中，机器人对初始位置变化和外界干扰的适应能力直接影响产线稳定性，选择自动化方案时应重点考察模型的泛化和抗扰动表现。

机器人行业正在从二维视觉路线转向三维原生世界模型，服务商可以围绕数据统一、仿真迁移和场景预测提供新解决方案。

1. 行业趋势：VLA路线被贾奎视为捷径学习，泛化效率低，UNITAS主张以三维物理空间为原生表示，代表具身智能技术路线的新动向，服务商需跟进这一方向的技术能力。

2. 客户痛点：机器人数据高度分散，仿真、真机、人工示教数据格式各异，不同机械结构难以互通，UNITAS将人手、夹爪和仿真操作端统一为三维点轨迹，解决多源数据联合训练的关键问题。

3. 解决方案参考：跨维智能构建仿真引擎DexVerse、数据集Aligned DexWorld和世界模型DexWorldModel闭环，服务商可以借鉴这套方法为客户提供仿真数据生产、异构数据对齐和模型评测服务。

4. 效果数据：九个数据来源混合预训练后，LIBERO-Plus整体成功率提升2.6个百分点，扰动场景从84.8%提升至94.2%，说明统一数据方案能实际提升模型鲁棒性，可作为服务商向客户展示价值的依据。

跨维智能的UNITAS展示了机器人能力平台的基础架构思路，平台商在整合具身智能生态时可借鉴其统一三维空间的设计方式。

1. 平台需求与问题：具身智能面临数据多源、机械结构异构的难题，需要建立统一的数据表达和训练框架，UNITAS用三维坐标系和点轨迹将不同操作端统一，是平台级数据互通的关键技术路线。

2. 最新做法：跨维智能形成仿真引擎DexVerse生产数据、Aligned DexWorld标准化对齐数据、世界模型DexWorldModel作为核心认知能力的闭环，平台商可沿此路径搭建集数据生成、对齐、训练、评测于一体的基础设施，吸引更多开发者接入。

3. 运营与风控：当前场景流监督主要来自仿真数据，在真实观测数据上仍需独立验证；平台在招商和开放数据接口时，应明确三维场景流数据的质量标准和来源标注，规避仿真到真实迁移不可复现带来的落地风险。

UNITAS的发布反映具身智能研究从二维视觉推断转向三维原生世界模型的新动向，并在数据统一、泛化评测和仿真迁移方面提出新问题。

1. 技术路线创新：模型以三维物理空间为表示基础，将位置关系、运动轨迹和场景变化在统一三维坐标系中联合建模，区别于传统VLA路线，为Physical AGI提供一种原生空间假设。

2. 数据与方法贡献：将人手、夹爪和仿真操作端统一为三维点轨迹，混合预训练九个数据来源，使LIBERO-Plus整体成功率提升2.6个百分点，初始状态扰动下从84.8%提升至94.2%，说明三维一致表达对提升泛化有实证价值。

3. 评测表现：1.7B参数的UNITAS在LIBERO取得99.8%平均成功率，超越参数更大的Fast-WAM、Motus等模型；RoboTwin双臂操作成功率分别为92.94%和93.04%，运动点平均轨迹误差降至0.427厘米。

4. 可复现性及政策启示：场景流监督主要来自仿真数据，但在DROID、RH20T真实观测数据上仅根据夹爪轨迹仍能预测合理场景变化，若被独立复现将为仿真到真机迁移提供可验证路径；产业资本高度密集投入这一方向，提示研究者应关注基础模型统一评测标准和数据共享机制的建设。

返回默认

声明：快读内容全程由AI生成，请注意甄别信息。如您发现问题，请发送邮件至 run@ebrun.com 。

我是品牌商卖家工厂服务商平台商研究者帮我再读一遍。

Quick Summary

Cross-dimensional Intelligence has released UNITAS, a task-native 3D world model that lets robots learn directly from three-dimensional physical space as the foundation, rather than inferring the 3D world from 2D images.

1. Core concept: UNITAS unifies a robot's observations, actions, and scene changes in a shared 3D coordinate system for joint modeling, enabling the robot to understand the world through spatial relationships instead of taking the imitation-style shortcut of the VLA approach.

2. Performance highlights: With only 1.7B parameters, it achieves a 99.8% average success rate on LIBERO, surpassing larger models; on real robots, the average success rate for tidying books and loading balls is 85%; scene motion prediction error drops from 0.656 cm to 0.427 cm.

3. Value for everyday users: Robots become more stable under initial-state perturbations and across different tasks — for example, daily operations like tidying books and loading balls reach 80%–90% success rates, laying the groundwork for more reliable home and service robots.

Cross-dimensional Intelligence builds brand differentiation through its technology route choice, and its product development, capital backing, and ecosystem layout can serve as a reference for tech brand building.

1. Product development and brand positioning: From its inception, Cross-dimensional Intelligence has focused on using AI to solve perception, decision-making, and execution in the 3D physical world. It has built a closed loop consisting of the simulation engine DexVerse, the dataset Aligned DexWorld, and the world model DexWorldModel, using UNITAS as the core product to establish a native 3D world model label and explicitly reject VLA-style shortcut learning, forming a distinctive technical mindset.

2. Financing and market endorsement: The Series B round raised over 1 billion RMB, with a post-investment valuation exceeding 10 billion RMB. Investors include Shenzhen Capital Group, Qianhai FOF, Lens Technology, ICBC Capital, and others, indicating strong capital market recognition of its technology route and commercial prospects, and providing credibility support for brand pricing and expansion.

3. Product trend implications: Embodied intelligence is moving toward Physical AGI, with robot manipulation capabilities improving significantly. Brands can watch for potential demand from task-native robot models for smart home and commercial service product upgrades, and note that 3D spatial understanding will become an important competitive dimension for next-generation intelligent products.

The embodied intelligence track is becoming a high-growth market of interest to capital and industry, and Cross-dimensional Intelligence’s latest model gives sellers a concrete window to observe new opportunities and assess risks.

1. Growth market signals: Cross-dimensional Intelligence’s Series B raised over 1 billion RMB with a valuation exceeding 10 billion RMB, reflecting concentrated capital inflows and signaling that robot operation intelligence is entering an acceleration phase. Sellers can watch for resulting demand for automation services.

2. Changing consumer demand: UNITAS enables robots to complete tasks such as tidying books and loading balls in real environments, with an average success rate of 85%, suggesting that logistics sorting, warehouse organization, and store services may adopt more reliable robots to lower labor costs.

3. Opportunities and risks: After unifying data from different sources into 3D point trajectories, model robustness clearly improves — success under initial-state perturbations rises from 84.8% to 94.2%. However, current scene-flow supervision mainly comes from simulation data, and real-world transfer still needs independent replication and validation. Sellers considering robot adoption should evaluate the risk of performance stability.

4. Learning points: Cross-dimensional Intelligence uses a generative simulation engine to continuously produce high-fidelity data and integrates human demonstration, real-robot teleoperation, and synthetic simulation. This multi-source data coordination approach is also instructive for sellers building flexible supply chains and cross-channel operations.

UNITAS provides a concrete technical reference for factory automation upgrades, especially regarding the precision of dual-arm manipulation and adaptability under interference.

1. Production task fit: The model supports single-arm, dual-arm, or humanoid forms completing tasks in a unified 3D coordinate system. In RoboTwin dual-arm manipulation tests, success rates reach 92.94% and 93.04% under Clean and Randomized settings, making it applicable to complex assembly and collaborative handling processes.

2. Data acquisition efficiency: The DexVerse generative simulation engine can produce diverse, high-fidelity scenes, and the Aligned DexWorld dataset achieves spatiotemporal action alignment across human, simulation, and real-robot data. This helps factories reduce the threshold and cost of collecting real operation data, enabling faster validation of automation solutions.

3. Implications for digital production: After mixed pretraining, the success rate under initial-state perturbations improves from 84.8% to 94.2%, indicating that in flexible production environments, a robot’s adaptability to initial position changes and external disturbances directly affects production line stability. When selecting automation solutions, factories should focus on the model’s generalization and anti-disturbance performance.

The robotics industry is shifting from 2D vision-based approaches to native 3D world models, and service providers can offer new solutions around data unification, simulation transfer, and scene prediction.

1. Industry trend: The VLA route is seen by Jia Kui as a shortcut with low generalization efficiency. UNITAS advocates using 3D physical space as the native representation, representing a new direction in embodied intelligence. Service providers need to follow this technical capability.

2. Customer pain points: Robot data is highly fragmented — simulation, real-robot, and human demonstration data have different formats, and different mechanical structures are hard to interoperate. UNITAS unifies human hands, grippers, and simulated end-effectors into 3D point trajectories, solving the key problem of multi-source data joint training.

3. Solution reference: Cross-dimensional Intelligence has built a closed loop with the simulation engine DexVerse, the dataset Aligned DexWorld, and the world model DexWorldModel. Service providers can adopt this approach to offer clients simulation data production, heterogeneous data alignment, and model evaluation services.

4. Performance data: After mixed pretraining on nine data sources, LIBERO-Plus overall success rate improves by 2.6 percentage points, and perturbed scenarios improve from 84.8% to 94.2%, showing that unified data schemes can genuinely enhance model robustness. This can serve as evidence for service providers to demonstrate value to clients.

Cross-dimensional Intelligence’s UNITAS demonstrates the foundational architecture approach for a robot capability platform, and platform companies can learn from its unified 3D space design when integrating the embodied intelligence ecosystem.

1. Platform needs and problems: Embodied intelligence faces challenges from multi-source data and heterogeneous mechanical structures, requiring a unified data representation and training framework. UNITAS unifies different end-effectors through a 3D coordinate system and point trajectories, representing a key technical route for platform-level data interoperability.

2. Latest practices: Cross-dimensional Intelligence has formed a closed loop in which the DexVerse simulation engine produces data, Aligned DexWorld standardizes and aligns data, and the DexWorldModel serves as the core cognitive capability. Platform companies can follow this path to build infrastructure integrating data generation, alignment, training, and evaluation, attracting more developers to connect.

3. Operations and risk control: Current scene-flow supervision mainly comes from simulation data, which requires independent validation on real observation data. When recruiting developers and opening data interfaces, platforms should define quality standards and source labeling for 3D scene-flow data to avoid landing risks from unreproducible simulation-to-real transfer.

The release of UNITAS reflects a new shift in embodied intelligence research from 2D visual inference to native 3D world models, raising new questions around data unification, generalization evaluation, and simulation transfer.

1. Technical route innovation: The model uses 3D physical space as its representational basis, jointly modeling positional relationships, motion trajectories, and scene changes in a unified 3D coordinate system. It differs from traditional VLA routes and offers a native spatial hypothesis for Physical AGI.

2. Data and methodological contributions: It unifies human hands, grippers, and simulated end-effectors into 3D point trajectories. Mixed pretraining on nine data sources improves LIBERO-Plus overall success by 2.6 percentage points, and under initial-state perturbations it rises from 84.8% to 94.2%, demonstrating empirical value of 3D-consistent representation for improving generalization.

3. Evaluation performance: The 1.7B-parameter UNITAS achieves a 99.8% average success rate on LIBERO, surpassing larger models such as Fast-WAM and Motus; RoboTwin dual-arm success rates are 92.94% and 93.04%, and the average trajectory error for motion points falls to 0.427 cm.

4. Reproducibility and policy implications: Scene-flow supervision mainly comes from simulation data, but on DROID and RH20T real observation data, reasonable scene changes can still be predicted from gripper trajectories alone. If independently replicated, this would provide a verifiable path for simulation-to-real transfer. Intense industrial capital investment in this direction suggests researchers should focus on building unified evaluation standards and data-sharing mechanisms for foundation models.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

【亿邦原创】大语言模型的能力，已经在海量文本训练与规模化扩展中得到验证。它之所以能够持续强化，一个重要基础是：海量来自书籍、网页还是对话的知识，最终都能被统一到“文本”这一输入输出空间中，模型才能在其中完成学习、组合与迁移。

那么，具身智能能否找到一条类似路径？机器人如何从不同来源的经验中学习，让感知、动作与物理世界的变化形成统一理解？走向Physical AGI，它的“原生统一空间”又应该是什么？

10月12日，跨维智能（DexForce）发布机器人任务原生3D世界模型UNITAS（A 3D-Native World Action Model for Embodied Manipulation）。该模型的核心主张是：机器人模型应直接以三维物理空间为表示基础，而非从二维视觉中推断三维世界。

机器人需要什么样的“原生空间”？

跨维智能成立于2021年6月，总部位于深圳南山，创始人贾奎为香港中文大学（深圳）数据科学学院终身教授。据公开资料，贾奎分别于西北工业大学、新加坡国立大学、伦敦大学获得学士、硕士和博士学位，连续五年入选斯坦福全球Top 2%顶尖科学家，研究方向涵盖可泛化深度学习、几何深度学习、三维语义感知与机器人灵巧操作。

贾奎此前在接受采访时表示，其个人相关研究始于2016年，公司2021年成立时确立的方向“用AI解决三维物理世界的感知、决策与执行问题”。2026年6月，跨维智能完成B轮融资，金额超10亿元，投后估值突破100亿元，投资方包括深创投、前海母基金、蓝思科技、工银资本等。

从技术积累来看，UNITAS并非孤立发布。跨维智能此前已推出具身仿真引擎DexVerse®、世界模型DexWorldModel，以及数据对齐技术Dexterity-BEV。贾奎在不同场合多次表达过一个判断：VLA路线的本质是“捷径学习”，希望机器人看到图像、听到指令后直接输出动作，但这一模式在泛化学习上效率极低。他表示，机器人需要的不是“看起来像在干活”的视频，而是“在一个绝对正确的坐标系下，让机械臂、双臂或人形本体真正动起来完成任务”。

这一判断与UNITAS的技术路线一脉相承。UNITAS的核心主张是：三维物理世界本身，就是机器人理解世界的原生表示。模型不再从二维视觉中“翻译”三维信息，而是直接在三维物理空间中学习，观察对应物体的三维位置关系，动作对应操作端的运动轨迹，场景变化对应物体受动作影响后的运动变化。三者被统一到同一个三维坐标系中进行联合建模。

让机器人经验成为同一种“语言”

现实中，机器人数据高度分散，有的来自仿真，有的来自真机采集，有的来自人类操作示范；不同机器人的机械结构与控制方式各异。

UNITAS的解法是将人手、机器人夹爪及仿真操作端统一转换为三维空间中的点轨迹，使不同来源的数据拥有共同表达方式。论文显示，将九个数据来源混合预训练后，在LIBERO-Plus测试中整体成功率提升2.6个百分点，面对机器人初始状态扰动时成功率从84.8%提升至94.2%。

这一“统一表示”思路，与跨维智能近年来在数据层面的布局高度契合。据公开信息，跨维智能自研的DexVerse生成式仿真引擎可高精度还原真实物理规则，通过生成式能力持续产出多样化、高真实度的全新场景与交互任务。此前发布的Aligned DexWorld数据集，则是全球首个贯通人类到仿真到真机的时空动作对齐大规模数据集，融合真人示教、真机遥操、仿真合成等多源样本，通过三维物理级对齐实现异构数据的互通与联合训练。

跨维智能表示，公司已形成“仿真引擎DexVerse与生成式仿真生产数据，Aligned DexWorld标准化对齐数据，世界模型DexWorldModel作为核心认知能力”的技术闭环。

目前，UNITAS已在多项评测中展现出竞争力。仅凭1.7B参数，它在LIBERO上取得99.8%的平均成功率，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RoboTwin 2.0双臂操作测试中，Clean和Randomized设置下成功率分别为92.94%和93.04%。场景运动预测方面，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%。真实机器人测试中，UNITAS在Organize Books任务上成功率达90%，Pack the Ball任务达80%，平均成功率85%。

此外，UNITAS的场景流监督主要来自仿真数据，但在DROID、RH20T等真实世界观测数据上，仅根据夹爪运动轨迹，模型仍能预测合理的场景变化。如果这一结果可被独立复现，将为“仿真数据训练-真实环境迁移”提供一条可验证的路径。

本文首发于亿邦动力官方网站

文章来源：亿邦动力

FAQ回顾

UNITAS是什么？

UNITAS是跨维智能发布的机器人任务原生3D世界模型，核心主张是机器人模型应直接以三维物理空间为表示基础，而非从二维视觉中推断三维世界。它将物体三维位置关系、操作端运动轨迹和场景变化统一到同一三维坐标系中进行联合建模。

跨维智能是一家什么样的公司？

跨维智能成立于2021年6月，总部位于深圳南山，创始人为香港中文大学（深圳）数据科学学院终身教授贾奎。公司专注于用AI解决三维物理世界的感知、决策与执行问题，2026年6月完成超10亿元B轮融资，投后估值突破100亿元，投资方包括深创投、前海母基金、蓝思科技、工银资本等。

UNITAS与VLA路线有什么区别？

贾奎认为VLA路线的本质是“捷径学习”，让机器人看到图像、听到指令后直接输出动作，但泛化学习效率极低。UNITAS则直接以三维物理空间为原生表示，在统一三维坐标系中对物体位置、动作轨迹和场景变化进行联合建模，让来自仿真、真机、人类示教等不同来源的数据拥有共同表达方式。

UNITAS在真实机器人测试中表现如何？

在Organize Books任务上成功率达90%，Pack the Ball任务达80%，平均成功率85%。在LIBERO上仅凭1.7B参数取得99.8%平均成功率，超越Fast-WAM、Motus等更大参数模型。RoboTwin 2.0双臂操作测试中，Clean和Randomized设置下成功率分别为92.94%和93.04%。