广告
加载中

自变量发布 X2Real 仿真评测基准 具身模型进入“自动化评测”时代

龚作仁 2026-10-19 17:47
龚作仁 2026/10/19 17:47

邦小白快读

EN
全文速览

这篇文章介绍了具身模型评测领域的新工具X2Real,核心干货是一套能自动、可信地评估机器人模型能力的仿真评测系统。

1. 它解决了真机评测贵、慢且仿真评测不真实、易被刷分的问题,通过仿真标定让虚拟机器人与真实硬件高度一致,测试显示仿真评测与真机评测的成功率相关性达0.96,进度分相关性达0.99,说明仿真结果能可靠反映真实能力。

2. 它设计了44个仿真任务,涵盖推理和操作两大类能力,包括泛化、视觉理解、语言理解、记忆、精细操作、双臂协同、工具使用、动态物体操作等,还专门设置堆叠积木、解决汉诺塔等挑战任务,为模型进步留出空间。

3. 它的核心创新是自动化:用大模型智能体驱动任务设计和回放分析,既能持续生成新题库,又能自动诊断模型卡在哪个步骤,形成了出题、判分、补题的自动化流水线。同时通过统一评测协议,让不同接口的模型都能接入评测,开发者无需本地搭建仿真环境,降低了使用门槛。

从品牌商视角看,这篇文章的核心干货在于具身智能产品能力评估基础设施的建立,有助于品牌对机器人产品进行更客观的选型和质量控制。

1. 品牌在采购或开发具身智能产品时,过去难以判断不同模型在真实场景中的表现差异。X2Real通过仿真标定和统一评测协议,让不同接口的模型可以在同一规则下比较,等于提供了一份可量化的产品能力体检报告,便于品牌筛选合作伙伴和产品方案。

2. 品牌关注产品研发效率和迭代质量。X2Real将评测变成自动化流程,能自动定位模型能力薄弱环节并生成针对性新任务,相当于让产品研发过程中的缺陷发现和改进更加系统化,减少人工评测成本,缩短迭代周期,有助于品牌持续优化机器人产品的用户体验和稳定性。

3. 品牌也关注消费者趋势和用户行为。文章透露具身模型正在走向精细化能力评价,包括视觉理解、语言理解、记忆、精细操作等维度,这提示未来具身智能产品竞争将从单一功能转向综合能力,品牌在定义产品卖点和传播时,可以引用多维度的评测数据来建立信任。

这篇文章对卖家而言,主要干货是具身智能评测基础设施带来的市场机会和产品迭代路径。

1. 市场变化与机会:仿真评测成为行业基础设施,意味着具身智能模型研发门槛降低。卖家如果提供相关机器人产品或服务,可以借助X2Real这样工具快速验证模型能力,减少真机测试成本和时间,从而更快响应客户需求,拓展业务。

2. 可学习点:文章展示了如何通过标准化任务设计、自动化诊断和无限题库来持续优化产品。卖家可以从中学到一种以评测驱动研发的方法:将客户需求拆解成可量化的任务,用自动化流程持续发现短板并改进,提升服务质量。

3. 合作方式与风险提示:X2Real统一了评测协议,开发者只需接入统一模型接口即可在线评测。这提示卖家可以基于该平台向客户提供模型接入、结果解读、性能优化等增值服务。同时需注意,不同模型接口差异较大,平台通过适配层统一翻译,卖家在对接客户时应当关注模型兼容性和评测环境与真实环境的差异,避免过度依赖仿真结果。

4. 政策与扶持:文章提到X2Real使用英伟达Isaac Lab-Arena支撑的Mana仿真生态系统,这一生态方向可能获得更多技术和资源支持,卖家可关注相关生态合作机会。

对于工厂来说,这篇文章的干货集中在机器人产品设计与生产需求的变化,以及数字化转型的启示。

1. 产品设计和生产需求:文章揭示具身模型的能力评价正在向精细化和自动化发展,涉及精细操作、双臂协同、工具使用、动态物体操作等任务。工厂在设计和生产机器人硬件时,需要关注这些能力对应的机械结构、关节响应速度、视觉传感和物理属性标定,确保硬件能与仿真环境高度对齐,满足自动化评测的要求。

2. 仿真与生产联动:仿真评测与真机评测相关性很高,意味着工厂可以在设计阶段利用仿真环境提前测试机械臂或移动平台的控制参数,减少物理样机试错成本。文章提到支持单臂Franka、双臂ArtiXon Arm-6A、ARX R5和移动操作平台Quanta X1等多种本体,说明工厂产品如果符合这些主流标准,更容易接入评测生态。

3. 数字化和电商启示:X2Real的自动化评测系统展示了工业数字化的一种方向,即把产品测试和迭代流程变成标准化、自动化的流水线。工厂可以类比这种思路,将生产质检、产品调试等环节拆解为可自动判定和持续优化的模块,利用数据和智能体提升效率。此外,工厂如果生产机器人配件或整机,可以主动向仿真评测平台提供硬件数据,参与生态共建,获取更多订单机会。

从服务商角度看,这篇文章的核心干货是具身智能评测领域的行业趋势、技术架构以及可切入的解决方案。

1. 行业趋势:具身智能模型评测正在从人力密集、环境随意的模式转向自动化、标准化和可进化模式。服务商如果提供模型评估、测试工具或机器人集成服务,X2Real的出现意味着行业对可信评测基础设施的需求正在爆发,是一个明确的业务方向。

2. 新技术和应用:文章详细介绍了仿真标定技术,包括控制参数标定、视觉特征和物理属性标定,以及毫米级轨迹一致性。服务商可以借鉴这套方法,为不同机器人本体做仿真校准服务,或者提供仿真环境搭建和定制化任务设计。

3. 客户痛点和解决方案:客户的核心痛点是真机评测贵且慢、仿真评测失真且易被刷分。X2Real通过三大原则解决:与真实硬件对齐、多维度任务设计、自动化评测与诊断。服务商可以利用这些原则,帮助客户建立自身的评测体系,尤其是针对特定行业场景定制任务图和诊断逻辑。

4. 生态延伸:文章提到智能体驱动任务设计和回放分析,以及统一评测协议,大幅降低评测门槛。服务商可以围绕这个平台,做模型接入适配、训练数据集服务、失败轨迹分析报告等增值服务,形成新的商业模式。

这篇文章对平台商的核心干货在于如何构建和运营具身智能评测平台,以及平台在商业化和管理上的关键设计。

1. 平台需求与问题:文章指出现有评测体系存在真机贵、仿真失真、题库易被刷分等问题,这为平台商提供了建立评测平台的需求依据。平台要解决的核心问题是让模型能力可度量、有可比性,这就需要统一评测协议、统一接口和标准化任务。

2. 平台最新做法:X2Real展示了完整的平台架构。它采用三个设计原则:可信性通过仿真标定实现,有效性通过44个任务覆盖多维能力,客观性通过任务多维随机化、严格分离训练与评测配置、持续生成新考题实现。平台商可以借鉴这套原则来设计自己的评测产品。

3. 运营管理:平台通过智能体驱动两个循环(任务设计循环和回放分析循环)实现自动化运营,同时利用统一评测协议让不同模型接入。平台商可以从中学习到如何降低开发者接入门槛,提供在线评测、判分和轨迹回放服务,形成标准化、可规模化的运营模式。

4. 风向规避:平台需要防止模型背题刷分。X2Real的做法是任务拆分为外观条件、任务设置、具身条件三个维度随机化,并提供近300小时隔离训练数据,确保训练与考试环境完全分离。平台商在制定规则时,应重点关注评测数据隔离和随机化策略,维护平台公信力。

从研究者角度看,这篇文章的干货提供了产业新动向、评测体系设计原则和可借鉴的自动化方法。

1. 产业新动向:具身智能正在进入自动化评测时代。研究者关注的关键问题——如何衡量模型进步、如何发现缺陷——在文中有了具体答案:通过仿真标定提升评测可信度,通过智能体自动化评测降低人力成本,形成持续进化的基础评测量规。相关性数据0.96和0.99为仿真替代真机评测提供了实证依据。

2. 政策法规建议与启示:文章隐含一个观点:评测是行业基础设施,需要建立可信、公平的规则。这对监管机构或行业组织有启示意义:可以推动制定统一的仿真评测标准,规范训练数据与测试数据的隔离要求,防止数据泄露和刷分行为,促进行业健康发展。

3. 商业模式与研究方法:研究者可以研究X2Real的三原则(可信性、有效性、客观性)如何落地。它通过硬件对齐的仿真标定、任务图拆解、节点自动判定、智能体生成考题等机制,构建了闭环改进循环。这种将大模型智能体用于任务设计和诊断的模式,为AI研发流程自动化提供了新思路:人类提供高层意图,智能体执行出题、判分、补题,可能成为未来AI系统开发的重要范式。

4. 研究空间:文章提到44个任务中大部分具身模型仍很难完成,说明当前模型能力上限有限,为研究者提供了明确的挑战清单和基准参照。

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

This article introduces X2Real, a new tool in the field of embodied model evaluation. The core value is an automated, credible simulation-based evaluation system for robot model capabilities.

1. It solves the problems of expensive and slow real-robot evaluation, as well as unrealistic simulation evaluation that is vulnerable to benchmark gaming. Through simulation calibration, virtual robots are made highly consistent with real hardware. Tests show that the success rate correlation between simulation evaluation and real-robot evaluation reaches 0.96, and the progress score correlation reaches 0.99, indicating that simulation results can reliably reflect real-world capabilities.

2. It designs 44 simulation tasks covering two major capability categories—reasoning and manipulation—including generalization, visual understanding, language understanding, memory, fine manipulation, dual-arm coordination, tool use, and dynamic object manipulation. It also includes challenge tasks such as stacking blocks and solving the Tower of Hanoi, leaving room for model improvement.

3. Its core innovation is automation: large-model agents drive task design and replay analysis, continuously generating new task pools and automatically diagnosing where a model gets stuck, forming an automated pipeline for question generation, scoring, and supplementation. Through a unified evaluation protocol, models with different interfaces can all be evaluated, and developers do not need to set up simulation environments locally, lowering the barrier to entry.

From a brand perspective, the core value of this article is the establishment of infrastructure for assessing embodied intelligence product capabilities, helping brands make more objective product selection and quality control decisions for robot products.

1. When procuring or developing embodied intelligence products, brands previously found it difficult to judge performance differences among models in real-world scenarios. X2Real, through simulation calibration and a unified evaluation protocol, enables models with different interfaces to be compared under the same rules—providing a quantifiable product capability health check that helps brands screen partners and product solutions.

2. Brands care about R&D efficiency and iteration quality. X2Real turns evaluation into an automated process that can automatically identify weak points in model capabilities and generate targeted new tasks. This makes defect discovery and improvement in product development more systematic, reduces manual evaluation costs, shortens iteration cycles, and helps brands continuously optimize the user experience and stability of robot products.

3. Brands also track consumer trends and user behavior. The article reveals that embodied models are moving toward fine-grained capability evaluation, including visual understanding, language understanding, memory, and fine manipulation. This indicates that future competition in embodied intelligence products will shift from single-function performance to comprehensive capability. When defining product positioning and messaging, brands can cite multi-dimensional evaluation data to build trust.

For sellers, the main value of this article lies in the market opportunities and product iteration pathways created by the embodied intelligence evaluation infrastructure.

1. Market change and opportunity: Simulation evaluation becoming an industry infrastructure lowers the barrier to embodied intelligence model R&D. Sellers offering related robot products or services can use tools like X2Real to quickly validate model capabilities, reducing real-robot testing costs and time, thus responding to customer needs faster and expanding their business.

2. Key takeaways: The article demonstrates how standardized task design, automated diagnosis, and an unlimited task pool can enable continuous product improvement. Sellers can learn an evaluation-driven R&D method: decompose customer needs into quantifiable tasks, and use automated processes to continuously discover weaknesses and improve service quality.

3. Cooperation model and risk warnings: X2Real unifies the evaluation protocol, so developers only need to connect through a unified model interface for online evaluation. Sellers can build value-added services on this platform, such as model integration, result interpretation, and performance optimization. However, different model interfaces vary significantly; the platform uses an adaptation layer for unified translation. When serving customers, sellers should pay attention to model compatibility and the gap between evaluation and real-world environments, avoiding over-reliance on simulation results.

4. Policy and support: The article mentions that X2Real uses the Mana simulation ecosystem supported by NVIDIA Isaac Lab-Arena. This ecosystem direction may gain more technical and resource support, and sellers should keep an eye on related ecosystem cooperation opportunities.

For factories, this article focuses on changes in robot product design and production requirements, as well as implications for digital transformation.

1. Product design and production requirements: The article shows that capability evaluation for embodied models is moving toward fine-grained and automated approaches, involving tasks such as fine manipulation, dual-arm coordination, tool use, and dynamic object manipulation. When designing and producing robot hardware, factories need to consider the corresponding mechanical structures, joint response speeds, visual sensing, and physical property calibration, ensuring that hardware is highly aligned with simulation environments to satisfy automated evaluation requirements.

2. Simulation-production linkage: The high correlation between simulation and real-robot evaluation means factories can use simulation environments during the design phase to test control parameters for robotic arms or mobile platforms in advance, reducing the trial-and-error costs of physical prototypes. The article mentions support for multiple embodiments—including the single-arm Franka, dual-arm ArtiXon Arm-6A, ARX R5, and the mobile manipulation platform Quanta X1—indicating that products conforming to these mainstream standards can more easily integrate into the evaluation ecosystem.

3. Digital transformation and e-commerce implications: X2Real's automated evaluation system demonstrates a direction for industrial digitalization—turning product testing and iteration processes into standardized, automated pipelines. Factories can apply this mindset by decomposing production inspection and product debugging into modules that are automatically judged and continuously optimized, using data and agents to improve efficiency. In addition, if factories produce robot accessories or complete machines, they can proactively provide hardware data to simulation evaluation platforms, participate in ecosystem co-building, and create more order opportunities.

From a service provider perspective, the core value of this article is found in the industry trends, technical architecture, and actionable solutions in embodied intelligence evaluation.

1. Industry trend: Evaluation of embodied intelligence models is shifting from labor-intensive and ad-hoc settings to automated, standardized, and evolvable models. For service providers offering model evaluation, testing tools, or robot integration services, X2Real signals a surge in demand for credible evaluation infrastructure—a clear business direction.

2. New technologies and applications: The article details simulation calibration technologies, including control parameter calibration, visual feature and physical property calibration, and millimeter-level trajectory consistency. Service providers can adopt these methods to offer simulation calibration services for different robot embodiments, or provide simulation environment setup and customized task design.

3. Customer pain points and solutions: The core pain points for customers are that real-robot evaluation is expensive and slow, while simulation evaluation is distorted and susceptible to benchmark gaming. X2Real solves this with three principles: alignment with real hardware, multi-dimensional task design, and automated evaluation with diagnosis. Service providers can leverage these principles to help customers build their own evaluation systems, especially customizing task graphs and diagnostic logic for specific industry scenarios.

4. Ecosystem extension: The article mentions agent-driven task design and replay analysis, along with a unified evaluation protocol, significantly lowering evaluation barriers. Service providers can build value-added services around this platform—such as model integration adaptation, training dataset services, and failure trajectory analysis reports—forming new business models.

This article's key value for platform operators lies in how to build and operate an embodied intelligence evaluation platform, as well as the critical design considerations for commercialization and management.

1. Platform demand and problems: The article identifies problems in existing evaluation systems—expensive real-robot testing, distorted simulation, and task pools that are easily gamed—providing the demand basis for platform operators to build such a platform. The core issue a platform must solve is making model capabilities measurable and comparable, requiring unified evaluation protocols, unified interfaces, and standardized tasks.

2. Latest platform practices: X2Real demonstrates a complete platform architecture. It adopts three design principles: credibility through simulation calibration, validity through 44 tasks covering multi-dimensional capabilities, and objectivity through multi-dimensional task randomization, strict separation of training and evaluation configurations, and continuous generation of new test questions. Platform operators can borrow these principles to design their own evaluation products.

3. Operations management: The platform uses agents to drive two loops—the task design loop and the replay analysis loop—for automated operations, while also leveraging a unified evaluation protocol to connect different models. Platform operators can learn how to lower developer onboarding barriers and provide online evaluation, scoring, and trajectory replay services, forming a standardized, scalable operating model.

4. Risk mitigation: Platforms need to prevent models from memorizing tasks and gaming the benchmark. X2Real's approach is to randomize tasks across three dimensions—appearance conditions, task settings, and embodiment conditions—and provide nearly 300 hours of quarantined training data to ensure complete separation between training and evaluation. When setting rules, platform operators should prioritize evaluation data isolation and randomization strategies to maintain platform credibility.

From a researcher's perspective, this article delivers industry trends, evaluation system design principles, and adaptable automation methods.

1. Industry trend: Embodied intelligence is entering the era of automated evaluation. The key questions researchers care about—how to measure model progress and how to identify defects—find concrete answers in the article: simulation calibration improves evaluation credibility, agent-driven automated evaluation reduces human costs, and together they establish an evolving baseline metric for evaluation. The correlation values of 0.96 and 0.99 provide empirical evidence for replacing real-robot evaluation with simulation.

2. Policy and regulatory implications: The article implicitly argues that evaluation is industry infrastructure requiring credible, fair rules. This has implications for regulators and industry bodies: they can advance unified simulation evaluation standards, standardize the separation of training and test data, prevent data leakage and benchmark gaming, and promote healthy industry development.

3. Business model and research methods: Researchers can study how X2Real's three principles (credibility, validity, objectivity) are applied. Through hardware-aligned simulation calibration, task graph decomposition, automatic node scoring, and agent-generated test questions, X2Real builds a closed-loop improvement cycle. This use of large-model agents for task design and diagnosis offers a new approach to automating AI R&D: humans provide high-level intent, while agents carry out question generation, scoring, and supplementation—potentially an important paradigm for future AI system development.

4. Research space: The article notes that most embodied models still struggle to complete most of the 44 tasks, indicating the current capability ceiling of models and giving researchers a clear challenge list and benchmark reference.

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.

当数据规模与模型能力持续增长,该如何衡量模型的能力进步,又该如何发现模型缺陷,加以改进?衡量模型能力需要一把客观可信的标尺,评测体系正是行业必不可少的基础设施。

然而目前的模型评测存在诸多问题:真机评测成本昂贵、耗时缓慢,难以满足研发高频次、大规模的测试需要。仿真评测虽然能大幅提升测试效率,但又面临仿真不够“真实”、题库不够“丰富”的问题,很容易被模型“摸透规律”刷分,评测能力难以伴随模型能力共同成长。

今日,自变量发布X2Real——一套可进化的仿真评测基准,旨在评估具身模型在真实物理世界中的性能。它提出三个设计原则:可信性(Faithfulness)、有效性(Effectiveness)和客观性(Fairness),分别通过与真实硬件对齐的仿真、多维度的任务设计,以及自动化的评测与诊断流程,让仿真评测成为模型性能的可信标尺。

更重要的是,X2Real是一套能够伴随模型成长不断进化的评测体系,其评测能力不会随着模型能力提升而过时。经过测试,X2Real仿真评测与真机评测的任务成功率相关性达到0.96、进度分相关性达到0.99,居于行业最前列水平。它真正实现了“让模型能力可度量,让研发进步可积累”,成为具身行业的关键基础设施。

仿真忠实还原真机表现,任务成功率相关性高达0.96

X2Real的第一个原则是可信性,即忠实反应模型在物理世界的真实能力。这要求X2Real能够以极高的仿真度模拟机器人在现实物理世界的动作。

但仿真与现实之间存在天然差异:关节的响应快慢、画面的质感、物理的重量与摩擦力,这些细节的仿真失真都会影响模型策略的表现。因此,要让仿真评测成为可信赖的“考场”,第一步就是将虚拟机器人调试到与真实机器人表现一致。

X2Real采用与真实硬件对齐的仿真标定,去校准仿真世界里的机器人视觉、物理和控制特性。在控制参数方面,自变量团队先记录下真实机械臂的运动轨迹,再将同样的控制指令输入仿真系统,比较两者的运动轨迹,反复调整关节的响应力度,直到两者一致。经测试,X2Real能够实现毫米级的轨迹一致性。

在视觉特征和物理属性上,X2Real也都对偏差做了细致标定,确保模型在仿真中看到的画面、感受到的接触反馈,都与现实一致。这些标定覆盖了多种主流机器人本体,包括单臂Franka、双臂ArtiXon Arm-6A、ARX R5和移动操作平台Quanta X1等。

测试显示,X2Real经过标定能够忠实反映模型能力:在通用操作任务上,仿真评测与真机评测的任务成功率相关性达到0.96、进度分相关性达到0.99,仿真评测足以成为模型能力的可信量化依据。

44个任务“海量题库”,全面考核推理与操作能力

X2Real的第二个原则是有效性,即综合全面地考察模型的能力和短板。它从任务的多样性入手,将通用操作能力分为推理与操作两个类别:推理即正确理解任务指令和视觉输入,作出正确的动作决策;操作即正确执行对应的动作,达到预期的结果。

X2Real共设计了10个类别的44个仿真任务,能够覆盖推理与操作中的多类关键能力。具体来说,4类推理能力包括泛化、视觉理解、语言理解、记忆;5类操作任务包括精细操作、双臂协同、工具使用、动态物体操作、移动操作。用于评测的任务由其中的一系列动作组合而成,比如拉开抽屉放入药瓶、将眼镜挂上木架子等。

此外,X2Real还专门设计了一类挑战任务,包括堆叠积木、按指定颜色序列按键、解决汉诺塔,分别考核模型的运动精度、长时记忆和推理规划能力。目前,这些任务对大部分具身模型都有很大的挑战,为模型未来进步留下充足的评测空间。

严格隔离训练与考试配置,杜绝模型“背题”刷分

X2Real的第三个原则是客观性:既减少模型靠“背题”获得高分的可能性,也保证各种模型在同一规则下可以比较。围绕客观性要求,X2Real做出三点设计:

一是任务的多维随机化。X2Real将任务拆分为三个维度:外观条件,包括背景、光照、桌面材质纹理;任务设置,包括任务资产、物体摆放、干扰物布局、语言指令;具身条件,包括夹爪类型、相机参数、桌面高度。这些维度分别随机化,确保模型无法靠“死记硬背”评测环境获得好成绩。

二是严格分离训练与评测配置。模型在测试前,都需要在X2Real提供的近300小时数据集上进行后训练。X2Real确保这些训练数据在统一的绿幕摄影棚中采集,与评测的考试环境完全隔离,能够检验模型对新配置的泛化能力,杜绝“考前泄题”的风险。

三是持续生成新考题。X2Real使用由英伟达Isaac Lab-Arena支撑的Mana仿真生态系统,将场景、机器人、任务步骤和环境都打包为可复用的模块,能够持续生成新的评测组合,相当于拥有无限的题库。

智能体驱动任务设计与评分,让评测成为“自动化流水线”

X2Real能持续扩充考题,靠的是一套让大模型智能体深度参与的工作系统。

智能体同时驱动两个循环:一是任务设计的循环,持续编写新任务、新场景,扩充评测体系;二是回放分析的循环:每次评测后自动读取失败节点与轨迹,定位能力薄弱的环节,生成新的针对性任务。

在任务设计循环中,智能体会驱动仿真软件搭建任务。通常,使用仿真软件需要处理复杂的底层接口,智能体难以直接使用。X2Real则不直接使用这些接口,而是给智能体提供一套标准化的“可选项”:选择场景、摆放物品、定义任务步骤、选择机器人形态。配置完成后,系统还会自动检测,剔除掉不合理的配置。

在回放分析的循环中,X2Real会将每个任务拆解为任务图:子任务按照先后关系连接成流程,每个节点的完成情况,都由物体姿态、接触关系、抓取状态等自动判定。即使模型未能完成最终任务目标,开发者也能知道它卡在哪一步,为改进模型提供清晰的诊断意见。

此外,智能体还会根据每次评测的节点记录,定位出模型能力的薄弱环节,针对性地生成新考题。可以说,在X2Real系统中,人类只提供高层意图,具体迭代都由智能体执行。这真正将“出题、判分、补题”的一整套评测流程,从人力密集工作变为可自动运行的流水线。

得关注的是,在传统评测中,由于不同模型拥有不同的接口,包括输入格式、动作表示、推理节奏等,难以直接互相比较。X2Real通过统一评测协议(Policy Space)将这些差异由适配层统一翻译,相当于给所有模型发了同样的“答题卡”。

通过统一评测协议,X2Real也大幅降低了评测门槛:开发者不需在本地搭建仿真环境,只需要接入统一模型接口、在本地运行模型推理服务,平台就能在线完成评测判分和轨迹回放工作。这使仿真评测变成开发者触手可及的行业基础设施。

仿真评测与模型能力共同成长

具身智能的持续迭代,需要一套能够与模型共同成长的评测体系。X2Real在真实反映模型能力的基础上,能够提供更加细致的诊断方案,持续参与模型的改进,让每一次评测都成为模型能力下一次增长的起点。

自变量团队相信,建立持续可信的评测反馈机制,是具身智能实现长期持续进步的前提。当模型能力能够被持续验证,失败经验就能被系统地积累,数据、算力与算法资源就能投入到更有针对性的研究方向,提升模型研发的效率。

X2Real希望为行业建立的,正是这样一个能够伴随模型持续成长的评测基础设施。在具身智能模型发展的道路上,当每一步足迹都可以被丈量、被记录、被评估,最终将汇集成一条通向“物理智能奇点”的康庄大道。

注:文/龚作仁,文章来源:Laborer,本文为作者独立观点,不代表亿邦动力立场。

文章来源:Laborer

广告
微信
朋友圈

FAQ回顾

X2Real是什么?它要解决什么问题?

X2Real是自变量于2026年发布的可进化仿真评测基准,用于评估具身模型在真实物理世界中的性能。它提出可信性、有效性、客观性三个设计原则,通过与真实硬件对齐的仿真、多维任务设计和自动化评测诊断流程,让仿真评测成为模型性能的可信标尺,是具身智能行业的关键基础设施。

具身智能模型评测目前存在哪些痛点?

目前面临双重痛点:真机评测成本昂贵、耗时缓慢,难以满足研发高频次、大规模的测试需要;仿真评测虽能提升效率,却存在仿真不够“真实”、题库不够“丰富”的问题,容易被模型“摸透规律”刷分,评测能力难以伴随模型能力共同成长。

仿真评测和真机评测的结果差距大吗?

X2Real采用与真实硬件对齐的仿真标定,在控制参数、视觉特征和物理属性上反复校准,实现了毫米级轨迹一致性。测试显示,在通用操作任务上,仿真评测与真机评测的任务成功率相关性达到0.96、进度分相关性达到0.99,居于行业最前列,能忠实反映模型能力。

如何防止模型在仿真评测中“背题”刷分?

X2Real通过三个设计防止“背题”刷分:任务多维度随机化,让模型无法死记硬背评测环境;严格分离训练与评测配置,训练数据在绿幕摄影棚采集,与考试环境完全隔离;借助Mana仿真生态系统持续生成新考题,相当于拥有无限题库。

X2Real仿真评测支持哪些机器人?

X2Real的仿真标定覆盖单臂Franka、双臂ArtiXon Arm-6A、ARX R5和移动操作平台Quanta X1等多种主流机器人本体。同时,开发者无需在本地搭建仿真环境,只需接入X2Real的统一模型接口,平台即可在线完成评测判分和轨迹回放,大幅降低评测门槛。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0