• 亿邦公众号
    • APP下载

从模型到生产力 具身智能还要跨过哪些坎?|WRC观点

作者:亿邦动力 文章来源:亿邦动力

2026-08-22 14:39

作者:亿邦动力 文章来源:亿邦动力

2026-08-22 14:39

邦小白快读

EN
全文速览

本文核心讲解当前具身智能行业的发展阶段、技术进展与商业化路径,核心干货整理如下:

1. 行业整体现状:目前具身智能已经进入从技术验证走向商业化落地的关键窗口期,全球18个月内涌现出超过一百款开源基础模型,技术每三个月就有明显进步,但行业仍处于销售主导阶段,主要靠卖整机硬件盈利,模型能力还没有形成独立的付费闭环。

2. 核心技术进展:头部企业星海图的基础模型在8个权威第三方榜单排名第一,采用统一架构打通了零样本泛化、通用抓取、长程任务三大核心能力,依托落地的强化学习技术,已经能把操作精度提升到毫米甚至亚毫米级,作业成功率高达99.9%。

3. 现存行业瓶颈:当前机器人作业效率仅为人类的70%-80%,核心短板是泛化能力不足,完成一项新任务目前需要10小时后训练,后续还有很大提升空间。

本文为具身智能领域的品牌商提供了清晰的产业趋势、研发方向与盈利策略参考,核心干货整理如下:

1. 产业发展趋势:具身智能未来价值重心会逐步从整机硬件向智能能力迁移,商业模式会分三步演进:当前卖整机的销售主导阶段、未来3年左右的方案订阅阶段、长期的物理世界Token销售阶段,品牌商需要提前将核心竞争力从硬件制造转向智能能力构建。

2. 定价与盈利趋势预判:当前行业整机毛利率在40%-60%,进入方案订阅阶段硬件毛利率会回落至20%左右,进入Token销售阶段整机甚至会进入负毛利,靠拉动智能消耗盈利,品牌商需要提前调整盈利结构,摆脱对硬件利润的依赖。

3. 产品研发方向:核心要聚焦泛化能力提升,降低新任务培训成本,对标速度、精度、泛化性三大核心指标优化产品,参考头部企业的“1+3+N”战略,拓展产业合作伙伴,优先落地电商零售、工业制造等刚需赛道。

对于具身智能行业的卖家,本文明确了行业增长机会、商业模式变化与风险提示,核心干货整理如下:

1. 市场增长机会:具身智能处于商业化早期增量市场,目前已经在电商前置仓、第三方仓储、工业装配等场景跑通落地案例,星海图今年已经实现千台级交付,明年目标数倍增长,电商零售、工业制造、快递物流、商业服务四大赛道都有明确的需求缺口,海外场景也有拓展空间。

2. 商业模式迭代机会:未来行业会从卖整机逐步转向卖方案、卖智能能力,卖家可以提前转型,针对不同垂直场景打造定制化落地方案,拓展新的盈利增长点,也可以和头部技术厂商合作,依托成熟模型能力降低自身技术门槛。

3. 风险提示:当前行业尚未形成模型能力的独立付费闭环,核心竞争力是泛化能力,如果模型培训成本过高、泛化能力不足,很难在后续竞争中存活,中小卖家需要避免盲目囤货布局硬件,优先聚焦场景落地服务。

对于具身智能产业链的工厂,本文明确了生产需求、商业机会与数字化转型启示,核心干货整理如下:

1. 产品生产设计需求:具身智能是软硬结合的领域,对硬件的精度要求远高于传统工业产品,当前旗舰产品要求力控精度0.1N、位置控制精度0.1mm,工厂需要升级自身的精密制造能力,才能满足行业的生产要求,适配高端具身智能产品的生产需求。

2. 产业链合作机会:当前行业已经进入千台级交付阶段,未来几年市场规模会有数倍增长,头部企业星海图计划在2027年将合作伙伴数量从当前的20余家拓展到100家,开放“1+3+N”产业合作体系,硬件工厂可以作为合作伙伴切入具身智能赛道,获取新的增长增量。

3. 自身数字化转型启示:具身智能已经能完成上百步的长程精细作业,精度达到毫米级,成功率高达99.9%,可以替代人工完成拣选、装配等工作,工厂自身可以试点引入具身智能机器人,提升生产环节的自动化水平,降低人力成本。

对于服务具身智能行业的服务商,本文明确了行业趋势、客户痛点与服务方向,核心干货整理如下:

1. 行业发展趋势:具身智能正处于从技术验证到商业化落地的关键转型期,未来1-3年会逐步进入方案订阅阶段,大量场景落地需要第三方服务商提供适配、优化、运营服务,随着产业规模扩张,对服务商的需求会快速增长,市场空间大。

2. 行业核心客户痛点:当前行业的核心痛点是模型泛化能力不足,完成一项全新任务需要10小时的后训练,客户普遍希望把训练时间压缩到1小时以内,同时当前机器人作业效率仅为人类的70%-80%,场景落地的适配成本高,这些都是服务商可以解决的痛点。

3. 解决方案方向:头部企业已经验证了基础模型+底层数据+强化学习的成熟技术路线,服务商可以围绕这个路线,针对不同垂直行业的场景需求,提供模型微调、场景适配、运营维护等服务,也可以面向开发者生态提供技术支持服务,切入快速增长的市场。

对于布局人工智能、机器人领域的平台商,本文明确了行业需求与平台布局方向,核心干货整理如下:

1. 行业对平台的核心需求:当前具身智能行业走开放合作的路线,头部企业都在快速扩张产业伙伴规模,同时还在布局开发者生态,需要平台提供合作对接、流量支持、生态建设等相关服务,帮助连接技术厂商和场景客户,共同推进落地。

2. 平台招商与运营方向:平台招商可以重点倾斜具备核心基础模型能力的头部企业,以及覆盖电商零售、工业制造、快递物流、商业服务四大赛道的场景合作伙伴,优先引入已经跑通真实落地案例的项目,比如全自主机器人前置仓、海外仓储履约等成熟项目。

3. 风向规避要点:当前行业还处于技术迭代的早期阶段,不同企业的技术实力差距很大,泛化能力不足、精度不达标的企业很难长期存活,平台需要建立核心指标筛选机制,优先引入技术指标达标、有落地成果的企业,规避落后产能带来的资源浪费。

对于具身智能领域的研究者,本文提出了产业发展的新动向、清晰的商业模式演进路径,核心干货整理如下:

1. 产业最新动向:过去18个月全球已经涌现超过一百款开源具身智能基础模型,技术能力每三个月就会出现一轮明显进步,行业已经进入从技术验证到商业化落地的关键窗口期,头部企业已经实现千台级交付,技术上已经实现单模型同时打通零样本泛化、通用抓取、长程任务三大核心能力,作业精度达到毫米级,成功率99.9%。

2. 商业模式演进规律:文章明确提出具身智能产业化会经历三个清晰阶段:第一阶段是当前的销售主导阶段,以整机销售为核心;第二阶段是方案订阅阶段,客户为场景解决方案付费,硬件只是载体;第三阶段是Token销售阶段,以销售物理世界智能能力为核心,价值重心完全从硬件转向智能。

3. 待突破的核心问题:当前行业尚未形成模型能力的独立付费闭环,机器人作业效率仅为人类的70%-80%,核心瓶颈是泛化能力不足,新任务培训成本偏高,这些都是后续产业研究和技术研究需要突破的核心方向。

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 outlines the current development stage, technical progress and commercialization path of the embodied intelligence industry, with key takeaways as follows:

1. Overall industry status: Embodied intelligence is now at a critical inflection point transitioning from technical validation to commercial deployment. Over 100 open-source foundation models have emerged globally in the past 18 months, with visible technical advancement every three months. However, the industry remains sales-driven, relying primarily on revenue from complete hardware sales, and model capabilities have not yet formed an independent closed-loop paid business model.

2. Core technical progress: Xinghaitu, a leading player in the space, ranked first on 8 authoritative third-party benchmarks with its foundation model. Its unified architecture enables three core capabilities: zero-shot generalization, universal grasping, and long-horizon task execution. Backed by deployed reinforcement learning technology, the model achieves operational accuracy at the millimeter even sub-millimeter level, with a 99.9% task success rate.

3. Existing industry bottlenecks: Current robot operational efficiency only reaches 70%-80% of human levels. The core shortcoming is insufficient generalization capability: a new task still requires 10 hours of post-training to complete, leaving substantial room for improvement.

This article provides brand players in the embodied intelligence industry with clear references for industry trends, R&D directions and profitability strategies, with key takeaways as follows:

1. Industry development trends: The future value center of embodied intelligence will gradually shift from complete hardware to intelligent capabilities. The business model will evolve in three stages: the current sales-driven stage focused on hardware sales, the solution subscription phase expected to arrive in around three years, and the long-term Token-for-physical-world intelligence stage. Brands need to shift their core competitiveness from hardware manufacturing to building intelligent capabilities in advance.

2. Pricing and profitability outlook: The current industry gross margin for complete hardware ranges from 40% to 60%. When entering the subscription stage, hardware gross margin will drop to around 20%, and in the Token sales stage, hardware may even operate at negative gross margin, with profit generated from driving consumption of intelligent services. Brands need to adjust their profit structure early to reduce reliance on hardware profits.

3. Product R&D direction: Brands should focus on improving generalization capability to lower training costs for new tasks. They should optimize products against three core metrics: speed, accuracy and generalizability, follow the "1+3+N" strategy of leading companies, expand industry partnerships, and prioritize deployment in high-demand sectors such as e-commerce retail and industrial manufacturing.

For sellers in the embodied intelligence industry, this article clarifies industry growth opportunities, business model changes and risk warnings, with key takeaways as follows:

1. Market growth opportunities: Embodied intelligence is an early-stage growth market that has already achieved proven deployments in scenarios such as e-commerce fulfillment warehouses, third-party logistics warehouses and industrial assembly. Xinghaitu has reached 1,000-unit delivery volume this year, with a target of multi-fold growth next year. There are clear unmet demand gaps across four major sectors: e-commerce retail, industrial manufacturing, express logistics and commercial services, with additional expansion opportunities in overseas markets.

2. Business model iteration opportunities: The industry will gradually shift from selling complete hardware to selling solutions and intelligent capabilities. Sellers can pursue early transformation to build customized deployment solutions for different vertical scenarios to open up new profit growth points, or partner with leading technology vendors to lower their own technical barriers leveraging mature model capabilities.

3. Risk warnings: The industry has not yet formed an independent closed-loop paid model for model capabilities, and generalization is the core competitive advantage. Sellers with high model training costs and insufficient generalization capabilities will struggle to survive in future competition. Small and medium-sized sellers should avoid blindly stockpiling hardware and instead prioritize scenario-based deployment services.

For factories in the embodied intelligence supply chain, this article clarifies product requirements, business opportunities and insights for digital transformation, with key takeaways as follows:

1. Product design and manufacturing requirements: Embodied intelligence is a hardware-software integrated field, with much higher requirements for hardware precision than traditional industrial products. Current flagship products require force control accuracy of 0.1N and position control accuracy of 0.1mm. Factories need to upgrade their precision manufacturing capabilities to meet industry production requirements and adapt to the production needs of high-end embodied intelligence products.

2. Industrial chain collaboration opportunities: The industry has now entered the 1,000-unit delivery phase, and the market size will grow multi-fold over the next few years. Leading player Xinghaitu plans to expand its partner base from the current 20+ to 100 by 2027 through its open "1+3+N" industry collaboration system. Hardware factories can enter the embodied intelligence track as partners to capture new growth.

3. Insights for in-house digital transformation: Embodied intelligence robots can already complete hundreds of steps of long-horizon fine operations with millimeter-level accuracy and 99.9% task success rate, and can replace human workers for picking, assembly and other tasks. Factories can run pilot programs to introduce embodied intelligence robots to improve production automation and reduce labor costs.

For service providers serving the embodied intelligence industry, this article clarifies industry trends, customer pain points and service directions, with key takeaways as follows:

1. Industry development trends: Embodied intelligence is at a critical turning point transitioning from technical validation to commercial deployment, and will gradually enter the solution subscription phase over the next 1-3 years. Large-scale scenario deployment will require third-party service providers to provide adaptation, optimization and operation services. Demand for service providers will grow rapidly alongside expanding industry scale, creating large market opportunities.

2. Core customer pain points: The core pain point in the current industry is insufficient model generalization: a completely new task requires 10 hours of post-training, and customers generally want to reduce this training time to under 1 hour. In addition, current robot efficiency only reaches 70-80% of human levels, leading to high adaptation costs for scenario deployment. All these are pain points that service providers can address.

3. Solution directions: Leading players have already validated the mature technical path of foundation model + underlying data + reinforcement learning. Service providers can build their offerings around this path: provide model fine-tuning, scenario adaptation, operation and maintenance services tailored to the needs of different vertical industries, or provide technical support for the developer ecosystem to enter this fast-growing market.

For platform players operating in the artificial intelligence and robotics space, this article clarifies industry needs and platform positioning directions, with key takeaways as follows:

1. Core industry needs for platforms: The embodied intelligence industry is pursuing an open collaboration route, with leading players rapidly expanding their partner bases and building developer ecosystems. Players need platforms to provide partnership matching, traffic support and ecosystem building services to connect technology vendors and scenario customers to advance joint deployment.

2. Platform merchant recruitment and operation directions: Platforms should prioritize recruitment of leading players with core foundation model capabilities, as well as scenario partners covering the four high-demand sectors of e-commerce retail, industrial manufacturing, express logistics and commercial services. They should prioritize onboarding projects with proven real-world deployment, such as mature fully autonomous robot e-commerce fulfillment warehouses and overseas warehouse fulfillment projects.

3. Risk mitigation: The industry is still in the early stage of technical iteration, and there is a huge gap in technical capability between different players. Companies with insufficient generalization and unqualified accuracy are unlikely to survive long-term. Platforms should establish a core indicator screening mechanism to prioritize onboarding companies that meet technical requirements and have deployment achievements, to avoid resource waste from outdated capacity.

For researchers in the embodied intelligence field, this article presents the latest industry developments and a clear path for business model evolution, with key takeaways as follows:

1. Latest industry developments: Over 100 open-source embodied intelligence foundation models have emerged globally in the past 18 months, with visible technical advancement every three months. The industry has entered a critical window transitioning from technical validation to commercial deployment, with leading players already achieving 1,000-unit delivery. Technically, a single model has already unlocked all three core capabilities: zero-shot generalization, universal grasping, and long-horizon task execution, with millimeter-level operational accuracy and 99.9% success rate.

2. Patterns of business model evolution: This article outlines three clear stages for embodied intelligence industrialization: the first is the current sales-driven stage centered on complete hardware sales; the second is the solution subscription stage, where customers pay for scenario-specific solutions and hardware acts only as a carrier; the third is the Token sales stage, centered on selling intelligent capabilities for the physical world, with the entire value center shifting from hardware to intelligence.

3. Core unaddressed problems: The industry has not yet formed an independent closed-loop paid model for model capabilities, robot operational efficiency only reaches 70-80% of human levels, and the core bottleneck is insufficient generalization with high training costs for new tasks. These are core directions for future industrial and technical research breakthroughs.

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】过去一年,具身智能行业进入从技术验证走向商业化落地的关键窗口期。2025年至今的18个月里,全球已涌现超过一百款开源基础模型,模型能力几乎每三个月就出现一轮可感知的进步。但行业整体仍处于“销售主导”阶段,大量交易以整机硬件为主,模型能力尚未独立形成付费闭环。

具身智能如何从模型真正走向生产力?它的产业化路径、商业终局和落地关键是什么?

在2026世界机器人大会上,星海图创始人、CEO高继扬表示,具身智能本质是人工智能事业,但绝非纯软件,一定是软硬结合、构建完整体系。基于这一前提,行业价值重心将逐步从整机向智能迁移,从卖硬件走向卖智能能力。这一迁移的终点,是物理世界Token销售成为商业主体,届时整机甚至会进入负毛利状态,因为行业需要更多终端部署在物理世界,拉动具身智能的智能消耗。

他指出,而决定这条路径能否真正走通的关键,在于泛化性:培训成本越低,泛化能力越强,培训成本极高、始终学不会任务,则代表泛化能力不足。

以下为演讲实录(有删减),由亿邦动力整理:

“生产力”,是公司自创办之初就确立的核心关键词。我们希望让机器人真正走进现实世界、创造生产力,解决真实问题。而实现这一目标的核心驱动力,来自基础模型。 接下来,我将阐述我们如何看待以基础模型为核心驱动力的具身智能产业化进程。

在我们的划分中,2026年之前,模型整体能力还比较弱,几乎看不到以模型能力为核心驱动的商业化落地,市场上大量交易以整机销售为主,属于销售主导的阶段。 过去一年,基础模型能力迭代速度极快。从2025年至今的18个月里,全球已经涌现一百多款开源基础模型,几乎每三个月就能看到明显技术进展。

从今年下半年、明年,直至2028年,具身智能商业化进程中将逐步出现方案订阅模式:在特定生产力场景域形成完整产业解决方案,客户为产业方案付费,整机只是方案的物理载体。 随着智能通用性持续提升,能力从单一场景走向多场景跨域,具身智能产业商业模式将迎来再次跃迁,演进为以物理世界Token销售为主体。

在整个产业化进程中,具身智能分为整机硬件与智能能力两部分,价值重心会逐步从整机向智能迁移。 去年到今年,行业内各类整机产品毛利率大致在40%‑60%。我们判断,到第二阶段方案订阅时期,硬件毛利率或将回落至20%左右;进入Token销售阶段,整机甚至会进入负毛利状态。我们需要更多终端部署在物理世界,拉动具身智能的智能消耗。以上,就是模型能力牵引的具身智能产业化完整路径。

经过三年奋战,星海图已经取得一系列成果,基础模型位列全球第一,在8个权威第三方榜单全部排名第一。 今年第二季度,我们发布G0.5模型;今年第一季度,发布Fast‑WAM全新世界模型架构,均属于全球第一梯队。 今年8月海外全新第三方评测榜单出炉,我们模型再次取得行业第一。

行业第一背后代表什么?可以概括为:我们创新采用统一自回归架构,单模型同时打通零样本泛化、通用抓取、长程任务三大核心能力。

第一,零样本泛化,通俗来讲就是“言出法随”。给模型一条语言指令,它就可以跟随指令完成对应作业。 比如指令“把面包从盘子拿起来放到面包机”,或是“把碗里的食材倒进锅里”,这类场景模型此前从未见过,依然可以完成执行,这就是零样本泛化能力。

第二,通用抓取。抓取是具身智能非常关键的能力。通用抓取的含义是:不经过二次后训练、不重新编程,就可以对大小、厚薄、宽窄、不同材质的物体,完成从一处拾取到另一处放置。 对应的Demo在我们展位现场,大家可以掏出随身物品,直接测试模型指令跟随效果。

第三,长程任务。现实世界几乎没有工作流只包含单个或少数动作,大多由几十甚至上百个动作串联而成。让模型连续完成几十、上百步长链条任务,是基础模型产生商业价值的关键前提。 叠纸盒Demo就可以体现该能力,整套流程包含上百步精细操作,现场展位也有实物演示,欢迎大家前往观看。

强大的基础模型预训练只是起点,还需要后续迭代优化。除监督微调、模仿学习微调之外,强化学习是非常关键的手段,目前我们已经完成相关落地。 它带来的效果是:将预训练完成的基础操作,进一步迭代至毫米级、亚毫米级精度,同时作业成功率大幅提升至99.9%。

现场两段演示视频可以直观看到: 第一段为多机协同+人机协同作业。当机器人作业出现异常,人类可以直接接管,这就是机器学习领域的接管机制。接管产生的数据会回流汇总进入模型迭代链路。 第二段演示精细装配:拾取螺丝、对准孔位、螺丝刀锁付,整套毫米级操作全部来自真机强化学习。

基础模型、底层数据、强化学习,是具身智能走向真实生产力的核心变量。 进入真实生产力落地,有三个关键指标,我们概括为:速度、精度、泛化性,共同定义最终作业效果。 坦率讲,当前机器人作业效率,大概只有人类的70%‑80%;精度已经可以稳定达到毫米级别;强化学习还在持续迭代优化。

而泛化性,通俗可以用培训成本衡量。就像人类新员工上岗需要培训,机器人的培训成本高低直接对应泛化能力强弱。培训成本越低,泛化能力越强;培训成本极高、始终学不会任务,代表泛化能力不足。 目前,完成一项新任务,大概需要10小时后训练;我们的目标是未来把训练门槛压缩到1小时,甚至仅需少量样本,即可完成全新长程任务适配。

关于星海图

星海图是一家以模型为核心的具身智能企业,其基础模型在8个权威第三方榜单全部排名第一,2025年一季度发布Fast-WAM世界模型架构,二季度发布G0.5模型,8月海外第三方评测再次取得行业第一。技术上,星海图采用统一自回归架构,单模型同时打通零样本泛化、通用抓取、长程任务三大核心能力;强化学习已落地,将操作精度提升到毫米级甚至亚毫米级,作业成功率提升至99.9%。

产品层面,星海图发布全新旗舰轮臂整机Nexo,力控精度0.1N、位置控制精度0.1mm,单臂最大负载10kg、双臂合计20kg,原生适配G系列基础模型,700—1000套算力即可支撑G0.5实时推理;同时布局双足人形Kengo系列和Lemo开发者系列,分别面向非结构化复杂环境与开发者生态。

业务与战略上,星海图围绕电商零售、工业制造、快递物流、商业服务四大赛道推进落地。公司与京东合作打造全球首个全自主机器人前置仓,现场完成100多单真实履约,支持数千至上万SKU,实现商品抓取、拣选、放置、打包全流程无人干预;并在美国洛杉矶3PL第三方仓储推进拣选、打包、退货上架等海外场景。今年星海图已开始交付千台级模型驱动的真实生产力单元,明年目标实现数倍增长。公司提出“1+3+N”战略:一套软硬结合完整体系,三大产品系列,N个产业伙伴;当前合作伙伴20余家,目标到2027年突破100家。

本文首发于 亿邦动力 官方网站

【版权提示】亿邦动力倡导尊重与保护知识产权。未经许可,任何人不得复制、转载、或以其他方式使用本网站的内容。如发现本站文章存在版权问题,烦请提供版权疑问、身份证明、版权证明、联系方式等发邮件至run@ebrun.com,我们将及时沟通与处理。

已点赞

+1

{{ praise_text }}

好文章,值得被更多人看到

FAQ回顾

具身智能行业当前发展阶段及未来商业模式有什么变化?

当前具身智能行业处于销售主导阶段,交易以整机硬件为主,行业整体毛利率约40%-60%。未来将逐步演进为方案订阅模式,硬件毛利率回落至20%左右,最终进入以物理世界Token销售为主体的阶段,整机甚至可能出现负毛利。

具身智能落地为真实生产力的核心判断指标有哪些?

核心指标包括速度、精度、泛化性,其中泛化性是落地的核心关键,可通过培训成本衡量,培训成本越低泛化能力越强。当前具身智能作业效率约为人类的70%-80%,精度已稳定达到毫米级别,强化学习还在持续迭代优化。

星海图在具身智能领域有哪些核心技术优势?

星海图是国内以模型为核心的具身智能企业,其基础模型在8个权威第三方榜单全部排名第一。采用统一自回归架构,单模型同时打通零样本泛化、通用抓取、长程任务三大核心能力,强化学习落地后作业成功率提升至99.9%。

具身智能当前主要的落地应用场景有哪些?

具身智能当前主要围绕电商零售、工业制造、快递物流、商业服务四大赛道推进落地,已落地全自主机器人前置仓履约、第三方仓储拣选打包、退货上架等场景,支持数千至上万SKU的全流程无人干预作业。

广告

下一篇

征集报名《品牌数字化服务商资源手册》

《品牌数字化服务商资源手册》展示数字化服务商风采,亿邦全年零售大会通发,触达更多资源,马上加入!

马上入驻

世界模型该如何演进?|WRC观点

黄冠判断,未来两到三年,大部分普通难度任务将被通用机器人解决,行业有望迎来类似GPT-3、ChatGPT的关键时刻,五到十年,机器人将进入本体与模型协同自进化的阶段。

具身智能 智能制造

2026-08-22 13:53:44