WAIC机器人狂卷落地:细分赛道爆发,VLA 与世界模型争做大脑The WAIC Robo-Sweep Hits the Floor: Niche Markets Erupt as VLA and World Models Vie for Brain Dominance


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/MNA8GxZKfQFDr9_SvWoLgg
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

图片作者:林帆 | 监制:罗超
图片

7 月 17 日,第九届世界人工智能大会(WAIC)在上海世博中心正式开幕,作为一年一度的 AI 行业盛会,今年大会依然聚集了国内外大模型公司、AI 应用、硬件厂商和机器人公司,雷科技(ID:leitech)旗下AI新媒体「雷科技AGI(leikejiagi)」也派出报道团,赴上海进行现场报道。

5b4e9b5f8bb853a79cdc3f1883b52631
图源:雷科技摄制

这次WAIC,能明显感觉到人形机器人行业彻底换了画风。

往年的各大展会里,机器人都主打一个观赏性,翻跟头、舞蹈、拳击等等,今年WAIC,机器人同时担任产品与工作人员,各种人形机器人分散在整个场馆角落,专门给观众指路、看管无人零售点位,这也是行业第一次,在大型展会公共区域大规模投放全尺寸人形机器人商用。

花哨表演退场,量产落地,成了今年全场的核心主线。

图片

行业开卷干活,

机器人厂商各有各的打法

智元:WAIC 2026“镇馆之宝”,定义部署态元年

智元是目前人形机器人行业里,少数同时在技术和交付两个维度都拿出了东西的公司。它不仅是展台上最热闹的,也是真实场景里跑得最快的。

智元给今年的WAIC定了一个主题词:“部署态元年”,因此,智元也带来了多个部署态产品与场景。其中包括远征A3 Ultra、精灵G2 Max、灵犀 X2 EDU版本、临界点OmniHand 3 Ultra-M和酷拓世界首个骑行机器人。

远征A3 Ultra是本届WAIC 10件镇馆之宝中唯一的人形机器人产品。其身高174厘米,全身自由度超50个,双臂负载38kg,搭载最新灵巧手,集成端到端VLA大模型,在真实工业场景中能稳定执行高难度操作任务。

走进智元展台的作业部署态区域,智元还原了料盘上下料产线部署态场景,精灵G2机器人可自主识别、抓取、转运芯片料盘,精准完成上下料全流程,适配半导体车间无尘、高重复度的作业需求。现场还设有均普智能的整条芯片加工产线,展示机器人连续完成芯片上料、成品装盒、整箱转运超长流程作业。

7c36c9c63ea59d14d20abc3fe706c7d7
图源:雷科技摄制

在交互部署态区域,智元还将广州地铁真实应用场景搬到WAIC现场,精灵G2可自主完成安检引导、咨询服务等常态化工作,成都宽窄巷子游客服务中心的“迷你版”也空降智元展区。

针对巡检场景,旗下子品牌智元酷拓也带来两大自研核心引擎 “灵越”、“灵途” 补齐四足机器人能力短板。灵越全地形运动引擎靠感知+运控双模架构,工业场站复杂地形通行无压力;灵途全域导航引擎解决巡检机容易丢定位的痛点,搭配自动充电,支持数月无人值守,如今这套方案已经落地电力、化工、园区安防等巡检项目。

在技术底座上,智元GO-2自研具身基座大模型做决策大脑,提出动作思维链与异步双系统,打通规划与执行。世界模型GE-2登顶2026 World Arena世界模型赛道总分冠军,从“世界模拟”升级到“世界学习”。

展区外围,智元还安排了60台机器人服务WAIC大会,由智元旗下RaaS平台擎天租负责。

60台机器人连续服务几天,换班、充电、故障处理全程无人干预,敢把60台全尺寸人形机器人扔到大型展会的公共区域连续跑,本身就是一次压力测试。

地瓜:560TOPS的算力底座,大幅缩短开发周期

地瓜有意思的地方在于,它不造机器人,在一个所有人都盯着人形机器人本体的时候,它选了另一条路:当水电煤。

因此,可能在普通用户看来,地瓜所展示的内容相当硬核。

其中,旭日600是面向具身智能机器人的大算力SoC,搭载CPU、DSP、GPU、NPU、ISP等算力单元,专为代表性机器人、操作、规划、感知和控制等智能任务打造,为人形机器人及具身智能系统提供强劲算力、高可靠性和低延迟响应。

IMG_256
图源:地瓜机器人

具体来看,旭日S600具备560TOPS(INT8)算力,搭载4核BPU Nash,支持模型高效推理。18核A78AE CPU拥有高性能CPU性能集群,用于任务灵活调度,6核R52+MCU面向实时控制场景,强调稳定可靠;同时采用256-bit LPDDR5x高带宽内存,带宽达204.8 GB/s,为具身机器人端侧计算提供高带宽数据支撑。

为了让机器人开发从数月压缩到数周,地瓜带来了三层工具链,云端RoboGo管数据生成、训练、量化与发布,PC端RDK Studio管设备连接与真机部署,板端Moss管任务执行,三层之间Moss AgentEngine统一调度。

值得一提的是,目前旭日S600已获得20+头部客户认可,开启量产验证,地瓜也已经联动100多家产业链伙伴完成协同适配,覆盖芯片、模组、传感器、机器人控制器、数据采集等环节。

魔法原子:三款机型覆盖全场景,从晚会演出落地实战场景

魔法原子在WAIC上全球首发三款战略新品:旗舰全尺寸MagicBot X1、工业轮式MagicBot D1、轻工业四足MagicDog T1,同步展出的还有九大场景解决方案和自研通用具身大模型Magic-VLA K02。

三款产品各司其职,X1身高180厘米,31个主动自由度,单关节峰值扭矩450N·m,能稳定完成扣篮、乒乓球、击剑等高动态任务,走通用能力极限突破路线。

D1已进入追觅智能制造工厂,轮式底盘加人形上肢,靠无轨自主导航和集群调度,打通仓储取料、跨区运输、产线上下料的无人周转闭环,T1则是补上了轻工业和园区的分散化场景。

1
图源:雷科技摄制

自研 Magic-VLA K02 模型现场演示了叠盒封胶、柔性衣物整理、行李箱完整收纳这类超长流程复杂任务,整套一体化作业成功率突破 90%。

这套模型采用双层协同逻辑,上层先读懂完整任务目标,下层连贯执行一连串精细动作,彻底摆脱早年机器人只能死板复刻单一步骤的局限,能直接以最终交付结果为导向自主规划全套操作流程。

商业化落地这块,魔法原子算是圈内经验很足的玩家,早前百台级机器人集群就上过央视春晚、海外赛事迎宾表演,今年 4 月更是直接切入公共安全赛道,和无锡市公安局达成合作,自研交管人形机器人“小麦” 投入无锡马拉松现场疏导交通,算得上人形机器人正式走进政务场景参与实战。

乐聚机器人:实打实进厂搬货,整机国产零部件占比超 95%

乐聚机器人在展区搭了一个显眼的“搬砖区”,1:1复刻真实产线,纸箱拆垛、塑料箱拆垛、小件上料,三台人形机器人吭哧吭哧地搬,遇到箱体歪斜自己校准,从头到尾不用人管。这些方案已经在工厂部署运行三到四个月,客户名单覆盖一汽、中兴、长虹等10多个行业的头部企业。

乐聚还展出了全链路开发工具链,能把部署周期大幅压缩,标准化展厅场景1到2天就能部署,工业场景1到2周,定制场景1到2个月。跟蚂蚁灵波合作的物流方案只用300条数据、1到2天完成微调,搬运成功率80%以上。

a9bbccd623fed2ea92eca974fbc05218
图源:雷科技摄制

值得一提的是,乐聚夸父系列的国产化率已超过95%,联合辉羲智能做的国产高算力主控方案,把全栈国产化最后一块拼图也补上了,同时,乐聚“面向汽车制造场景的人形机器人智能拣运”方案也已入选工信部人工智能应用典型案例。

千觉机器人:补上触觉感知,靠触感分辨球鞋真假

千觉是WAIC上最容易被忽略的一家,但它做的事恰恰是行业最短的那块板。它想解决的是机器人触觉这个更底层的物理问题。机器人碰到物体之后,还是看得见、摸不准、控不住,没有触觉反馈,跨不过精细化作业的门槛。

87a5bc0399657f9684ac18ef96d71fdb
图源:雷科技摄制

千觉机器人在WAIC期间发布X-TouchMind V1,首个面向触觉智能的VTLA具身触觉模型,外加一个千小时视触觉多模态数据集TacVerse 1k。

展台上同时摆着双臂长序列纸盒成型和耳机精密装配两个Demo,前者让机器人在柔性纸板上连续完成抓取、展开、对位、折边、压实全套动作;后者切入3C装配场景,让机器人在精密操作中靠触觉反馈自主完成精准对位和动态纠偏。

在得物App的展台上,这套触觉智能已经跑起来了。机器人搭载千觉自研的VTLA和触觉传感器,5秒内完成鞋子全自动质检,鉴别结果跟人类专家吻合度超99.9999%。

这个数据纯靠视觉完全无法获取,千觉的力感知精度达到毫牛级,表面形貌感知达到微米级,夹持和按压全程柔性自适应,对鞋品零损伤。

另外,千觉还展出了自研的XTac UMI G1穿戴式视触觉数采夹爪,能把真人实操中的接触变化、受力反馈、滑移趋势转化为高质量具身数据集。整手触觉解决方案则通过指尖传感器和电子皮肤模组,搭建了从指尖到掌面的全域触觉反馈网络。

启元机器人:变形金刚进客厅,主打开源 DIY 改装玩法

上纬新材旗下启元机器人,在WAIC上让启元T1完成了公开首秀,全球首个可变形个人机器人,Transformer跨形态一体架构,轮足人形和四足形态自主切换,室内是人形助手,出门变机器狗,它能联动Insta360等运动相机,语音拍摄、智能跟随、轨迹运镜。

启元Q1探索者版同步发布,外观拿了国际A' Design Award设计金奖,结构件全面开源,支持3D打印个性化改装。

启元卖的不只是一台机器人,同时也是一个用户能自己改外壳、加配件、改功能的硬件平台,更接近消费电子的逻辑。

IMG_256
图源:启元

可变形、开源、3D打印改装,这些设计语言在消费电子行业是基本功,在机器人行业是冒险,但新物种就得需要这种勇于探索的态度。目前启元Q1已正式开启北美、欧洲市场布局,以设计为桥梁向全球用户传递产品价值。未来启元Q1将继续以“人人拥有专属机器人”为目标,走进全世界更多人的生活中。

图片
机器人大脑成焦点,

VLA和世界模型双线并行

不管各家推出什么新奇机型、花哨方案,归根到底目标只有一个,让机器人落地干活。

机器人想要顺利完成作业,离不开两大核心环节:一是智能大脑梳理任务逻辑、规划操作步骤;二是机械机身精准执行动作。

今年整个大会从业者聊得最多、分歧最大的话题,就是机器人智能大脑该走哪条技术路线。

目前行业主流分为 VLA(视觉-语言-动作模型)、世界模型两条赛道,二者各有短板,暂时还没有诞生能完美融合两者的统一框架。

VLA 模型优势在于听懂人类自然语言、快速适配多样场景,但短板十分明显,很难预判物体受力形变、摩擦运动等物理规律;世界模型可以精准推演各类物理交互结果,缺陷则是解读文字、理解人类复杂指令的能力偏弱。

大晓机器人就是纯粹世界模型路线代表,但其研发负责人也公开表示:现阶段世界模型技术尚且不够完善,面对标准化固定任务,还是 VLA 学得更快、运行更稳,短期两条路线会同步并行发展。

他预判,等机器人真实实操数据积累到千万小时规模,世界模型会逐步吸纳 VLA 全部能力,最终形成大一统基座大模型。

1
图源:雷科技摄制

在WAIC主论坛上,智元合伙人、觅蜂科技的姚卯青也点明了现在世界模型存在两大硬伤:

第一是数据类型错配。互联网视频和机器人需要的物理世界数据根本是两回事。互联网上的内容可能反物理规律,人飞起来都没问题,但机器人需要的是推拉拧拽、柔性物体、摩擦力、流体这些真实物理交互数据。

这类数据在网上稀缺,现有物理引擎也很难模拟。

第二是世界模型这条路,瓶颈在数据,不在架构。大语言模型预训练数据已经100万亿Token,相当于100亿小时的说话量。物理世界数据的信息密度远低于文本,要实现高阶物理常识判断,可能需要亿小时级别的真实世界数据。

姚卯青的判断是:高频刚需、环境可控、确定性强、有容错空间的场景最先落地,更开放的环境和任务等通用能力再往前迈一步。

c5ddba1f6da6e5db3270f2fd86317806
图源:雷科技摄制

因此,智元机器人选择双线并行的折中路线,一边依托 GO-2 自研具身基座大模型充当决策大脑,搭配动作思维链、异步双系统保障 VLA 模型在真实场景稳定落地;另一边持续迭代拿下赛道冠军的 GE-2 世界模型,完成从虚拟世界模拟到真实世界自主学习的升级。

同时推出 Genie Evolver 1.0 真机闭环强化学习框架,走出第三条差异化思路:不再依赖互联网杂乱的公开视频数据训练,而是让机器人在真实作业场景持续交互、沉淀实操经验,自主迭代进化。

世界模型什么时候能够成熟?按地瓜CEO王丛的判断来说:当人们不再刻意提“世界模型”这四个字时,才说明这条路真跑通了,就像今天没人纠结自动驾驶用了哪种技术流程,只看实际跑起来怎么样。

图片
机器人迎来“宁德时刻”,

供应链走向台前

今年逛展还有一个特别直观的变化:芯片、关节、传感器这类上游供应链厂商,不再依附整机品牌当配角,大大方方走到台前展示自家核心产品。

其中国产化推进速度肉眼可见,乐聚夸父系列人形机器人整机零部件国产化率突破 95%,同时联合辉羲智能打造国产高算力主控方案,补全人形机器人全栈国产化最后一块核心硬件短板,他们针对汽车工厂打造的人形拣运方案,还入选工信部人工智能典型落地案例。

6ff63839b539d3e0644822a5c6ba4108
图源:雷科技摄制

地瓜机器人旭日 S600 算力模组联动上百家上下游产业链企业完成软硬件协同适配,覆盖芯片、核心模组、各类传感器、机器人运动控制器、多模态数据采集全环节。

此外,智身科技还自研 CHAMP 冠军系列关节模组,除了供给自有整机生产线,还对外供货全行业厂商。过去关节部件长期占据人形机器人近 50% 整机成本,海外企业垄断核心技术与产能多年,国产自研关节模组直接打破这项卡脖子难题。

当前人形机器人全品类核心零部件整体国产化率已经突破 75%,从底层算力芯片、驱动关节、多模态感知传感器,一套完整国产自主供应链体系逐步成型。这也意味着人形机器人彻底摆脱实验室零散进口零件拼凑的试验样机身份,变成可以大批量稳定生产、标准化交付的成熟工业品。

除了整机、大模型、算力基座之外,仿人灵巧手也是当下厮杀格外激烈的细分赛道,各家路线、感知方案各有亮点,在本届 WAIC 集中亮出了自家核心硬件。

以灵心巧手为例,这次一次性展出两条主流技术路线的成熟产品:一个是面向复杂精细操作、适配各类具身操作大模型训练的直驱款 Linker Hand O30 系列;还有一个是迭代升级后的全新腱绳传动灵巧手 Linker Hand L30 Pro。

两种方案分别对应不同抓取、操控场景,覆盖科研与工业柔性作业需求。

图片

图源:灵心巧手

临界点则带来旗舰机型 OmniHand 3 Ultra-M 完成国内首次亮相,这款全直驱灵巧手内置 20 个主动自由度,每根指尖都搭载一体化视触觉传感器,整只手掌铺满三百余个三维触觉感知点位,能同步捕捉物体压力、形变与滑动细节,是做高精度接触式操作的核心硬件。

触觉底层配套方面,他山科技搭建起一套完整的触觉感知产业生态,从自研触觉芯片、传感元件到成套落地解决方案全部打通。

其中全场核心展品绿宝石 E10A 芯片被官方定义为全球首款动态触觉感知芯片,拥有微秒级超高时间分辨率,待机功耗仅 0.9μA;依托这款芯片的技术优势,企业相关触觉产品在人形机器人赛道市场占有率已经突破八成,是行业里触觉硬件的核心供应商。

机器人行业的供应链正在复制电动汽车走过的路,当年宁德时代和比亚迪也是从被嘲笑一路走到全球第一,机器人的“宁德时刻”,未必很远。

图片
没有全能选手,产业链加速分工

文章看到这,相信大家也能发现一个现象:再也没有企业一门心思包揽产业链全部环节。

两年前人形机器人赛道,几乎所有企业宣传核心都是“全栈自研”,从底层芯片、驱动关节、算法模型到整机装配,恨不得所有零部件、生产工序全部自主完成。

但2026 WAIC呈现出完全不一样的全新产业格局,每家企业找准一小块细分赛道深耕细作,不再贪心通吃整条产业链。

做芯片的专心做芯片,地瓜机器人不造整机,它在做算力底座和开发工具链,旭日S600联动100多家产业链伙伴,优必选Cruzr Y1供“大脑”。

做关节的专心做关节,智身科技的CHAMP冠军系列关节模组,不仅自用,还往外卖。关节曾占人形机器人总成本近50%,长期被国外卡着,现在有人专门干这件事,还做到百万台级的年产能规划。

做触觉的专心做触觉,千觉机器人成立不到两年,三轮融资,让机器人“摸得着”,VTLA模型、XTac UMI G1数采夹爪、整手电子皮肤方案,不造整机,专攻触觉这个“最后一厘米”。

3282a5995c9773d0c9fa2958b9ba3df9
图源:雷科技摄制

做整机的专心做场景落地,智元、魔法原子、乐聚把精力集中在把机器人塞进工厂、地铁、公安、仓储、产线。它们不用从头造芯片和关节,产业链上有人替它们做了。

做运营的专心做运营,擎天租不卖机器人,但客户付月费,机器人就能上岗。

一家公司从芯片到整机到运营全包的时代结束了。这不是坏事,智能手机行业就是最好的参照,高通做芯片,索尼做传感器,苹果做系统,富士康做组装。没有谁从头吃到尾,但每家公司都在自己那一环做到了极致。

机器人行业正在走到这条路上,分工不是退步,是成熟的标志,2026年的WAIC,每一家都找到了自己的位置,这才是这个行业最大的好消息。

图片
写在最后

逛完整场 2026 WAIC,最大的感受就是今年所有人的讨论重心,全都落在产能、交付、落地成本、场景复现率这些非常现实的产业问题上,人形机器人已经走出实验室,正式走进我们的生产、公共服务甚至日常生活。

当然行业依旧存在不少待解难题:家用场景复杂环境适配能力有限、部分核心零部件成本还有下降空间、大一统融合模型尚未成型。但不可否认,如今整个行业已经走完最艰难的从零到一的摸索阶段,进入从一到百的规模化扩张窗口期。

各家企业不再闭门造车、盲目内卷,而是依托完整国产产业链协同互补,有人做底座、有人做硬件、有人做整机、有人做运营,各司其职稳步推进。

2026 这场大会更像一道分界线,人形机器人已经上岗就业,成为了持续创造实际价值的生产力工具,属于人形机器人规模化商用的黄金阶段,真的拉开帷幕了。

#WAIC#机器人#具身智能#智元#地瓜#魔法原子

主题为「智能伙伴·共创未来」的WAIC 2026正在火热开展中!
AI叙事路线从模型参数堆叠,转变到Agent生产力落地;异构协同、光子计算持续提高计算上限;具身智能加速应用,机器人到家进厂让物理AI变为现实。
雷科技WAIC探展团已抵达上海,直击AI产业化落地年度巅峰时刻,敬请关注!‍
图片
End
解锁车主专属圈,同系车主,都在这里!

入群畅聊爱车,欢迎大家扫码添加小助理拉群图片

 

Source: https://mp.weixin.qq.com/s/MNA8GxZKfQFDr9_SvWoLgg If there are copyright issues, please notify us via email: 13951431913#139.com or submit a comment on the right.

Author: Lin Fan | Supervising Editor: Luo Chao图片

图片

On July 17, the 9th World Artificial Intelligence Conference (WAIC) officially opened at the Shanghai Expo Center. As the annual grand event for the AI industry, this year's conference still brought together domestic and international large-model companies, AI application developers, hardware manufacturers, and robotics firms. Leitech's (ID: leitech) AI new media outlet "Leitech AGI (leikejiagi)" also sent a reporting team to Shanghai for on-site coverage.

5b4e9b5f8bb853a79cdc3f1883b52631

Photo source: Leitech

At this WAIC, it was clear that the humanoid robot industry had completely changed its style.

At major exhibitions in previous years, robots were mainly about spectacle—doing flips, dancing, boxing, and so on. This year at WAIC, robots served as both products and staff members, with various humanoid robots scattered throughout the venue, specifically guiding visitors and watching over unattended retail points. This was also the industry's first time deploying full-size humanoid robots at scale for commercial use in the public areas of a large exhibition.

Flashy performances have exited the stage, and mass production and real-world deployment have become the core theme of the entire event.

图片

The industry has shifted to practical work,

and robot makers each have their own approach

Zhiyuan: WAIC 2026 "Treasure of the Hall," defining the first year of deployment

Zhiyuan is currently one of the few companies in the humanoid robot industry that has delivered results in both technology and deployment. It was not only the liveliest booth at the show, but also the fastest-moving in real-world scenarios.

Zhiyuan set this year's WAIC theme as "Year One of Deployment," so it brought multiple deployment products and scenarios, including the Yuanzheng A3 Ultra, Elf G2 Max, Lingxi X2 EDU edition, Critical Point OmniHand 3 Ultra-M, and KUTUO's world-first riding robot.

The Yuanzheng A3 Ultra is the only humanoid robot product among WAIC's 10 "treasures of the hall." Standing 174 cm tall, with over 50 degrees of freedom across the body, dual-arm payload of 38 kg, the latest dexterous hands, and an integrated end-to-end VLA large model, it can stably perform high-difficulty operational tasks in real industrial scenarios.

In the operational deployment area of the Zhiyuan booth, Zhiyuan recreated a chip tray loading/unloading production line deployment scenario. The Elf G2 robot can autonomously identify, grasp, and transfer chip trays, accurately completing the entire loading and unloading process, suited to the dust-free, highly repetitive operational needs of semiconductor workshops. The site also featured a full chip processing line from Joupu Intelligent, demonstrating robots continuously completing ultra-long workflow operations including chip loading, finished product packaging, and full-box transfer.

7c36c9c63ea59d14d20abc3fe706c7d7

Photo source: Leitech

In the interactive deployment area, Zhiyuan also brought Guangzhou Metro's real application scenario to the WAIC site, where the Elf G2 can autonomously complete routine tasks such as security check guidance and concierge services. A "mini version" of the Chengdu Kuanzhai Alley tourist service center also made a special appearance at the Zhiyuan exhibition area.

For inspection scenarios, Zhiyuan's sub-brand KUTUO also brought two self-developed core engines, "Lingyue" and "Lingtu," to fill the capability gaps of quadruped robots. The Lingyue all-terrain motion engine uses a dual-mode perception + motion control architecture, allowing smooth passage over complex terrain in industrial stations. The Lingtu full-domain navigation engine solves the pain point of inspection robots easily losing positioning, and with automatic charging, it supports unattended operation for months. This solution has already been deployed in power, chemical, and campus security inspection projects.

On the technology foundation side, Zhiyuan's GO-2 self-developed embodied foundation large model serves as the decision-making brain, proposing an action chain of thought and asynchronous dual systems to connect planning and execution. The world model GE-2 topped the 2026 World Arena world model track overall ranking, advancing from "world simulation" to "world learning."

Outside the exhibition area, Zhiyuan also arranged 60 robots to serve the WAIC conference, managed by Zhiyuan's RaaS platform Qingtianzu.

Sixty robots worked continuously for several days, with shift changes, charging, and fault handling requiring no human intervention throughout. Daring to put 60 full-size humanoid robots in the public areas of a large exhibition for continuous operation was itself a stress test.

D-Robotics: A 560 TOPS computing power foundation that greatly shortens development cycles

What makes D-Robotics interesting is that it does not build robots. While everyone else was focused on humanoid robot bodies, it chose a different path: serving as the water, electricity, and gas infrastructure.

As a result, what D-Robotics showcased may seem quite hardcore to ordinary users.

Among its offerings, the Sunrise 600 is a high-computing-power SoC for embodied intelligent robots. It integrates computing units such as CPU, DSP, GPU, NPU, and ISP, and is purpose-built for intelligent tasks including representative robot operation, planning, perception, and control, providing humanoid robots and embodied intelligent systems with strong computing power, high reliability, and low-latency response.

IMG_256

Photo source: D-Robotics

Specifically, the Sunrise S600 delivers 560 TOPS (INT8) of computing power and features a 4-core BPU Nash for efficient model inference. Its 18-core A78AE CPU provides a high-performance CPU cluster for flexible task scheduling, while the 6-core R52+MCU is aimed at real-time control scenarios, emphasizing stability and reliability. It also uses 256-bit LPDDR5x high-bandwidth memory with 204.8 GB/s of bandwidth, providing high-bandwidth data support for on-device embodied robot computing.

To compress robot development from months to weeks, D-Robotics brought a three-layer toolchain: cloud-based RoboGo handles data generation, training, quantization, and release; PC-side RDK Studio handles device connection and real-machine deployment; and board-side Moss handles task execution, with Moss AgentEngine providing unified scheduling across the three layers.

Notably, the Sunrise S600 has already gained recognition from 20+ leading customers and entered mass-production validation. D-Robotics has also coordinated with more than 100 supply chain partners to complete collaborative adaptation, covering chips, modules, sensors, robot controllers, and data collection.

Magic Atom: Three models covering all scenarios, from gala performances to real combat scenarios

Magic Atom held the global launch of three strategic new products at WAIC: the flagship full-size MagicBot X1, the industrial wheeled MagicBot D1, and the light-industrial quadruped MagicDog T1. It also showcased nine scenario solutions and its self-developed general-purpose embodied large model, Magic-VLA K02.

Each of the three products has its own role. The X1 stands 180 cm tall, has 31 active degrees of freedom, and delivers a peak joint torque of 450 N·m. It can stably complete high-dynamic tasks such as dunking, table tennis, and fencing, following a route of pushing the limits of general capability.

The D1 has already entered Dreame's smart manufacturing factory. Combining a wheeled chassis with a humanoid upper body, and relying on trackless autonomous navigation and cluster scheduling, it completes an unattended closed loop for warehouse picking, cross-zone transport, and production line loading/unloading. The T1 fills the gap for decentralized scenarios in light industry and industrial parks.

1

Photo source: Leitech

The self-developed Magic-VLA K02 model demonstrated ultra-long, complex tasks such as box stacking and sealing, flexible clothing organization, and complete luggage storage, with overall integrated operation success rates exceeding 90%.

This model uses a two-layer collaborative logic: the upper layer first understands the complete task objective, while the lower layer continuously executes a series of fine actions. It completely breaks away from the early limitation where robots could only rigidly replicate single steps, and can directly plan the full set of operations autonomously, oriented toward the final delivered result.

On commercial deployment, Magic Atom is a player with considerable industry experience. Earlier, a cluster of over 100 robots appeared on the CCTV Spring Festival Gala and at overseas event welcome performances. In April this year, it directly entered the public safety track, partnering with Wuxi Public Security Bureau to deploy its self-developed traffic management humanoid robot "Xiaomai" at the Wuxi Marathon for traffic guidance—marking the first time a humanoid robot formally entered government scenarios for real-world participation.

Leju Robotics: Genuinely working in factories and moving goods, with domestic components accounting for over 95% of the entire machine

Leju Robotics set up a conspicuous "brick-moving zone" at its exhibition area, replicating a real production line 1:1. With carton depalletizing, plastic box depalletizing, and small-part feeding, three humanoid robots worked away steadily, self-correcting when boxes were skewed, with no human supervision from start to finish. These solutions have already been deployed and running in factories for three to four months, with a client list covering leading companies across more than 10 industries, including FAW, ZTE, and Changhong.

Leju also showcased a full-link development toolchain that can greatly compress deployment cycles: 1 to 2 days for standardized exhibition hall scenarios, 1 to 2 weeks for industrial scenarios, and 1 to 2 months for customized scenarios. The logistics solution developed with Ant Lingbo requires only 300 data samples and 1 to 2 days of fine-tuning, achieving a handling success rate of over 80%.

a9bbccd623fed2ea92eca974fbc05218

Photo source: Leitech

Notably, the domestic content rate of Leju's Kuafu series has exceeded 95%. Together with HuiXi Intelligent, it developed a domestic high-computing-power main control solution, completing the final piece of the fully domestic supply chain. In addition, Leju's "humanoid robot intelligent picking and transport solution for automotive manufacturing scenarios" has been selected as a typical AI application case by the Ministry of Industry and Information Technology.

Qianjue Robotics: Filling the gap in tactile perception, distinguishing genuine and fake sneakers by touch

Qianjue is the easiest company to overlook at WAIC, but what it does is precisely the industry's shortest board. It aims to solve the more fundamental physical problem of robot touch. After a robot touches an object, it can still see it but cannot feel it accurately or control it well. Without tactile feedback, it cannot cross the threshold of fine-grained operations.

87a5bc0399657f9684ac18ef96d71fdb

Photo source: Leitech

During WAIC, Qianjue Robotics released X-TouchMind V1, the first VTLA embodied tactile model for tactile intelligence, along with TacVerse 1k, a thousand-hour visuotactile multimodal dataset.

The booth featured two demos: a dual-arm long-sequence carton forming task and a precision headphone assembly task. The former has the robot continuously complete the full set of actions—grasping, unfolding, aligning, folding edges, and pressing—on flexible cardboard. The latter enters the 3C assembly scenario, where the robot relies on tactile feedback to autonomously complete precise alignment and dynamic correction during precision operations.

At the Dewu App booth, this tactile intelligence system was already running. Equipped with Qianjue's self-developed VTLA and tactile sensors, the robot completes fully automatic shoe inspection within 5 seconds, with identification results matching human experts at a rate exceeding 99.9999%.

This data is completely unobtainable through vision alone. Qianjue's force perception accuracy reaches the millinewton level, and its surface morphology perception reaches the micron level. Gripping and pressing are fully flexible and self-adaptive throughout, causing zero damage to the shoes.

In addition, Qianjue also showcased its self-developed XTac UMI G1 wearable visuotactile data collection gripper, which can convert contact changes, force feedback, and slip trends from real human operations into high-quality embodied datasets. The full-hand tactile solution uses fingertip sensors and electronic skin modules to build a full-domain tactile feedback network from fingertips to palm.

Qiyuan Robotics: Transformers enter the living room, focusing on open-source DIY modification play

Qiyuan Robotics, under Shangwei New Materials, gave the Qiyuan T1 its public debut at WAIC. It is the world's first transformable personal robot, with a Transformer cross-form unified architecture that autonomously switches between wheeled-foot humanoid and quadruped forms. Indoors it is a humanoid assistant; outdoors it becomes a robot dog. It can connect with action cameras such as Insta360 for voice-triggered shooting, intelligent tracking, and trajectory-based camera movement.

The Qiyuan Q1 Explorer Edition was also released at the same time. Its design won the international A' Design Award gold prize, and its structural components are fully open source, supporting personalized modification through 3D printing.

What Qiyuan sells is not just a robot, but also a hardware platform where users can change the shell, add accessories, and modify functions themselves—closer to consumer electronics logic.

IMG_256

Photo source: Qiyuan

Transformability, open source, and 3D-printed modification are basic skills in the consumer electronics industry, but they are a gamble in the robotics industry. Yet a new species needs exactly this kind of exploratory attitude. The Qiyuan Q1 has already begun its expansion into the North American and European markets, using design as a bridge to convey product value to users worldwide. In the future, the Qiyuan Q1 will continue toward the goal of "a personal robot for everyone," entering the lives of more people around the world.

图片

The robot brain has become the focus,

with VLA and world models advancing on two parallel tracks

No matter what novel models or flashy solutions companies launch, the ultimate goal is only one: getting robots to work in the real world.

For a robot to complete tasks smoothly, two core elements are indispensable: first, an intelligent brain that organizes task logic and plans operational steps; second, a mechanical body that precisely executes actions.

This year, the most discussed and most divisive topic among attendees was which technical route the robot brain should take.

The industry currently splits into two main tracks: VLA (vision-language-action models) and world models. Each has its own shortcomings, and no unified framework perfectly merging the two has yet emerged.

The advantage of VLA models lies in understanding natural human language and quickly adapting to diverse scenarios, but their shortcoming is obvious: they struggle to predict physical laws such as object deformation under force and frictional motion. World models can accurately reason about the results of various physical interactions, but their weakness is a relatively poor ability to interpret text and understand complex human instructions.

Daxiao Robotics is a pure representative of the world model route, yet its R&D lead also publicly stated that world model technology is still not mature enough at this stage. For standardized fixed tasks, VLA still learns faster and runs more stably, and the two routes will develop in parallel for the short term.

He predicted that once real robot operational data accumulates to the scale of ten million hours, world models will gradually absorb all of VLA's capabilities, ultimately forming a unified foundation large model.

1

Photo source: Leitech

At the WAIC main forum, Yao Maoqing, a partner at Zhiyuan and Mifeng Technology, also pointed out two major shortcomings in current world models:

First is data type mismatch. Internet video and the physical world data robots need are fundamentally different things. Online content may violate physical laws—people flying around is no problem—but robots need real physical interaction data involving pushing, pulling, twisting, flexible objects, friction, and fluids.

Such data is scarce online, and existing physics engines also struggle to simulate it.

Second, for the world model route, the bottleneck lies in data, not architecture. Large language models have already been pretrained on 100 trillion tokens, equivalent to 10 billion hours of speech. Physical world data has far lower information density than text, and achieving high-level physical common sense judgment may require billions of hours of real-world data.

Yao Maoqing's judgment: scenarios with high-frequency demand, controllable environments, strong certainty, and room for error tolerance will land first, while more open environments and tasks will have to wait for general capabilities to advance another step.

c5ddba1f6da6e5db3270f2fd86317806

Photo source: Leitech

Therefore, Zhiyuan Robotics chose a compromise route of running both tracks in parallel. On one hand, it relies on the self-developed GO-2 embodied foundation large model as the decision-making brain, paired with an action chain of thought and asynchronous dual systems to ensure stable deployment of VLA models in real scenarios. On the other hand, it continues iterating on the GE-2 world model that won its track championship, completing the upgrade from virtual world simulation to autonomous learning in the real world.

It also launched Genie Evolver 1.0, a closed-loop reinforcement learning framework on real robots, charting a third differentiated path: instead of relying on messy public video data from the internet for training, it lets robots continuously interact and accumulate hands-on experience in real operational scenarios, autonomously iterating and evolving.

When will world models mature? According to D-Robotics CEO Wang Cong: only when people stop deliberately mentioning the phrase "world model" will it show that the route has truly worked—just as today no one fusses over which technical process autonomous driving uses, only how it actually performs on the road.

图片

Robots are ushering in their "CATL moment,"

and the supply chain is stepping into the spotlight

Another very visible change at this year's exhibition: upstream supply chain players such as chip, joint, and sensor makers are no longer playing supporting roles attached to complete-machine brands. They are stepping boldly into the spotlight to showcase their core products.

The pace of domestic substitution is visibly accelerating. Leju's Kuafu series humanoid robots have surpassed 95% domestic content across all components, and together with HuiXi Intelligent, Leju has built a domestic high-computing-power main control solution, completing the final core hardware short board in the fully domestic humanoid robot supply chain. Its humanoid picking and transport solution for automotive factories has also been selected as a typical AI application case by the Ministry of Industry and Information Technology.

6ff63839b539d3e0644822a5c6ba4108

Photo source: Leitech

D-Robotics' Sunrise S600 computing module has coordinated with over a hundred upstream and downstream supply chain companies to complete software-hardware collaborative adaptation, covering chips, core modules, various sensors, robot motion controllers, and multimodal data collection across the entire chain.

In addition, Zhishen Technology has self-developed the CHAMP series of joint modules, supplying not only its own complete-machine production lines but also selling to manufacturers across the industry. Joint components have long accounted for nearly 50% of the total cost of a humanoid robot, and overseas companies monopolized core technology and production capacity for years. Domestic self-developed joint modules have directly broken this bottleneck.

Currently, the overall domestic content rate of core components across all categories of humanoid robots has exceeded 75%. From foundational computing chips and drive joints to multimodal perception sensors, a complete domestic autonomous supply chain system is gradually taking shape. This also means humanoid robots have fully shed their identity as experimental prototypes assembled from scattered imported parts in laboratories, and have become mature industrial products that can be mass-produced stably and delivered in a standardized way.

Beyond complete machines, large models, and computing foundations, anthropomorphic dexterous hands are also an especially fiercely contested niche track this year. Each company's route and perception approach has its own highlights, and they all concentrated on showing their core hardware at this WAIC.

Take Lingxin Clever Hand as an example: this time it exhibited mature products across two mainstream technical routes at once—one is the direct-drive Linker Hand O30 series, aimed at complex fine operations and adapted for training various embodied operation large models; the other is the newly iterated and upgraded tendon-driven dexterous hand, the Linker Hand L30 Pro.

The two solutions correspond to different grasping and manipulation scenarios, covering research and industrial flexible operation needs.

图片

Photo source: Lingxin Clever Hand

Critical Point brought its flagship OmniHand 3 Ultra-M for its domestic debut. This fully direct-drive dexterous hand has 20 active degrees of freedom, with integrated visuotactile sensors on every fingertip and more than 300 three-dimensional tactile sensing points spread across the entire palm. It can simultaneously capture object pressure, deformation, and sliding details, making it core hardware for high-precision contact-based operations.

On the supporting tactile foundation side, Tashan Technology has built a complete tactile perception industry ecosystem, covering self-developed tactile chips, sensing elements, and complete turnkey solutions.

Among them, the show's core exhibit, the Green Gemstone E10A chip, has been officially defined as the world's first dynamic tactile perception chip, featuring microsecond-level ultra-high time resolution and standby power consumption of just 0.9μA. Leveraging this chip's technical advantages, the company's related tactile products have already captured over 80% market share in the humanoid robot track, making it the core supplier of tactile hardware in the industry.

The robotics industry's supply chain is replicating the path electric vehicles once took. Back then, CATL and BYD also went from being ridiculed to becoming world number one. The robot industry's "CATL moment" may not be far off.

图片

No all-round players exist; the industrial chain is accelerating its division of labor

By this point in the article, readers can surely notice a phenomenon: no company is trying to cover every link of the industrial chain by itself anymore.

Two years ago, in the humanoid robot track, almost every company's marketing core was "full-stack self-development," from underlying chips, drive joints, and algorithm models to complete machine assembly—they wanted to do all components and production processes themselves.

But WAIC 2026 presented a completely different new industrial landscape. Each company has found a small niche track to cultivate deeply, rather than greedily trying to dominate the entire industrial chain.

Chip makers focus on chips. D-Robotics does not build complete machines; it builds computing foundations and development toolchains, with the Sunrise S600 connecting more than 100 industry chain partners, while UBTech's Cruzr Y1 supplies the "brain."

Joint makers focus on joints. Zhishen Technology's CHAMP series joint modules are not only used internally but also sold externally. Joints once accounted for nearly 50% of the total cost of a humanoid robot and were long constrained by foreign suppliers. Now someone is dedicated to this, with annual production capacity planned at the million-unit level.

Tactile specialists focus on touch. Qianjue Robotics, founded less than two years ago with three funding rounds, makes robots "able to feel." With its VTLA model, XTac UMI G1 data collection gripper, and full-hand electronic skin solution, it does not build complete machines but focuses solely on touch—that "last centimeter."

3282a5995c9773d0c9fa2958b9ba3df9

Photo source: Leitech

Complete-machine makers focus on scenario deployment. Zhiyuan, Magic Atom, and Leju concentrate their efforts on getting robots into factories, subways, public security, warehouses, and production lines. They do not need to build chips and joints from scratch—someone in the supply chain does it for them.

Operations specialists focus on operations. Qingtianzu does not sell robots, but customers pay a monthly fee and the robot goes to work.

The era when one company handled everything from chips to complete machines to operations is over. This is not a bad thing. The smartphone industry is the best reference: Qualcomm makes chips, Sony makes sensors, Apple makes systems, and Foxconn does assembly. No one covers everything from start to finish, but every company excels at its own link.

The robotics industry is heading down this same road. Division of labor is not regression; it is a sign of maturity. At WAIC 2026, every company found its own position—and that is the biggest piece of good news for the industry.

图片

Final thoughts

After walking through the entire WAIC 2026, the biggest takeaway is that this year everyone's focus has shifted entirely to very practical industrial issues such as production capacity, delivery, deployment costs, and scenario replication rates. Humanoid robots have walked out of the laboratory and officially entered our production, public services, and even daily life.

Of course, the industry still faces many unresolved challenges: limited adaptability to complex home environments, room for further cost reduction in some core components, and the absence of a unified merged model. But there is no denying that the entire industry has now completed the hardest stage of exploration from zero to one and entered a window of large-scale expansion from one to one hundred.

Companies are no longer working behind closed doors or blindly competing. Instead, they rely on a complete domestic industrial chain for coordinated complementarity—some build foundations, some build hardware, some build complete machines, and some run operations, each doing its part and advancing steadily.

This 2026 conference is more like a dividing line. Humanoid robots have already gone to work and become productive tools that continuously create real value. The golden stage of large-scale commercial use for humanoid robots has truly begun.

#WAIC #Robotics #EmbodiedIntelligence #Zhiyuan #D-Robotics #MagicAtom

WAIC 2026, themed "Intelligent Partners · Creating the Future Together," is now in full swing!

The AI narrative route has shifted from stacking model parameters to Agent productivity deployment; heterogeneous collaboration and photonic computing keep raising the ceiling of computing power; embodied intelligence is accelerating into applications, with robots entering homes and factories, turning physical AI into reality.

The Leitech WAIC exploration team has arrived in Shanghai, covering the annual pinnacle moment of AI industrialization. Stay tuned!

图片

End

Unlock the exclusive owner community—same-model owners are all here!

Join the group to chat about your car. Scan the QR code to add our assistant and be invited to the group.图片

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry