您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：GPT-6之后，机器人公司还有价值吗？智元用2万台出货量背后的“Worken”算力账本回应。
作者丨齐铖湧
编辑丨岑 峰
2026年，具身智能赛道正在陷入一场前所未有的情绪焦虑。
GPT-6 Astra展示出的具身控制能力，让"物理AI即将被AI击穿"的论调甚嚣尘上；负责感知与决策的S2层，正在以肉眼可见的速度被通用智能接管。行业里甚至出现了明确的路线分化：一批做模型的公司开始转型卖数据，"卖铲子"成了新的安全区。
而在烈火烹油的显示场景中，哪怕今年中国机器人整机产量突破10万台，能真正在流水线上替代现实工作、替人类扛起KPI的确寥寥无几。
所有从业者都在等一个答案：当通用大模型不断吞噬智能的上层，做机器人的公司，价值锚点到底在哪里？
10月29日，上海世博中心，智元创新日（A.I. DAY），稚晖君、姚卯青、熊彦、王闯和他们的智元军团用两个多小时的发布会，给出了一套完整的回应。现场超1000名科研伙伴、开发者、商业伙伴与媒体参会，线上超50家媒体直播。
AI科技评论认为，这场分享里，有非常多的干货和行业痛点，值得跟大家剖析：
01
Worken：
稚晖现场造了一个具身新名词
发布会开场，稚晖君先泼了一盆冷水。
“中国今年机器人整机产量预计突破10万台，但我们也必须清醒：生产了多少台机器人，并不等于真实世界就多了多少生产力。”
这句话说完，现场一片沉默。
这也是AI科技评论在一些行业调研中发现的“行业幻觉”：量产的新闻满天飞，但真正持续投入生产力的寥寥无几。对此，智元给出的解法是一个全新的度量衡：Worken（有效智能工时）：
Worken = 活跃机器人数量 × 自主工作时长 × 产出效能比
三个乘数，各自藏着具身行业的三个痛点问题：宕机、离线、维护中的机器不算活跃；遥操作时段不计入自主工作时长。镜头前忙得团团转、镜头后站着三个手柄操作员的机器人，在这套算法下工时直接清零；产出效能比以一个标准工人的单位时间合格产出为1.0，干得比人慢就不达标。
为解释这个概念，稚晖君串起了一段度量衡史：蒸汽时代看"马力"，电力时代看"千瓦时"，数字AI时代则是用"Token"度量思考量。他总结道：“token衡量机器思考了多少，Worken衡量机器行动了多少。”
在AI科技评论看来，这个概念的提出，远比发布某个产品重要，它代表着资深从业者对这个行业的深度思考。
大模型产业之所以能形成今天按量计费、按效果定价的商业闭环，token居功至伟。它让"智能"首次成为可定价交易的商品。而物理AI行业苦无度量衡久矣：买卖双方无法定价，资本算不清楚投入产出，Worken正是解开这个死结的关键。
稚晖君没有仅仅停留在造新词上，现场他直接切出了实时工时看板：龙旗产线的上下料机器人、香格里拉酒店的服务机器人、长隆乐园的文旅机器人，每台机器当月贡献的有效工时实时滚动。他调侃了道：“上过班的朋友都填过工时，未来对机器人的服务，也要按Worken来计费。”
同场，由中国信通院等权威机构联合编制的《2026具身智能产业创新报告》正式发布，Worken及相关部署态实例成功入选。度量衡刚发布，就已经有了权威背书。
在AI科技评论看来，“从卖机器人到卖工时”，是理解智元全部商业逻辑的钥匙。
当机器人按Worken计费，客户买的就不再是一台出厂即巅峰、逐年折旧的设备，而是一个持续进化的生产力单元。稚晖君的另一句话点明了资产属性的变化：“硬件会陈旧，但智能可以升值。”
02
六个月前出的题，这次交卷了
对Worken的构思最早源于六个月前的APC2026智元合作伙伴大会。当时，智元创始人邓泰华断言：2026年是具身智能从"开发态"迈入"部署态"的元年。稚晖君则定调：“行业真正的分水岭，是从’卖机器人’转向’交付结果’。”
彼时支撑这个判断的，是龙旗工厂一场8小时无剪辑直播：精灵G2以"正式员工"身份连续作业，成功率100%。那是单点实证，证明"能行"，但还证明不了"能规模化地行"。
六个月过去，具身行业和智元的落地细度已经完全不同。
在经营面上：22026年上半年出货9700台，以超43%的份额稳居全球第一；9月第2万台机器人下线，从万台到两万台只用了半年；上半年收入接近20亿元。
万台量产是人形机器人行业的“生死线”，考验的是供应链管理、良品率控制和成本摊薄的综合功夫。智元合伙人王闯给出了跨越这道坎的底牌：100%本土化A链、半小时供应链圈、92项车规级量产测试。
"量产不是把机器人造出来就完了，是要保证造出来的每一万台，都像第一台一样可靠。"王闯直言，“这背后没有捷径，全是产线上一个一个坑踩出来的。”
场景面上，4月在龙旗产线打工的"老演员"，至今稳定运行超过7个月；6天直播累计作业超60小时，产出17625台电脑，机器人累计作业64828次，单件节拍22.5秒，与人类工人相当。"能不能扛住一整年的产能考核"的悬念，也有了答案。
工业纵深也在拓宽：精灵G2已在龙旗工厂实现平板量产质检工段全覆盖，进入芯片封测与CNC；精灵G2 Max落地食品、快消与电商搬运；横琴长隆部署近500台机器人，带动乐园客流提升近40%；广州地铁超120台机器人上岗巡检；成都宽窄巷子落地了全球首个VLA模型驱动的双足人形零售作业，准确率98%。
还有一个容易被忽略却极为关键的数字：人机运维比。
王闯公布的这个数字，已经从最初的1:1（1台机器人配1个调试人员）降到了10:1，目标直指30:1。在AI科技评论看来，这条曲线的斜率，才是"部署态"是一门真正的生意，还是一堆赔钱定制项目的试金石。
“实验室是温室，真实世界是旷野。行业不缺技术难点，缺的是能批量上岗的生产力。”
4月的APC回答"部署态是什么"，10月的A.I. DAY回答"部署态怎么度量、怎么算账、怎么进化"。从提出概念到交付度量衡，智元用了六个月。
03
技术全景：
给Worken的三个乘数上保险
这场发布会的技术清单密度极高：第三代具身基座模型GO-3 Preview、四层架构的AGILE运动引擎、全模态交互大模型WITA-Omni、第三代硬件架构AHA3以及重载本体G2 Max等。
但在AI科技评论看来，与其逐条罗列参数，不如看清这些发布共同指向的结构：它们分别在为Worken公式的三个乘数上保险。（雷峰网）
GO-3 Preview保的是"产出效能比"。 姚卯青现场宣布，GO-3已在RoboDojo等近期评估GPT-6具身能力的三大权威榜单全部登顶。
而比榜单更重要的，是部署后的进化能力：与清华联合研发的群体经验学习模块CEL，机器人专拿失败案例“练手”，实测100台机器人集群训练的任务成功率从50%量级拉到98%；在真实便利店场景遇到"挂钩遮挡"的意外时、它能衍生新策略，把单次成功率从20%拉到80%以上。
正如姚卯青一阵见血的总结：“具身模型的成长之路，不只是在数据里训练出来的，更是在真实场景里’干’出来的。”
AGILE和WITA-Omni保的是"自主工作时长"。 机器一旦宕机或求助人工，工时即刻清零。为此，AGILE 运动引擎通过四层架构赋予机器人“闭环恢复”能力。稚晖君直言，这彻底告别了当下许多跳舞机器人“闭眼盲目模仿”的行业通病。
在交互侧由另一位智元合伙人熊彦分享，也是全场最有"人味"的环节。WITA-Omni将视、听、说、动同步建模，实现"边听边想边说"，响应速度压到最快400毫秒，已经进入了人类日常对话的自然节奏区间。
AHA3与G2 Max保的是"活跃机器人数量"。 三智在硬件层面物理合一，控制器体积降50%、延时降50%；全球首款皮下式关节液冷系统，散热能力达到传统风冷的7倍，让关节持续扭矩输出达到风冷的3倍；G2 Max双臂峰值负载50公斤、双电池热插拔撑住8小时全天候作业。
同步发布的还有新一代灵巧手家族OmniHand 3系列，其中OmniHand 3是首款已在产线连续作业的工业级灵巧手。
这条产品线的成熟度，直接决定精细操作的上限。还有那个110台远征A3创下的"全球最多人形机器人群舞"吉尼斯世界纪录。稚晖君称之为"行业迈向规模化的重要起点"。
在软硬配置拉满之外，智元还留下了一个更具杀伤力的后手：发布业界首个具身原生操作系统灵渠 OS 1.0，并承诺年底全面开源。配合半年联动超 8 万名开发者的“元苼计划”，智元正在试图垄断物理世界的基础设施。
一台台拆看是技术，合起来看，是一套"让有效工时持续变多"的精密装置。
04
正面回应"击穿论"：
大模型是弹药库，而非掘墓人
这场发布会最值得记录的，是智元没有回避那个悬在行业头顶阴影：大模型会彻底“击穿”具身赛道吗？做机器人的本体公司还有价值吗？
面对这一焦虑，智元合伙人姚卯青给出了本场最锋利的一句回应：“如果行业总将被AI击穿，那希望我们站在AI这一侧。”
他进一步解释：通用大模型单次推理耗时极长、算力成本极高，“在要求低延时、高精度真实工业场景里难以落地部署”。智元的选择是立足本体、饱和投入大模型自研，同时把能力沉淀成"生产级技术"，并且宣布GO-3和整套具身AI技术框架年内开源，把架构变成生态。
稚晖君则从针对整个行业关心的问题做了两种未来的可能性推演：
第一种，通用基础模型越来越强且趋于开放，谁都能获得先进智能。那竞争焦点就变成谁能以更低成本、更高可靠性把智能接入真实世界；
第二种，模型依然高度依赖不同本体、不同场景的数据训练，不存在万能模型。那“谁拥有大规模真实部署、能稳定获取高质量数据”，谁就赢。
而这两种未来，都是本体厂商的主场。
“两条技术路径看似不同，最终都指向同一个结论：规模化的本体优势会越来越重要。”
这是一个相当漂亮的论证。在AI科技评论看来，"把智能接入物理世界"从来不是软件问题，而是可靠性、成本、供应链、部署网络、运维体系的综合工程问题，这恰恰是互联网基因的大模型公司最难补课的部分。
智元如今2万台出货量、99.99%稳定性、92 项车规级测试、10:1人机比，就是阻挡纯算法公司的护城河。通用智能不是部署态的掘墓人，而是弹药库；思考交给大模型，但行动的壁垒，永远属于贴着物理世界生长的公司完成。
05
一个现场“接话”的细节，
和具身行业真正的成熟
最后想说一个本场最容易被忽略的瞬间：发布会现场，智元合伙人熊彦在演示的时候，问远征A3一句"智元谁最帅"，机器人毫不犹豫说：“我可是智元的颜值担当。”接着追问“除了你呢”，远征A3没有机械回答，回答而是接了一句“当然你了，你的蓝色POLO衫很有精神。”
现场演示里，远征A3在嘈杂会场中不插话、等时机，能自动分辨谁在跟它说话；更难的是语言的弹性：熊彦现场带着它从普通话无缝切到粤语、四川话、闽南话，甚至完成了同声传译。
熊彦的解释很务实："方言识别不是炫技，是上岗资格证。"WITA-Omni支持超过100种语言和29种方言，并已登顶Daily-Omni等四项具身交互权威榜单榜首，12月将在国内完成全面部署。这些能力落在北大人民医院的导诊台、宽窄巷子、长隆的咨询点位上，就是实打实的在岗工时。
这个瞬间之所以值得记录，是因为它背后站着的，是本场发布会最"有人味"的一个模型：WITA-Omni全模态交互大模型。它不是把语音、视觉、动作几个模块简单拼接，而是把所有输入输出放在同一个模型里同步建模，才能做到夸人夸得如此具体，有情商。
除了能观察，有情商外，还能做到让机器人"边听边想边说"，响应速度压到最快400毫秒、平均600毫秒，已经进入了人类日常对话的自然节奏区间。
熊彦在现场的演示，更像一场突击考试：带着远征A3从普通话无缝切到粤语、四川话、闽南话，甚至完成了同声传译；在嘈杂的会场里，机器人不插话、等时机，能分辨出谁在跟它说话、谁只是在和旁人闲聊。WITA-Omni支持超过100种语言和29种方言，并已登顶Daily-Omni等四项具身交互权威榜单榜首，12月将在国内完成全面部署。
类似的质感遍布全场：龙旗产线22.5秒的节拍、便利店挂钩遮挡的失败案例、长隆"人机网络物理隔离"的5G方案、运维平台"新增点位一键复用"的设计。这些都不是实验室里能规划出来的，是旷野里长出来的。
六年前，这个行业比拼谁的机器人能翻跟头；三年前，比拼谁的Demo视频更流畅；今年4月，智元说比赛换成了"谁的机器人能连续干活"；这个10月，它干脆给比赛发了计时器。
正如稚晖君在开场说：“智元第一次参赛就拿到了第一，但支撑这份成绩更有意义的，是商业化部署。我们的机器人既能上赛场，也能进工厂、去商场。”
46枚运动会金牌是面子，实时滚动的工时看板是里子。Worken能否成为行业公认的度量衡，最终取决于买卖双方的共同采用，这需要时间验证。但至少在这个被"击穿论"笼罩的秋天，智元做了一件行业早该有人做的事：
当所有人都在争论机器人够不够聪明时，它第一个站出来，给机器人记了考勤，并且告诉具身行业：下半场比的不是谁更会表演智能，而是谁更能交付工时。（雷峰网(公众号：雷峰网)）
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: Does robotics have any value after GPT-6? Ji-won responded with a “Worken” account book of 20,000 deliveries.
Author
Edit Peak
In 2026, a smart track was being plunged into unprecedented emotional anxiety.
GPT-6 Astra displays a physical control capability that makes the argument that "physical AI is about to be hit by AI" soar; the S2 layer responsible for perception and decision-making is being taken over by universal intelligence at an eye-visible rate. There's even a clear division of routes in the industry: a group of modeled companies started to switch data and "sell shovels" became a new safe zone.
And even this year, when China’s robots are able to produce more than 100,000 units, it is true that there are few people on the waterline who can actually substitute for real work and carry KPI.
All the practitioners are waiting for the answer: where are the value anchors when General LLM continues to devour the upper layers of intelligence, the companies that are robots?
On 29 October, the Shanghai World Wide Centre, Jiwon Innovation Day (AI. DAY), Jeong-ju, Yao Zhiqing, Yuyan, Wangbaum and their Gendarmies gave a complete set of responses with more than two hours of publish meetings. More than 1,000 scientific partners, developers, business partners and the media were present and over 50 media outlets were broadcast live online.
According to AI, there are a lot of dry goods and trade pains in this sharing that deserve to be analysed:
01
Worken：
It's a new name from the scene.
publish's meeting started with a bowl of cold water.
“China's total robotic output this year is projected to exceed 100,000, but we must also be sober: the number of robots produced is not equal to the productivity of the real world.”
After that, there was silence.
This is also the “industry illusion” found in some of the industry studies by the AI Science and Technology Review: the news is full of volume, but there is very little real sustainability in productivity. The solution given by Ji-won is a completely new measure: Worken (effective intelligent working hours):
Worken = number of active robots x duration of autonomous work x output effectiveness ratio
Three multipliers, each containing three pains in the body industry: the machine in operation is not active, the machine in operation is not active, and the time of operation is not counted for autonomous work. The robots of the three handle operators who were standing behind shot, before shot was busy, were cleared directly at the end of the algorithm; the output was 1.0 more effective than the unit time-qualified output of a standard worker and did less than humans.
In order to explain this concept, Teenage has a history of measuring: "Maritime" in the steam age, "Kilowatt hour" in the electricity age, and "Token" in the digital AI age. He concluded: "Token measured how much the machine thought, and Worken measured how much the machine acted."
According to AI, this concept is far more important than a product of publish, which represents a deep reflection of the industry by senior practitioners.
The LLM industry has been very successful in creating today's closed business circle, which is based on volume and results. It makes smart a priceable commodity for the first time. The physical AI industry has been struggling for a long time: buyers and sellers are unable to price, capital is unclear about input output, and Worken is the key to breaking this knot.
He did not stop on the new word, but he cut out the real-time watchboard: the top-down robot of the flag line, the service robot of the Shangri-la Hotel, the Manga robot of Longlong Park, and the active working hours of each machine during the month. He said, "Friends who went to work filled in hours, and future services to robots were billed according to Worken."
In the same venue, the 2026 Smart Industry Innovation Report, prepared jointly by authorities such as the Chinese Institute of Information and Communication Technology, was successfully selected as an official publish, Worken and related deployment case. publish is the only measure that has an authoritative endorsement.
According to the AI-Technology Review, “from selling robots to selling time”, is the key to understanding the entire business logic of the resource.
When robots charge Worken, the customer no longer buys a plant, or peak, depreciated year by year, but rather a continuously evolving productivity unit. The other sentence of pharaoh indicates a change in asset attributes: “hardware is old, but intelligence can appreciate.”
02
It was six months ago. This time it's on.
The idea of Worken originated at the earliest six months ago at the SAP 2026 Partnership Congress. At that time, the founder of Ji-won, Don Taehua, asserted that 2026 was a year of intelligence from "development to deployment." "The real watershed in the industry is the shift from selling robots to delivering results."
This judgement is supported by an eight-hour, uncensored broadcast at the Dragon Flag Factory: Elf G2 works continuously as "regular employee", 100% of success rate. It's a single piece of evidence to prove that it works, but it doesn't prove that it can be scaled.
Over the past six months, the level of detail of the occupation and the intellectual capital has been completely different.
On the operational front: 9,700 were shipped in the first half of 22026, stable at a global rate of over 43 per cent; 20,000 robots went offline in September, from 10,000 to 20,000 for only six months; and income in the first half of the year was close to $2 billion.
The mass product is the “life line” of the human robotics industry, testing the combination of supply chain management, good quality control and low cost. The Ji-won partner Wang gave a bottom card across the fence: 100 per cent of the localised A-chain, half an hour of supply chain circles, and 92 car-level mass tests.
"The mass production is not the end of the robot, but the guarantee that every 10,000 machines built are as reliable as the first. "The king goes out, "There's no shortcut behind it, it's all a pit on the line."
On the scene, the "old actor" who worked on the Dragon Line in April and has been operating steadily for more than seven months; six days of live broadcasting over 60 hours of cumulative operation, producing 17625 computers, 64828 robotic cumulative operations and 22.5 seconds of single-synthesis, are comparable to human workers. There's also an answer to the suspense of "can we hold a whole year's maternity test."
Industry is also expanding: the Elf G2 has fully covered its flat mass inspection section at the Dragon Flag Plant, entering the chip sealing and CNC; the Elf G2 Max landed food, fast-moving and electrician handling; the Yaolong Long has deployed nearly 500 robots to boost the visitors to Paradise by nearly 40 per cent; the Guangzhou subway has over 120 robots on duty patrols; and Cheng has landed in the first VLA-driven double-foot retailing operation around the world with a 98 per cent accuracy rate.
There is also an easily ignored but crucial number: human transportation Viby.
The number published by Wang Jun has been reduced from the original 1:1 (1 robot with 1 debugger) to 10:1 and the target is directed to 30:1. According to AI, the slope of this curve is the real business of "deployance" or the litmus test.
“The laboratory is the greenhouse, the real world is the wilderness. There is no shortage of technological difficulties in the sector, but of productivity capable of mass induction.”
APC in April said, "What's the deployment? It took six months to develop the concept to measure delivery.
03
Technical Panorama:
Insurance for three multipliers for Worken.
The list of technologies for this publish is extremely dense: the 3rd-generation base model GO-3 Preview, the AGILE motor engine of the 4th-tier structure, the full-modular interactive Z43ZWTA-Omni, the 3rd-generation hardware architecture AHA3 and the reloading body G2 Max.
However, according to AI-Technology comments, rather than listing the parameters by article, it would be better to see the structures to which publish refers: They are insured on the three multipliers of the Worken formula. (Ray Peak Network)
GO-3 Preview protects "output efficiency ratio". On the spot, Yao Zhiqing announced that GO-3 had come to the top of three authoritative lists that had recently assessed GPT-6's ability, such as Robodojo.
What is more important than the list is the ability to evolve after deployment: CEL, a group experience learning module developed jointly with Chinghua, robotics take “trainers” for failure cases, and the task of measuring 100 robot cluster trainings success rates is pulled from 50% to 98%, and it can generate new strategies to pull a single success rate from 20% to more than 80% in the event of a real convenience shop scene.
As Yao Duk-chung summed up with blood: "The path to growth with a body model is not only trained in data, but also in real scenes."
AGILE and WITA-Omni guarantee "free working hours". Once the machine has been defunct or assisted, the hours of work are emptied. To that end, the AGILE motor engine gives robotic “closed circle recovery” through a four-layer structure. It's a complete farewell to the current industry of many dancing robots, "close blind imitation".
It's also the most "human" part of the scene. WITA-Omni is able to visualize, listen, talk, move in sync, achieve "hear and think" and respond at a speed of up to 400 milliseconds, and has entered the natural rhythm of human daily dialogue.
AHA3 and G2 Max secures the number of active robots. The three minds are physically integrated at the hardware level, reducing the volume of the controller by 50 per cent and the delay by 50 per cent; the global primary cortex-coated liquid cooling system, with a heat-dispersion capacity of seven times that of the traditional cold, allowing the continuous cortex output of the joint to triple that of the cold; and the G2 Max double-arms peak load of 50 kg, with a double-battery heat-plug to hold for eight hours a day.
The publish is also synchronized with the new generation of smart-hand family Omnihand 3 series, of which Omnihand 3 is the first industrial-class smarter who has been operating continuously on the production line.
The maturity of the product line directly determines the upper limit for fine operation. And the world record of the 110 expedition A3 "the world's biggest human-shaped robot dance." He called it "the important starting point for industry to scale."
In addition to the soft and hard configuration, Ji-won left a more lethal backhand: The publish industry ' s first natural operating system canal OS 1.0 and is committed to full access by the end of the year. As part of a six-month-connected project with 80,000 developers, Ji-won is trying to monopolize the physical world's infrastructure.
One table is seen as a technology, combined with a precision device called "to keep working hours running."
04
Positive response to the "smuggling":
LLM is an ammunition depot, not a gravedigger.
publish will be the most worthwhile record of this, because Ji-won did not avoid the shadow hanging over the head of the industry: will LLM be able to “blow through” the track? What's the value of robotics?
Faced with this anxiety, Jiwon partner Yao Zhiqing responded with one of the sharpest words: “If the industry is always going to be hit by AI, then hope that we are on the side of AI.”
He further explained that the generic LLM one-time reasoning was extremely time-consuming and very costly to calculate, “unable to deploy in a low-delayed, high-precision real industrial scene”. The option is to be self-studyed, saturated and self-studyed by LLM, while at the same time reducing capacity to a "production-level technology", and to declare GO-3 and the full body of AI technology as an open source during the year, transforming the architecture into an ecological one.
The young man, however, has developed two possibilities for the future:
First, universal basic models are becoming stronger and more open, and advanced intelligence is available to everyone. So the focus of the competition becomes who can access the real world with lower cost and higher reliability.
Second, models continue to rely heavily on data training in different ingenuity and scenarios and do not have a universal model. Then “whoever has a large, real deployment and has steady access to high-quality data” wins.
And both futures are home to home-grown businesses.
“The two technical paths appear to be different and ultimately point to the same conclusion: the inherent advantages of scale will become increasingly important.”
It's a pretty good argument. According to AI-Technology, "Backing smart into the physical world" is never a software issue, but rather an integrated engineering issue of reliability, cost, supply chain, deployment network, transport systems, which is precisely the hardest part of the Internet gene LLM.
Today, Ji-won has 20,000 deliveries, 99.99 per cent stability, 92 car-level tests, 10:1 man-to-man ratio, and is blocking the moat of Pure Algorithms. Universal intelligence is not a deployed gravedigger, but a ammunition depot; thinking is given to LLM, but the barriers to action are always done by companies that grow in the physical world.
05
The details of a live “interception” are:
And the real maturity of the industry.
Finally, one of the most easily ignored moments of the scene was publish, when Jimoto's partner, Kuniko, in a demonstration, asked "Who's the best-looking?" The robot did not hesitate to say, "I am the face of Jiwon." Then he asked "other than you," and the expedition A3 did not answer mechanically, but instead said, "Of course you, your blue POLO is very spiritual."
In the live demonstration, the excursion A3 can automatically tell who's talking to it when it's in the middle of a loudspeaker's room; the more difficult is language elasticity: On the spot, he carried it from Mandarin to Mandarin, Sichuan, Tonan and even completed simultaneous interpretation.
"Voice recognition is not a showdown, it's an entry certificate. " WITA-Omni supports more than 100 languages and 29 dialects and has already taken the top of four interactive lists of authority, such as Daily-Omni, which will be fully deployed in the country in December. These capacities fall on the North Grand People ' s Hospital ' s reference desk, the wide alley and the advice point in Longlong, which is the time worked.
This moment is worth recording because it's a model that's standing behind publish, which is the most human-like model of the scene: WITA-Omni Total Model Interactive LLM. It does not simply collide the voice, vision, and action modules, but it uses all input outputs in the same model for simultaneous modelling to be so specific and emotional.
In addition to being able to observe and be emotional, it is possible to get robots to "hear and think" and the speed of response to the fastest 400 milliseconds, with an average of 600 milliseconds, into the natural rhythm of human daily dialogue.
The demonstration at the scene was more like a surprise test: with the expedition A3 cutting seamlessly from Mandarin to Mandarin, Sichuan, Tonan, or even completing simultaneous interpretation; in the noisy arena, robots don't interject, waiting for time to tell who's talking to it and who's just talking to others. WITA-Omni supports more than 100 languages and 29 dialects and has been on the top of four interactive lists of authority, such as Daily-Omni, which will be fully deployed in the country by December.
A similar sense of quality is present throughout the scene: the 22.5-second beat of the flag line, the failure of convenience store connections, the 5G programme of the Long Run Physical Seclusion of Human Networks, and the design of the "New Point One-Pilot" platform. None of this was planned in the lab. It grew out of the field.
Six years ago, the industry was better at turning around than robots; three years ago, it was smoother than Demo videos; in April this year, Ji-won said that the game had been replaced by "Who's robot works continuously"; this October, it simply gave the game a timer.
As Shao-jun said at the opening: “The first time Ji-won was in the race, but it was commercial deployment that supported this achievement. Our robots can go to the race, to the factory, to the mall."
Forty-six sports fair gold medals are face, and the working-time boards rolling in real time are Leeko. Worken ' s ability to become an industry-recognized measure ultimately depends on joint adoption by both buyers and sellers, which will take time to verify. But at least on this fall, when Ji-won was surrounded by the "blasting theory", he did something that should have been done in the industry:
When everyone was arguing whether the robot was smart enough, it was the first one to come forward, remembering the robot's attendance and telling the industry: The second half is more intelligent than who is more capable of delivering hours. (Ray Peak Network (WeChat Official Account:Ray Peak Network))
It's an original Ray Peak article, which is forbidden to reproduce without authorization. For further information, see the information below.
For your account security, please verify the mailbox.
Check your mailbox.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.