新智元报道
这两天，机器人界「令人心动的offer」，彻底出圈了！
过去24小时，全网超200万人在线监工，围观6台智元精灵G2在真实便利店打工。今天上午十点，机器人「新员工」们在24小时便利店的「转正」挑战正式结束！
整整24小时的直播，我们全程追了下来。
这次直播，只为回答一个问题——面对真实世界的混乱难题，机器人到底行不行？
11月2日，智元决定打这场硬仗。
6台搭载智元最新具身Agent框架Genie Master的精灵G2机器人，以「便利店试用期员工」的身份，开启一场连续24小时的「转正考核」。
全程直播，全网监工。没有滤镜，没有剪辑。
就在刚刚，结果出了——挑战收官，4台成功「转正」！
「机器人打工不能停」24小时便利店直播，直接创下纪录。
6台精灵G2机器人累计完成1730项任务，包括接待顾客683次、商品取送369件、咖啡制作147次、早餐服务126次、补货巡检277次及桌面清洁128次。
这次直播，标志着一个分水岭——具身智能跳出了仿真环境、实验室，能在真实场景打工了。
把机器人放进营业中的便利店，是一个危险的决定。
因为不到 100 平米内，全是真实世界的物理混乱：机械臂得抓硬瓶、捏软袋、穿挂孔；系统要并发补货、做咖啡、拖地任务；顾客还会随时走动、插话、改单。
在这里，更致命的是多机冲突。
补货机停在货架前，挡住客户拿饮料；保洁机拐进窄道，堵死外卖动线。调度一错，机器在过道「死锁」，门店当场瘫痪。
所以才硬性要求「24 小时连续作业」。不为测电池，只为用时间放大隐患：10 分钟看不出的偶发卡顿，如何在一天内滚成任务雪崩 ？机器死机，系统怎么排队补位？
为了接住这个副本，智元派出了6位性格各异的G2「实习生」。
它们不仅分工明确，甚至被赋予了真实的职场人设：
咖少：早八人续命主理人。精准操作咖啡机，负责饮品制作。
阿豪：社牛门面担当。主攻语言交互，接待顾客，提供情绪价值。
火宝：无情搬砖机。负责取货与打包，效率至上，动作利落。
倦王：摸鱼式理货员。负责全店巡视与货架补货，默默填补空缺。
冷姐：赛博卫生委员。专属技能是清理桌面，保持就餐区整洁。
叮酱：热食区投喂官。负责加热熟食，精准掌控火候与出餐。
这6位实习生，在11月2日早上10点准时打卡。
迎接它们的，不是单次任务的Demo演示，而是连续24小时、四班倒的魔鬼拉练。
午间高峰（13:00-14:25）：周边白领涌入，订单高并发。取货、打包、出餐必须一气呵成。
晚高峰（19:00-20:00）：经过白天的消耗，硬件与算力迎来疲劳期。
夜班实验（22:00-06:55）：没有人类店长在旁，机器人需要独立值守全店。
早高峰（07:00-09:00）：这是「咖少」与「叮酱」的终极战场。咖啡与热食的需求量达到顶峰，对机械臂的执行精度和节拍速度，将是一场极限压榨。
为了让这场考核绝对真实，直播现场设置了严苛的「考官天团」。
主播化身「HR总监」，死盯四大转正KPI：连续在岗时间、真实服务数量、四班次抗压能力、顾客好评率。
技术专家坐镇「带教Mentor」，各路达人则扮演难搞顾客，随时抛出刁钻难题。
干得好，现场贴纸点赞；干得差，弹幕公屏处刑。
从「实验室里单次完成」到「真实便利店连续运转」，面对连续客流、突发需求和长时间物理磨损，它们能否不掉线、不摸鱼？
这家独特的便利店一开始营业，就迎来了不少人类的围观，很多小姐姐来合影。
毕竟，我们梦想许久的「人类享受，机器人干活」的画面终于来了，看起来极度舒爽。
而整整24小时的直播中，我们见证了不少惊心动魄的场面，看到机器人险些翻车。
好在，他们大多数都顺利完成了任务！
比如在取货时，火宝会自己识别挂钩角度，以巧妙的角度抬起工具，方便自己抓取，最终稳稳地把货物放到篮子里。
然后，它丝滑地给人类送了过来。
有趣的是，技术人员爆料：发现火宝干活时会偷懒，所以给它加了定时巡检的功能。
接着，就是冷姐登场了。
它识别出来，旁边桌子是干净的，所以略过了。来到充满人类垃圾的桌前，它挑拣出该扔的垃圾，开始清理。
遗憾的是，冷姐在这个过程中被小扣了几分，因为误把人类的食物当垃圾扔掉了。最终，它会打扫到认为桌面没有垃圾，才结束。
直播了大半天，到快晚上的时候，智元具身研究中心副总裁任广辉带着外国小哥体验官来探店了。
火宝给送来一瓶可乐，但是动作小扣分——递东西时不够轻柔。
什么任务对火宝最难？广辉老师透露说，就是开柜子。
人类的小孩，可能到三四岁才会开柜子，而机器人跟小孩子一样，一开始只会抓握，经过训练后才能学会的开柜门。
在便利店门口，社牛门面阿豪侃侃而谈，一会飙英文，一会说上海话，把小哥逗得哈哈大笑。
辛勤的倦王全场巡视，货物放得不对，就会立刻纠正。
外国小哥很皮地把口香糖调换了位置，倦王火眼金睛，立马发现不对，用夹爪重新摆放了货物，在背后，都是Genie Master大脑在指挥。
随后，小哥来到咖少这里，人生第一次尝试了机器人给做的咖啡。
旁边的叮酱，轻车熟路地给小哥递上了热包子。
外行看热闹，内行看门道。
观众在直播间看到的是：机器人咖少平稳地做了一杯咖啡，火宝把一袋薯片装进了塑料袋。
动作看似简单，但在技术圈，这叫「跨越鸿沟」。
这就是著名的「莫拉维克悖论」 ：对人类来说极难的高阶逻辑推理，对计算机来说易如反掌；但对人类来说极易的无意识技能，比如抓取软体物体、感知空间，对机器人来说却难于登天。
以「给货架补货」为例。
当「倦王」走向货架，它至少要完成三大任务。
首先，它需要空间理解与环境感知。
便利店的光线在变，顾客在随意走动，货架上的商品可能被放歪了。它必须通过视觉模型，实时构建3D空间，认出「这是薯片，那是挂钩」。
其次，它需要任务规划。Agent大脑要决定先拿哪一包，用什么角度拿，拿到之后手腕怎么旋转。
最后，它需要动作执行与反馈。底层的大模型（如GO-3 Preview）将规划转化为机械臂的扭矩和关节角度，并根据抓取时的触觉和视觉反馈，实时微调力度。
如果没有全栈协同，机器人只会像个瞎子一样乱抓，或者直接撞翻货架。
机器人真正上岗，绝不是靠攒几个硬件、套一个开源模型就能实现的。它需要底层算力、通用大模型、具身智能体框架的深度咬合。
一个便利店里的普通动作，背后一整套全栈技术体系。机器人大规模落地，靠的是全栈协同：
完成任务，需要模型与本体配合；
稳定运行，需要适应环境、处理异常；
规模化应用，还要降低训练成本、加快部署、复用能力。
真正的上岗，绝非拼凑硬件。
智元AI WAVE全栈布局的核心价值就在于此：把便利店作为应用入口，展现冰山下的庞大系统：数据、平台、本体、Agent、仿真、模型，缺一不可。
AGIBOT WORLD 2026 第四期，首个真实场景的 Physical Agent 数据集，专攻决策，覆盖九类逻辑推理场景；
作业平台Genie Studio 2.0完整覆盖「采‑训‑测‑推‑编‑管」全生命周期，最快7天即可完成全新作业场景从0到1的部署验证；
重载具身工业机器人精灵 G2 Max，在行业内率先实现长程、多任务作业；
Genie Master 让机器人在真实场景中持续进化；
Genie Sim 仿真平台承担后训练和部署前的预演验证；
GO-3 Preview 是智元新一代基础模型，采用具身原生的WRAM(World Reasoning Action Model)架构，将任务推理、未来预测与动作生成连接起来，可利用少量示范进一步学习，完成长程精细任务；
群体经验学习（CEL）范式，打通经验获取、经验利用、规模化在线学习完整链路，把真实作业中的每一次成功与失败都变成群体共同的学习素材。
在真实世界里，错误无法回避，没有不出错的机器人。
如果部署100家便利店，需要100支工程师团队去现场当「保姆」，那机器人有何商业价值？
如何打破这个魔咒？
智元给出的杀手锏是：Genie Master具身Agent框架。
10月27日，智元正式发布了这款Genie Master框架。它的核心使命只有一个：让机器人实现「边上岗，边解锁新技能」。
它不仅能组织长程任务、维护进度，还能在异常处理后自主接续任务。
当G2遇到没见过的状况，它不需要人类来救场。它能自己定位失败原因，自己调整策略。
它是怎么做到的？
核心在于，Genie Master框架在机器人的「大脑」和「肢体」之间，搭起了一座桥梁——具身执行协议 EEP（Embodied Execution Protocol）。
肢体执行的结果与反馈，将直接成为后续决策与能力优化的燃料。
过去，AI大脑和机械肢体相互割裂。
大脑下达指令「把货挂上去」，肢体照做。但要是没挂上，肢体无法告诉大脑「为什么没挂上」，大脑也无法给出更微观的指导。
执行完全是开盲盒。
EEP协议打通了任督二脉。它统一了「能力、引导、反馈、连续性」四类接口。
现在，肢体的执行结果会实时、精准地反馈给大脑。
一次失败，能不能让机器人下一次做得更好？这就涉及这场直播最硬核的技术落点：Physical RSI，从部署到迭代的持续进化。
在Genie Master框架和底层模型GO-3 Preview的支撑下，智元为机器人打造了独有的双线进化机制。
一条路线让机器人「记住经验」；另一条路线让机器人「长出肌肉记忆」。
第一条进化路线：Harness RSI（让经验可积累）
你可以把它理解为机器人的「工作日记」与「大脑记忆库」。
在白天的打工过程中，G2会将成功与失败的案例随做随记。哪里做得顺，哪里卡壳了，全部以数据形式保留。
当深夜便利店打烊，系统背后的大模型开始工作。它会统一总结一天的数据，把其中可复用的规律提取出来，沉淀进长期记忆中。
以「挂钩被挡住」的案例为例。经过夜间的总结，系统生成了一条可复用的策略提示：「遇到视线受阻的细小目标，优先调用腕部相机进行辅助定位。」
第二天，当另一台机器人去挂毛巾、挂耳机线时，它不需要再经历「失败-分析-调整-重试」的试错过程。
第二条进化路线：Policy RSI（让能力可固化）
在这次执行与恢复的过程中，系统产生了大量包含「问题、调整、结果」的真实反馈数据。
这些高质量的「纠错数据」经过后处理，会被直接回流，用于重新训练动作策略模型。
有效的改进不再仅仅是一条挂在大脑里的「提示语」，而是深深烙印在了模型的底层权重中，转化为了后续可直接调用的「基础操作能力」。
原本需要大脑「思考分析」才能完成的恢复动作，现在变成了底层模型的「肌肉记忆」。
两条路线，双剑合璧。
昨天失败的记录，是今天进化的养料；今天工作的经验，是明天进化的起点。
它证明了机器人并非只会把预设的流程跑得更熟，而是真正学会了「如何改进自己」。
具身智能的落地价值，不取决于资本市场PPT画得有多美，也不取决于宣传视频剪得有多炫。
它只取决于三件事：
真实任务能否稳定完成？人工干预成本能否降下来？改进能力能否迁移到新任务？
11月2日开启的这场24小时直播，就是要把这三个问题，毫无保留地撕开给所有人看。
从「如何上岗」到「上岗后如何持续进步」。
把商品精准挂上货架，把一杯热咖啡平稳递出，把掉落的残渣清理干净。
这些对人类而言微不足道的动作，是机器人跨越半个世纪的技术狂奔。
这次直播证明了机器人如何把错误变成进化的阶梯，我们将见证一个真正属于具身智能的「打工时代」，硬核开启。
编辑：Aeneas 大卫
不喜欢
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。
New Ji-won.
In these two days, the robot world is "the heart-momenting offer." It's out of the loop!
Over the past 24 hours, more than 2 million online supervisors have been placed on the Internet, covering six Jigen elves G2 in real convenience stores. At 10:00 this morning, the robot "new employees" officially ended their 24-hour "go-round" challenge!
24 hours of live air, we've been tracking down.
This live broadcast is just to answer one question — can robots do it in the face of the chaos of the real world?
On 2 November, Ji-won decided to fight the battle.
Six elves G2 robots with the latest Agent framework, Genie Master, opened a 24-hour "referral test" in their capacity as "practition workers at convenience stores".
All the way live, all the supervisors. No filters, no clips.
Just now, there's a result -- challenge to recruit, four successful "go-round"!
"Bob jobs can't stop," a 24-hour convenience store, which is a direct record.
Six Elf G2 robots performed a cumulative total of 1730 missions, including 683 customer receptions, 369 commodity deliveries, 147 coffee productions, 126 breakfast services, 277 catch-up inspections and 128 desktop cleaning.
This live broadcast marked a watershed — intelligently leaping out of a virtual environment, a laboratory, working in a real scene.
Putting robots in a convenience store is a dangerous decision.
Because less than 100 square metres, it's all physical chaos in the real world: machine arms have to grab hard bottles, squeeze soft bags, piercing; systems have to be filled with supplies, make coffee, drag land; and customers can walk, interject, change orders.
Here, more deadly is multi-machine conflict.
The patch is parked in front of the shelf, blocking the customer ' s access to drinks; the cleaning machine goes into the narrow and blocks the offline. In the wrong movement, the machine was "dead lock" in the hallway, and the door shop collapsed.
This is why it is mandatory for "24 hours of continuous work". Instead of measuring batteries, only to magnify the hazard over time: how can an avalanche be rolled in a day for an impervious carton in 10 minutes? The machine's dead. How can the system line up?
In order to catch the copy, Ji-won sent six different G2 "interns".
They not only have a clear division of labour, but have even been given real jobs:
Quaker: Eight people left in the morning. Precision of the coffee machine and preparation of drinks.
Ah Hao: I'm in charge. Main language interaction, hospitality and emotional value.
Fireballs: The heartless lifting of bricks. It collects and packs, works efficiently and works well.
The Tired King: Fish-carrying barter. Responsible for shop-wide inspection and shelf replenishment, filling vacancies in silence.
Leng Sister: Cyber Health Commissioner. The exclusive skill is to clean the desktop and keep the dining area clean.
Clinking sauce: Hot food area feeding officer. It is responsible for heating and cooking, with precision control of the fire and meals.
The six interns were given a card promptly at 10 a.m. on 2 November.
It's not a single mission, Demo demonstration, but a 24-hour, four-shift devil.
The midday peaks (13:00-14:25): the influx of surrounding white collars, with high orders issued. The goods must be collected, packed and eaten.
Late peak (1900-2000): after daytime consumption, exhaustion of hardware and arithmetic occurs.
Night work experiments (22:00-06:55): Without the head of a human store, robots need to guard the shop independently.
Early peaks (07:00-09:00): This is the ultimate battleground between curry and clink sauce. The demand for coffee and hot food peaks, and the execution precision and beat speed of the arm will be a maximum squeeze.
In order to make this test absolutely real, the live broadcast set up a tough "Citizens."
The main broadcaster, Director-General HR, is focused on the four main shifts of KPI: continuous working hours, number of real services, four shifts of resistance, customer ratings.
The technologists sit in town "to teach Mentor", while the Lutherans play a hard-to-reach role and throw themselves out.
Well done, stickers on the spot; bad work, execution on screen.
From "one-time work in a laboratory" to "real convenience store" running, can they not lose line or touch fish in the face of continuous traffic, sudden demand and prolonged physical wear?
The unique convenience store opened at the beginning of its life and brought a lot of humanity and a lot of little sisters to take pictures.
After all, the picture of what we've dreamed of for so long, "Enjoy humans, work robots" finally comes, and it looks like it's really good.
And in 24 hours of live air, we've witnessed a lot of excitement, and we've seen robots turn over.
Fortunately, most of them did their job well!
For example, when picking up a shipment, Fireball identifies its own angle of linkage, lifts the tools in a clever way, makes it easier for itself to catch them, and eventually places the goods in the basket.
Then it slipped over to humans.
Interestingly, the technologists came up with the idea that a firebath would be lazy when it was found working, so it was given a time-check function.
Then the cold sister showed up.
It identified it, the table next to it was clean, so it slipped. It comes to a table full of human garbage, and it picks out the garbage to throw and starts cleaning.
Unfortunately, Leng had been slashed a few points in the process because of the mistaken disposal of human food as garbage. In the end, it does not end until it thinks the desktop is empty.
For half of the day, the Vice-President of the Center for the Study of Ji-won's Personalities, Wang Fai, was visiting the store with a foreign brother.
Fireballs bring a bottle of Coke, but the action points -- not soft enough to deliver.
What mission is the hardest for Fireball? Mr. Hiromi revealed that it was the opening of the cabinet.
Human children may not open cabinets until they're three or four years old, and robots, like children, can only hold them at first and learn to open doors after training.
In front of the convenience store, the club's face, Ah Hao, was spoken in English, the Shanghai language was spoken, and my brother laughed.
The hard-working, tired King was on the road, and the goods were put in the wrong place, and they were corrected immediately.
The foreign boy changed his position in a pitiful fashion, tired of the King's eyeballs, found it wrong, repositioned the goods with his claws, and behind him was Genie Master's brain.
Then he came here, and for the first time in his life, he tried a robotic coffee.
Clinky sauce next to the car handing over hot buns.
Out-of-the-box, inside doorway.
The audience saw the robot curry making a smooth cup of coffee and the flamingo putting a bag of chips in a plastic bag.
The move seems simple, but in the technology circles, it's called "crossing the divide."
This is the famous Moravik paradox: very difficult high-level logic for human beings is easy for computers; but easy unconscious skills for humans, such as capturing soft objects, sensory space, are difficult for robots.
For example, "for the shelf".
When the Lord is on the shelf, it has at least three tasks.
First, it requires spatial understanding and environmental awareness.
The light in the convenience store is changing, the customers are moving at will, and the goods on the shelf may have been diverted. It has to build 3D space in real time through visual models, recognizing "this is chips, that's hooks."
Secondly, it requires mission planning. Agent's brain has to decide which bag to take first, which angle to take, and how to rotate the wrist after that.
Finally, it needs action implementation and feedback. The bottom LLM (e.g. GO-3 Preview) translates planning into a twist and joint angle of a mechanical arm and fine-tunes in real time based on touch and visual feedback at the time of capture.
Without full co-operation, robots would just scratch like blinds or crash the shelf.
A robot's real in-service is by no means possible by saving a few hardware, with an open source model. It requires a bottom-up calculator, a generic LLM, and a deep AI agent frame.
A common move in a convenience store, with a whole technology system behind it. The robot landed on a massive scale, thanks to the whole house:
(a) To fulfil its mandate by requiring model and physical alignment;
Stable operation, which requires adaptation to the environment and handling anomalies;
Scaled applications also need to reduce training costs, accelerate deployment and reuse capacity.
Real induction is not a piece of hardware.
That is the core value of the Jiwon AI WAVE layout: using the convenience store as an application portal to display the vast system under the iceberg: data, platform, body, Agent, simulation, model.
AGIBOT WORLD 2026, No. 4, First Real Age data set, dedicated to decision-making, covering nine types of logical reasoning;
The platform, Genee Studio 2.0, fully covers the entire life cycle of the "Technology" and completes the deployment validation of the new operating scenario from 0 to 1 in a maximum of 7 days;
G2 Max, a reloading industrial robot, takes the lead in long-range, multi-mission operations within the industry;
Genee Master keeps robots evolving in the real scene.
Training and pre-deployment pre-certification of the Genie Sim simulation platform;
GO-3 Preview is the basic model of the new generation of intellectuals, using the original WRAM (World Reasing Action Model) structure, linking mission reasoning, future predictions and action generation, with a small amount of demonstration to further study and perform long-range fine tasks;
The CEL model provides a complete link to learning through experience acquisition, experience utilization and scale-up online, and translates every success and failure in a real operation into learning asset in community.
In the real world, mistakes cannot be avoided, there is no wrong robot.
What is the commercial value of robots if 100 convenience stores are deployed and 100 teams of engineers are required to be on the scene as “sitters”?
How to break this spell?
The killer given by Ji-won is Gene Master with the Agent framework.
On 27 October, Ji-won officially introduced the Genie Master Framework. It has only one mission at its core: to get robots "on board, unlock new skills".
It not only organizes long-range missions, maintains progress, but also autonomously succeeds after an anomaly.
When G2 encounters an unknown condition, it does not need a human rescue. It can figure out why it failed, adjust its own strategy.
How did it do that?
The core is that the Genie Master framework creates a bridge between the robot's "brain" and "body" - the EEP (Embodied Exchange Production) protocol.
Results and feedback from physical implementation will be a direct fuel for subsequent decision-making and optimization.
In the past, AI's brain and mechanical limb were cut apart.
The brain gives the order "to hang the goods", and the body does it. But if you don't, you can't tell the brain why you don't, and you can't give more microdirection.
The execution was completely blind.
The EEP agreement provided a second line of supervision. It harmonizes the four types of interface: "Capacity, guidance, feedback, continuity".
Now, the results of body execution are fed back to the brain in real time and with precision.
One failure. Can you make the robot do better next time? This relates to the strongest technology site on the air: Physical RSI, which is continuously evolving from deployment to iterative times.
With the support of the Genie Master framework and the bottom model GO-3 Preview, Jigen created a unique dual evolutionary mechanism for robots.
One route allows robots to "remember experience" and the other to "remember muscle."
First evolutionary route: Harness RSI (to allow experience to accumulate)
You can interpret it as the robot's work diary and brain memory.
During daytime work, the G2 keeps a record of successes and failures. It's good. It's good. It's good.
When late-night convenience stores closed, LLM, behind the system, began work. It summarizes the data for one day, extracts the patterns that can be used again and settles into long-term memory.
For example, in the case of the "coup-blocked" case. After a summary at night, the system produces a reusable policy hint: "Stick targets are blocked by sight and priority is given to the use of wrist cameras to assist in locating."
The next day, when another robot went to wear a towel and a headset line, it no longer had to go through a "failure-analysis-adjustment-retry" error.
Second evolutionary route: Policy RSI (to make capacity solid)
During this implementation and recovery process, the system generated a large number of real feedbacks containing "problems, adjustments, results".
These high-quality "correcting data" are processed and re-routed for re-training in action strategy models.
The effective improvements are no longer just a "tip" hanging in the brain, but are deeply branded in the bottom weight of the model, which translates into a "basic operating capability" that can then be directly used.
The recovery action that would have required brain "thinking analysis" has now become the "mustual memory" of the bottom model.
Two paths, two swords.
The record of yesterday ' s failure is today ' s evolutionary nutrients; the experience of today ' s work is the beginning of tomorrow ' s evolution.
It proves that robots don't just get to know the process better, but they really learn how to improve themselves.
Smart landing values do not depend on how beautiful the capital market PPT is, or how fancy the publicity video is.
It depends on three things:
Can real task stabilize? Can the cost of manual intervention be reduced? Can improved capacity be relocated to new mandates?
This 24-hour live broadcast, which was launched on 2 November, was intended to tear these three issues apart for everyone without reservation.
From "how to get on duty" to "how to make progress after you get on duty."
Put the merchandise on a shelf, smooth out a hot cup of coffee and clean up the fall-off.
These insignificant actions for humans are robots running through half a century of technology.
This live broadcast shows how robots turn mistakes into evolutionary ladders, and we will witness a truly smart "working age" with hard cores.
Editor: Aeneas David
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.