Xiaomi Is Testing a Robot That Tidies Its Own Staff Apartments
小米让机器人在自家员工公寓里收拾桌子
小米把 7 月发布的具身基座模型投入自家员工公寓试运营：机器人可自主完成桌面整理、垃圾清扫、物品归位与房间巡检，连续作业超 30 分钟、覆盖 3 个房间、50 多个任务。
A model released in July is now doing desk tidying, rubbish clearing and room checks in Xiaomi's staff housing — more than 30 minutes of continuous work across three rooms and 50-plus tasks.
Xiaomi has started using a robot in its own staff apartments to tidy desks, clear rubbish and put things back where they belong — a modest-sounding job that is, for machine learning, unusually hard. The company published a video of the trial on 3 November.
The system is built on Xiaomi-Robotics-1, an “embodied foundation model” released in July and open-sourced in August. Where a chatbot predicts text, an embodied model predicts movement: it takes camera images and a spoken instruction, and outputs the joint commands that carry them out. Xiaomi says the loop — collecting data, running inference, executing the task — held up for more than 30 minutes across three rooms and more than 50 separate tasks without a reset.
That duration is the point. A robot that picks up one cup is a demonstration; a robot that keeps working for half an hour is a system, because errors compound. Every failed grasp leaves the room slightly more cluttered than the model expects to find it.
The training numbers are large. Xiaomi says the model was pretrained on 100,000 hours of real-world manipulation footage — 1,700-odd scenes, 2.4 million interaction clips and more than 260 task types, drawn from homes, offices, restaurants, outdoor settings and factories — then aligned on over 10,000 hours recorded from different robot bodies. The upshot, the company claims, is that a newly deployed robot can follow plain-language instructions, and a genuinely new task needs only a small amount of fine-tuning.
The more convincing evidence may come from a car factory rather than an apartment. On Xiaomi’s electric-vehicle line, robots handle self-tapping nut loading, cover-plate sorting and bin folding at success rates of 94–98%, according to the company — figures that matter because a production line punishes slowness and mistakes alike. Xiaomi says a two-arm nut-loading station improved from 90.2% to 98%.
The model also topped two public benchmarks, RoboCasa 365 and RoboDojo, in July, and picked up an award at China’s World Internet Conference this week.
It is worth keeping the framing honest. A pilot in staff housing is a controlled environment, and the tasks are chosen to be achievable. But the direction is clear: the gap between a robot demo and a robot shift is being closed with data rather than hand-written rules.
中文翻译（全文）
小米开始在自家员工公寓里用机器人收拾桌面、清理垃圾、把东西放回原位——这活听起来不起眼，对机器学习来说却异常困难。公司于 11 月 3 日发布了试运营视频。
这套系统建立在 Xiaomi-Robotics-1 之上，这是一个 7 月发布、8 月开源的「具身基座模型」。聊天机器人预测的是文字，具身模型预测的是动作：它接收摄像头画面和一句口头指令，输出执行该指令所需的关节命令。小米称，从采集数据、模型推理到执行任务这个闭环，在 3 个房间、50 多个不同任务上连续运转了 30 分钟以上，中间没有重置。
时长正是关键。能拿起一只杯子的机器人是演示；能连续工作半小时的机器人是系统——因为误差会累积。每一次抓取失败，都会让房间比模型预期的状态更乱一点。
训练数据量很大。小米称该模型在 10 万小时真实世界操作视频上做了预训练——约 1700 种场景、240 万段交互片段、260 多种任务类型，取材于家庭、办公室、餐厅、户外和工厂——随后又在 1 万多小时来自不同机器人本体的数据上做了对齐。按照公司的说法，结果是：新部署的机器人能听懂自然语言指令，而面对真正的新任务只需少量微调。
更有说服力的证据可能来自汽车工厂，而不是公寓。据小米介绍，在其电动汽车产线上，机器人承担自攻螺母上件、盖板排序、料箱折叠等工序，成功率 94%–98%——这些数字之所以重要，是因为产线对「慢」和「错」同样不容忍。小米称，双机械臂螺母上件工位已从 90.2% 提升到 98%。
该模型 7 月还在 RoboCasa 365 与 RoboDojo 两个公开榜单上排名第一，本周又在中国世界互联网大会上获奖。
不过还是把话说准一点：在员工公寓里试运营，是一个受控环境，任务也是挑过、确保能完成的。但方向很清楚——从「机器人演示」到「机器人上班」之间的差距，正在靠数据而不是手写规则来填。
重点词汇
- embodied /ɪmˈbɒdid/ adj. — existing in a physical body that can sense and act, rather than only in software 具身的 An embodied model predicts movement, not text.
- inference /ˈɪnfərəns/ n. — the process of a trained model producing an answer from new input 推理 The loop runs data collection, inference and execution.
- compound /kəmˈpaʊnd/ v. — to make something worse by adding to it 累积；加剧 Errors compound over a long run.
- grasp /ɡrɑːsp/ n. — an act of taking hold of something with a hand or gripper 抓取 Every failed grasp changes the room.
- manipulation /məˌnɪpjuˈleɪʃn/ n. — handling and moving objects with skill 操作；抓取物体 It was trained on 100,000 hours of manipulation footage.
- fine-tuning /ˌfaɪn ˈtjuːnɪŋ/ n. — further training of an existing model on a small, specific dataset 微调 A new task needs only a little fine-tuning.
- benchmark /ˈbentʃmɑːk/ n. — a standard test used to compare systems 基准测试 It topped two public benchmarks.
- pilot /ˈpaɪlət/ n. — a small trial run before wider use 试点 The apartment deployment is a pilot.
- upshot /ˈʌpʃɒt/ n. — the final result or practical conclusion 结果；结论 The upshot is that deployment is simpler.
- cluttered /ˈklʌtəd/ adj. — filled with things in an untidy way 凌乱的 A failed grasp leaves the room more cluttered.
阅读测验
共 5 道题目
What is Xiaomi's robot doing in its staff apartments?
What does an embodied model predict, unlike a chatbot?
A robot that works continuously for half an hour is described as a demonstration, not a system.
Which details does the article give? (Select all that apply)
A robot that picks up one cup is a demonstration; a robot that keeps working for half an hour is a _____.
评论
0 条讨论
登录后发表评论
立即登录暂无评论
成为第一个分享想法的人吧！