[{"data":1,"prerenderedAt":362},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"wiki-embodied-ai":9},{"tools":4,"reviews":5},78,26,{"tools":4,"reviews":5,"playbooks":7,"news":8},22,24,{"id":10,"title":11,"body":12,"category":342,"cover":343,"description":329,"extension":344,"lastVerified":343,"meta":345,"navigation":346,"path":347,"published":348,"relatedModels":349,"relatedTools":352,"seo":353,"slug":354,"stem":355,"summary":356,"tags":357,"updated":348,"__hash__":361},"wiki\u002Fwiki\u002Fembodied-ai.md","具身智能",{"type":13,"value":14,"toc":328},"minimark",[15,19,31,37,40,102,106,113,151,158,162,165,170,194,198,201,205,212,243,246,250,253,283,286],[16,17,18],"h2",{"id":18},"什么是具身智能",[20,21,22,26,27,30],"p",{},[23,24,25],"strong",{},"具身智能（Embodied AI）"," 指的是把人工智能\"装进身体里\"——让模型不仅能聊天、写代码，还能在真实的物理世界里",[23,28,29],{},"看、听、理解、然后动手做","。它区别于我们熟悉的\"数字 AI\"（ChatGPT、Claude、Copilot 都只在屏幕里运行），核心差异一句话：",[32,33,34],"blockquote",{},[20,35,36],{},"数字 AI 改变的是信息；具身智能改变的是物理世界。",[20,38,39],{},"一个具身智能系统通常由三层组成：",[41,42,43,59],"table",{},[44,45,46],"thead",{},[47,48,49,53,56],"tr",{},[50,51,52],"th",{},"层",[50,54,55],{},"角色",[50,57,58],{},"对应技术",[60,61,62,76,89],"tbody",{},[47,63,64,70,73],{},[65,66,67],"td",{},[23,68,69],{},"大脑",[65,71,72],{},"理解任务、做决策、规划步骤",[65,74,75],{},"多模态大模型 \u002F VLA 模型 \u002F 世界模型",[47,77,78,83,86],{},[65,79,80],{},[23,81,82],{},"小脑",[65,84,85],{},"把决策翻译成平稳的关节运动",[65,87,88],{},"运控算法 \u002F 强化学习 \u002F 模仿学习",[47,90,91,96,99],{},[65,92,93],{},[23,94,95],{},"身体",[65,97,98],{},"执行，并持续从环境获得反馈",[65,100,101],{},"机器人本体（双足 \u002F 轮式 \u002F 机械臂）+ 传感器",[16,103,105],{"id":104},"为什么-2026-被称为具身智能拐点","为什么 2026 被称为\"具身智能拐点\"",[20,107,108,109,112],{},"2025 年被业界称为\"人形机器人量产元年\"，而 2026 年上半年则是",[23,110,111],{},"从 research demo 走向工业规模部署","的转折点。几个标志性信号（均来自 2026 上半年真实行业事件）：",[114,115,116,123,129,135,141],"ul",{},[117,118,119,122],"li",{},[23,120,121],{},"宇树科技","冲刺 A 股\"具身智能第一股\"，科创板上市委 73 天过会，并被 NVIDIA 选为 Isaac GR00T 开放参考机器人平台；",[117,124,125,128],{},[23,126,127],{},"智元机器人（AGIBOT）"," 2026-03-30 下线第 1 万台人形机器人，从 5000 到 10000 台只用了三个月；",[117,130,131,134],{},[23,132,133],{},"Figure AI"," 以约 $39B 估值完成约 $1B 的 C 轮，并在 BMW  Spartanburg 工厂部署 50+ 台 Figure 02\u002F03；",[117,136,137,140],{},[23,138,139],{},"NVIDIA"," 发布 GR00T N1.7\u002FN2，首次提出\"机器人灵巧度 Scaling Law\"，并推出 Cosmos 3 世界模型与 Newton 1.0；",[117,142,143,146,147,150],{},[23,144,145],{},"波士顿动力 Atlas"," 进入量产，集成 Google Gemini；",[23,148,149],{},"Google DeepMind"," 推出可在机器人本机本地运行的 Gemini Robotics On-Device。",[20,152,153,154,157],{},"黄仁勋在 CES 2026 的主题演讲中直接宣布：\"机器人领域已正式迎来属于自己的'ChatGPT 时刻'。\" 这句话被反复引用，因为它点破了关键变化——",[23,155,156],{},"能力开始可复制、可规模化，而不是每换一个任务就重新写一套控制代码","。",[16,159,161],{"id":160},"大脑-vs-小脑具身智能的两大技术栈","大脑 vs 小脑：具身智能的两大技术栈",[20,163,164],{},"理解具身智能，先理解\"大脑\"和\"小脑\"的分工：",[166,167,169],"h3",{"id":168},"大脑让机器人懂事","大脑：让机器人\"懂事\"",[20,171,172,173,178,179,183,184,188,189,193],{},"由 ",[174,175,177],"a",{"href":176},"\u002Fwiki\u002Fvla.html","VLA 模型（视觉-语言-动作）"," 与 ",[174,180,182],{"href":181},"\u002Fwiki\u002Frobot-foundation-model.html","机器人基础模型"," 承担。它接收自然语言指令 + 摄像头画面，输出\"下一步该做什么\"的语义决策。代表：Google ",[174,185,187],{"href":186},"\u002Fmodels\u002Fgemini-robotics.html","Gemini Robotics","、NVIDIA ",[174,190,192],{"href":191},"\u002Fmodels\u002Fisaac-gr00t.html","Isaac GR00T","、Figure Helix、宇树 WVLA 2.0。",[166,195,197],{"id":196},"小脑让机器人做得稳","小脑：让机器人\"做得稳\"",[20,199,200],{},"负责把\"抓取杯子\"这种高层指令，转成几百个关节电机在毫秒级的协同运动。传统做法是手写运控，现在越来越多由模仿学习 \u002F 强化学习从数据里\"练\"出来。续航、负载、平衡都属于这一层——也是当前最现实的瓶颈（主流人形机器人单次充电只能 active 运行 2–4 小时）。",[16,202,204],{"id":203},"具身智能与本站关注的-ai-coding-是什么关系","具身智能与本站关注的 AI Coding 是什么关系",[20,206,207,208,211],{},"这俩不是两件事，而是",[23,209,210],{},"同一波浪潮的两面","：",[213,214,215,226,237],"ol",{},[117,216,217,220,221,225],{},[23,218,219],{},"造机器人需要写大量代码","：仿真（Isaac Sim \u002F Cosmos）、数据采集 pipeline、模型训练、OTA 升级，全靠 ",[174,222,224],{"href":223},"\u002Fcoding\u002Fcli\u002Fclaude-code.html","AI 编程工具"," 提效；",[117,227,228,231,232,236],{},[23,229,230],{},"机器人本身是 AI 编程的新执行环境","：Agent 不再只改 git 仓库，而可以\"在真实工厂里搬箱子、在机场地勤作业\"——这是 ",[174,233,235],{"href":234},"\u002Fwiki\u002Fagentic-coding.html","Agentic Coding"," 的物理延伸；",[117,238,239,242],{},[23,240,241],{},"VLA \u002F 世界模型 与 LLM 同源","：它们共享 Transformer、共享 Scaling Law 思想，区别只是输出从\"token\"变成了\"动作 \u002F 视频帧\"。",[20,244,245],{},"AIHO 把\"AI Coding 与具身智能\"作为两大支柱，正是因为它们共享同一套底层技术范式，只是落地形态一个在数字世界、一个在物理世界。",[16,247,249],{"id":248},"现实瓶颈别被-demo-带偏","现实瓶颈（别被 demo 带偏）",[20,251,252],{},"行业热，但落地有温差：",[114,254,255,265,271,277],{},[117,256,257,260,261,264],{},[23,258,259],{},"商业满意度低","：Morgan Stanley 对潜在工业买家的调研显示，仅 ",[23,262,263],{},"23%"," 对现有产品满意；",[117,266,267,270],{},[23,268,269],{},"续航硬约束","：2–4 小时 active 续航是普遍现状，固态电池量产才可能把阈值拉到 6–8 小时；",[117,272,273,276],{},[23,274,275],{},"长尾任务弱","：日常基础任务表现不错，但开放世界的罕见指令 \u002F 极端环境适配仍是短板；",[117,278,279,282],{},[23,280,281],{},"数据飞轮刚启动","：真实场景训练数据稀缺，仿真（如 GR00T-Dreams）能补约 40% 成功率，但仍靠真实数据兜底。",[16,284,285],{"id":285},"延伸阅读",[114,287,288,293,298,308,319],{},[117,289,290,291],{},"核心技术路线：",[174,292,177],{"href":176},[117,294,295,296],{},"通用能力底座：",[174,297,182],{"href":181},[117,299,300,301,304,305],{},"代表模型：",[174,302,303],{"href":186},"Google Gemini Robotics"," \u002F ",[174,306,307],{"href":191},"NVIDIA Isaac GR00T",[117,309,310,311,304,315],{},"行业进展：",[174,312,314],{"href":313},"\u002Fnews\u002F2026\u002Fembodied-ai-h1-2026.html","2026 上半年具身智能十大进展",[174,316,318],{"href":317},"\u002Fnews\u002F2026\u002Funitree-ipo-2026.html","宇树 IPO：具身智能第一股",[117,320,321,322,304,326],{},"数字世界侧的同源技术：",[174,323,325],{"href":324},"\u002Fwiki\u002Fai-agent.html","AI Agent",[174,327,235],{"href":234},{"title":329,"searchDepth":330,"depth":330,"links":331},"",3,[332,334,335,339,340,341],{"id":18,"depth":333,"text":18},2,{"id":104,"depth":333,"text":105},{"id":160,"depth":333,"text":161,"children":336},[337,338],{"id":168,"depth":330,"text":169},{"id":196,"depth":330,"text":197},{"id":203,"depth":333,"text":204},{"id":248,"depth":333,"text":249},{"id":285,"depth":333,"text":285},"concept",null,"md",{},true,"\u002Fwiki\u002Fembodied-ai","2026-07-23",[350,351],"gemini-robotics","isaac-gr00t",[],{"title":11,"description":329},"embodied-ai","wiki\u002Fembodied-ai","让 AI 拥有身体、在物理世界里感知-决策-行动的技术范式：大模型'大脑' + 运控'小脑' + 机器人本体，是 AI 从数字世界走向物理世界的载体。",[11,358,359,360,182],"Embodied AI","机器人","VLA","PZXcNZparNNIz_Pl-_dguX2QTACB3B8MbS9IHXDji30",1784900539377]