[{"data":1,"prerenderedAt":290},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"model-gemini-robotics":9},{"tools":4,"reviews":5},78,26,{"tools":4,"reviews":5,"playbooks":7,"news":8},22,24,{"id":10,"title":11,"apiCompatible":12,"benchmarks":14,"body":21,"category":258,"contextWindow":259,"cover":260,"description":261,"extension":262,"lastVerified":260,"maxOutput":263,"meta":264,"navigation":265,"path":266,"pricing":267,"published":268,"relatedTools":269,"releaseDate":270,"seo":271,"slug":272,"stem":273,"strengths":274,"updated":268,"useCases":279,"vendor":284,"vendorEn":284,"weaknesses":285,"__hash__":289},"models\u002Fmodels\u002Fgemini-robotics.md","Google Gemini Robotics",[13],"google",[15,18],{"name":16,"score":17},"语言指令零样本泛化","优于传统分层控制（定性）",{"name":19,"score":20},"本机推理延迟","显著低于云端往返（On-Device 版）",{"type":22,"value":23,"toc":243},"minimark",[24,28,40,43,59,62,67,76,80,92,96,99,102,117,120,204,207,210,213],[25,26,27],"h2",{"id":27},"概述",[29,30,31,35,36,39],"p",{},[32,33,34],"strong",{},"Gemini Robotics"," 是 Google DeepMind 把 Gemini 多模态能力\"装进机器人身体\"的成果。它解决的核心问题是：让机器人",[32,37,38],{},"听懂一句人话 + 看懂眼前画面，然后自己决定怎么动","——而不是工程师为每一个动作预先写死控制代码。",[29,41,42],{},"基于 Gemini 2.0 的视觉-语言底座，它能做到两件传统机器人做不到的事：",[44,45,46,53],"ol",{},[47,48,49,52],"li",{},[32,50,51],{},"现场生成控制代码","：面对没专门编程过的场景，模型即时产出可执行的机器人指令；",[47,54,55,58],{},[32,56,57],{},"本机本地推理（On-Device）","：2026 年推出的 Gemini Robotics On-Device 把模型直接跑在机器人自身的算力上，断网也能工作，延迟更低、在真实部署里更稳。",[25,60,61],{"id":61},"核心能力",[63,64,66],"h3",{"id":65},"多模态大脑","多模态\"大脑\"",[29,68,69,70,75],{},"和纯运控机器人不同，Gemini Robotics 继承了 Gemini 对图像、视频、语言的原生理解。你用自然语言下指令，它结合摄像头画面做语义决策——这正是 ",[71,72,74],"a",{"href":73},"\u002Fwiki\u002Fvla.html","VLA 模型"," 想达成的\"端到端统一\"。",[63,77,79],{"id":78},"生成式控制code-on-the-fly","生成式控制（code on the fly）",[29,81,82,83,86,87,91],{},"最被看重的特性是：",[32,84,85],{},"模型能为新任务实时生成控制代码","。这意味着\"教机器人做一件新事\"从\"养团队半年\"变成\"说一句话\"，是 ",[71,88,90],{"href":89},"\u002Fwiki\u002Frobot-foundation-model.html","机器人基础模型"," 泛化能力的直接体现。",[63,93,95],{"id":94},"on-device-本地推理","On-Device 本地推理",[29,97,98],{},"云端推理在工厂 \u002F 家庭里有两个致命问题：延迟、断网即瘫。Gemini Robotics On-Device 把模型下沉到机器人端，是 2026 年具身智能从 demo 走向部署的关键一步（同类思路也见于 NVIDIA Jetson Thor）。",[25,100,101],{"id":101},"在产业里的落点",[103,104,105,111],"ul",{},[47,106,107,110],{},[32,108,109],{},"波士顿动力 Atlas","：2026 年进入量产的新一代 Atlas 集成了 Google Gemini，作为\"大脑\"驱动决策；",[47,112,113,116],{},[32,114,115],{},"研究生态","：作为开放能力被多家机器人本体与高校采用，与 NVIDIA GR00T、Figure Helix 同列 2026 年主流技术路线之一。",[25,118,119],{"id":119},"与其他机器人模型的定位差异",[121,122,123,144],"table",{},[124,125,126],"thead",{},[127,128,129,133,135,141],"tr",{},[130,131,132],"th",{},"维度",[130,134,34],{},[130,136,137],{},[71,138,140],{"href":139},"\u002Fmodels\u002Fisaac-gr00t.html","NVIDIA Isaac GR00T",[130,142,143],{},"Figure Helix",[145,146,147,162,176,190],"tbody",{},[127,148,149,153,156,159],{},[150,151,152],"td",{},"主体定位",[150,154,155],{},"模型 + 生态",[150,157,158],{},"开放基础模型 + 全栈工具链",[150,160,161],{},"双系统本体智能",[127,163,164,167,170,173],{},[150,165,166],{},"强项",[150,168,169],{},"多模态理解 + 生成控制代码",[150,171,172],{},"跨本体训练 + 仿真 + Scaling Law",[150,174,175],{},"双机协作、动态环境",[127,177,178,181,184,187],{},[150,179,180],{},"推理位置",[150,182,183],{},"云端 \u002F 本机 On-Device",[150,185,186],{},"本机（Jetson Thor）",[150,188,189],{},"本机",[127,191,192,195,198,201],{},[150,193,194],{},"开放度",[150,196,197],{},"企业授权为主",[150,199,200],{},"开放权重 + 工具链",[150,202,203],{},"闭源",[25,205,206],{"id":206},"国内使用",[29,208,209],{},"与所有 Google 模型一样，Gemini Robotics 在大陆无官方直连，需 Vertex AI 或中转，且机器人本体侧还要解决端算力与部署合规。研究团队多用海外云 + 真机 \u002F 仿真（Isaac Sim）组合。",[25,211,212],{"id":212},"延伸阅读",[103,214,215,221,226,231],{},[47,216,217,218],{},"技术路线：",[71,219,220],{"href":73},"VLA 模型（视觉-语言-动作）",[47,222,223,224],{},"能力底座：",[71,225,90],{"href":89},[47,227,228,229],{},"同赛道：",[71,230,140],{"href":139},[47,232,233,234,238,239],{},"行业背景：",[71,235,237],{"href":236},"\u002Fwiki\u002Fembodied-ai.html","具身智能"," \u002F ",[71,240,242],{"href":241},"\u002Fnews\u002F2026\u002Fembodied-ai-h1-2026.html","2026 上半年具身智能十大进展",{"title":244,"searchDepth":245,"depth":245,"links":246},"",3,[247,249,254,255,256,257],{"id":27,"depth":248,"text":27},2,{"id":61,"depth":248,"text":61,"children":250},[251,252,253],{"id":65,"depth":245,"text":66},{"id":78,"depth":245,"text":79},{"id":94,"depth":245,"text":95},{"id":101,"depth":248,"text":101},{"id":119,"depth":248,"text":119},{"id":206,"depth":248,"text":206},{"id":212,"depth":248,"text":212},"multimodal",1000000,null,"Google DeepMind 基于 Gemini 2.0 的机器人多模态模型家族，能把自然语言指令和摄像头画面直接转成机器人动作；2026 年进一步推出 On-Device 本机推理版本，让模型跑在机器人自己身上。","md",8192,{},true,"\u002Fmodels\u002Fgemini-robotics","通过 Vertex AI \u002F Google AI Studio 按 Gemini 2.0 计价；机器人定制版为企业授权，具体以官方为准","2026-07-23",[],"2025-03-25",{"title":11,"description":261},"gemini-robotics","models\u002Fgemini-robotics",[275,276,277,278],"基于 Gemini 2.0 多模态底座，原生理解图像\u002F视频\u002F语言","能现场生成机器人控制代码（code generation on the fly），不需为每个场景预编程","On-Device 本机本地推理版本，断网也能跑，延迟低、鲁棒性强","与 Google 生态（AndroidXR \u002F Project Astra）协同，面向消费级机器人场景",[280,281,282,283],"未知场景的零样本操作（没训练过的指令也能尝试）","仓储 \u002F 工厂的非结构化抓取与分拣","家庭 \u002F 服务场景的语义级交互（说人话就能指挥）","作为'大脑'驱动波士顿动力 Atlas 等本体","Google DeepMind",[286,287,288],"国内无官方直连，需 Vertex AI 或中转","真实开放世界长尾任务仍弱于专用运控","本机推理对机器人端算力（如 Jetson 级）有门槛","Y70zEk9FGX18ctk7m-vgF7U4IX4Igshh_MXjaDsPvVw",1784900533518]