Gemini Robotics-ER 1.5 開發指南

发布于 2025-09-26 23:47:21（微信公众号导出记录）。

本文来自公众号后台的“导出文章内容”功能。博客正文由导出长图进行本地 OCR 转写，并保留原始排版图用于逐段核对。

原文链接：查看原文

OCR 转写有效文字约 17384 字；代码、流程图和版式以文末原始排版图为准。

正文（本地 OCR 转写）

original wex7ce EI V1ajero 2825年9月26日 23:47 新加坡 Gemin1 Robot1cs-ER 1.5 (Gemin1 Robotics-Embodied Reasoning 的桶) 是— 視凭語言模型(VLM),可Gemini的代理功能带入概器人领域。Gemini Robotics-ER

1.5是一细富思考的模型,能翔分析现實世界、原生呼叫工具,以及规割遥朝步骤来完成任

移。

Learnmoreabout Gemini Robotfcs 1.5: 04:13 deepmind.google/robotics

Gem1n1Robotics-ER 1.5與其他Gemin1模型频似,但專為提升概器人感知能力和现置 世界互勤而打造。这项技可解雜的視觉資料、敦行空同推理,根據自然語言指令规劃勤 作,籍此提供进港推理功能,解决實體间题。 在谨作方面,Gemin1Robotics-ER 1.5可奥现有的機器人控制器和行為搭配使用。這项功 能可以依序呼叫概器人的API,模型協這些行為,助器人完成长期任務。 有了Gemin1Robot1cs-ER 1.5,忽就能建横機器人噻用程式,敦行下列励作:

• 使用者透過自然證言措派视工作,轻题使用概器人。

• 镶概器人能狗在闻放式環境中推理、通感及因惠愛化,进而提升自主性。

GeniniRobotics-ER 1.5提供统一模型,適用於各種器人工作:

• 找出撒别物件:

確指出環境中的各稚项目,亚定款定界框。

• 腺解物件酮保:

根擦空醒配置和環项情境进行推理,做出明智决黄、

• 规劃抓取和轨踪

生成扒取黏和款畸,以操控物件。

• 解镇助彪場景

分析影片影格,追辉物件腺解一段時間内的勤作。

• 協调长期任猜

将自然语言指令分解為一系列退和子任,或呼叫现有携器人行為的函式,

• 人機互助

透過文字或語音理解自然語言指令。 開始使用:导找場景中的物件 片和文字提示傳透至模型,以取得已證别物件的清單,以及對鹰的2D贴。模型會德回图片中 别项目的贴,傅回這些项目的標维化2D座槽和标箍。 您可以将這项输出内容搭配提器人API使用,或呼叫视贺語言動作(VLA)模型或任何其他第

三方使用者定教函式,為横器人生成要執行的勤作,

RESTful API 1# First, ensure you have the lnage file locally. # Encode the inage lo base54 IHAGE_BASE64=s(basc64 -w 3 my-inage.png) curl -X P0ST \

• https:/generativelanguage-googleapis.con/vibeta/models/genini-rcba

H "x-goog-api-kcy: $GEMINI_API_KEY" uos/uotaeotidde :ad-suaquo. H-

• d{

]+,sjuojuos, "parts*: [ "inlineData":

• nineType":inage/png"

• data*: "*${INAGE_BASE64]"**

}. { "text*: *Point to no more than 1o itens 1n the image. The

1 “generationconfig”:[ "temperature*: 0.5. "thinkingBudget":

Python fron google.genai inport types Inport IPython fron PIL inpart Inage #Initializc the GenAI client and spccify the model HoDEL_ID = "genini-robotics-er-1.5-preview PROMPT = *** Point to no more than le itcms in the inage. The labcl rctur shoutd be an 1dentifying name for the object detected. The answer should folloa the json fornat: [f"point*: <point> "label*: <label1>), ---]. The points are in [y, x] fornat normalized to 9-1000. client = genai.client() abeuT Jno. peo7 # 1 5 ( bud #eut-Au,)uado*#fewI = 6u ing = img.resize((8ee, int{8gg * ing,size[1] / ing.size[9])), Inage,Re Inage_response = cllent.models.generate_content( nodel=MODEL_ID, 1mg, PROKPT 1, config = types.GenerateContentConf1g( temperature=.5, thinking_config=types, Thinkingconfig(thinking_budget=θ) print (image_response. text) 输出内容是包含物件的JSON障列,每物件都有point(棵化[y,x]座标)和用 於别物件的Label。 {"point": [376. 508]. "label': snall banana}, {"point": [223, 303]. "label':“pink starfruit*}, '(,brq Jaded, {“point": [435, 172]. "label*:* {"point": [z7e, 786], "label′: “green plastic bowl"), {"point": [488, 775]. "label': netal measuring cup"}, {"point": [673, 58e]. "label: dark blue boul}, {"point": [525, 429]. "label′: *lime"}

運作方式 GeniniRobotics-ER1.5可读機器人通用空翻解功能,睡解實腊世界在其中通作。 项功能會接收圆片/影片/音訊醋入内容和自然韬言提示,然後:

• 解物件和場景眠络

:鳞别物件,以及物件具场景的關係(包括可供性)。

• 解工作指令

:解流以自然語言下邃的工作,例如找出香蕉!。

• 空周和時翻推理

:度解动作序列,以及物體如何随时與堤景互动。

• 提供结模化输出

:傳回代表物件位置的座標(贴或定界框)。 遗项技術可旗機器人透過程式辅助「看見及「解周遭瑜境。 Gemin1Robotics-ER1.5也是代理式模型,可將工作(例如“把果故造硫理:)分 解為子工作,以旗长期工作:

• 子工作排序

:将指令分解為符合谨疆的步序列。

• 函式呼叫/程式碼执行

:呼叫現有的楼器人函式/工具或敦行生成的程式碼,雍此敦行步显。 使用Gemini Robotics-ER1.5的思考预算 Gen1n1Robot1cs-ER1.5具有弹性的思考预算,可您控管延还與確度之間的取舍。射 於物件值测等空圈理解工作,模型可以谨用少量思考预算,達到高成效。封龄計数和重量估算等 蛟翘的推理工作,较大的思考预算有助於提升维碰度。這樣一来,您就能在需要低延回应的 同時,兼顾较困鞋任的高確度结果。 機器人代理能力 本将逐步说明GeniniRobotics-ER1.5的各种功能,示前如何將模型用於機器人感 知、推理和规劃愿用程式。 本前鼓例窖展示各项功能,包括在图片中指出及最找物、规制轨臻,以及榕调長期任移。為蘑 化起見,程式碼片段已减,只示提示和对generate content API的呼r叫。如需完整的 可款行程式码和更多箍例,請参圈「楼器人食错」。 指向物件 在圆片或影片影格中指出辱找物件,是機器人领域中祝觉舆韬言模型(VLM)的常見用途。在 下列舱例中,我們要求模型找出医片中的特定物件,亚传回图片中的座标。 1ua5 1odug a6o6 sog fron google.genai inport types #Initialize the Gen/I cllent and specify the mode[ matnaid-g t-Ja-sattuqou -Tutuab, = 01 130on client = genai.client() # Load your Inage and set up your pronpt with open( 'path/to/image-with-objects.jpg'."rb') as f: image_bytes = f,read() queries - 1 "bread", "starfruit", ""eueueq. pronpt = f"= Get all points natching the following objects: [', .join(querles] label returned should be an identifying nane for the object detect The answer should follow the json format: [&#123;&#123;"point: <point>, Label: <label1>\}, ...1. The points are [y. x] format nornalized to e-1eee. 1I 11 Inage_response = cllent.models-generate_content( model=M00EL_ID, contents=[ types.Part.fron_bytes{ data=inage _bytes, mime_type=*inoge/ipeg', prompt 1, conf1g = types.GenerateContentConf1g( temperature=.5, thinking_config=types , ThinkingConfig(thinking_budget=θ) print (image_response,text) 翰出内容與人門箱例频似,也就是包含找到的物件座標和标的JSON。 {"point": [671, 317]. "label': *bread"}.

• bread"},

{"point": [629, 307]. “labet′: *bread"}. {“point": [833, 8BB], “label′: *bread"}, {"point": [609, 663]. "label': *banana"}, {“point":[77e, 483], “lahel′:*starfruit"}

使用下列提示要求模型解镇抽象频别(例如「水果:),而非特定物件,找出画片中 的所有例项。

• = duoud

Get all points for fruit. The label returned should be an ldent. name for the object detected. ". + The answer should follow the json format: [{"point: <point>, Label: <labell>}。 --.]. The points are 1n [y. x] fornat normalized to e-1eee.* 如幂其他国像虑理技術,請参阳国像理解真面。 追影片中的物件 Genin1 Robotics-ER 1.5也能分析影片影格,追物件随時間的化。如需支援的影片格 式清單,错少阳影片输入」。 以下是基提提示,用於在模型分析的每個影格中导找特定物件: # Define the objects to find queries = [ "pen (on desk), "pen (in robot hand)", "laptap (opened}, '(pasolp) doidel base_prompt = f** The answer should follow the json format: [&#123;\f"point*: spoint>, "label*: <label1>&#125;&#125;, ...1. The points are in [y, x] format normalized to e-1e88, If no objects are found, return an empty JsoN List [1. 翰出结果售示在影片影格中追践的筆和筆電。 物件侦测和定界框 除了单一黏之外,模型也盲傅回2D遗界框,提供封用物件的矩形匠域。 这因舱例膚要求取得桌上可辨薄物件的2D遗界框。模型會收到指令,输出内容限制为25 個物件,亚為多個敦行個體命名專属名。 fron google inport genai # Initialize the GenAI client and specify the modeT HODEL_Ip - "genini-robotics-er-1.5-preview" client = genai.Client() g5ear JnoA pea7 # and set with open('path/to/image-with-objects.ipg',rb') as f: Image_bytes = f.read() prospt = ± Return bounding boxes as a JsoN array with labels. Never return 1 or code fencing. Limit to 25 objects. Incluce as many objects as can identify on the table. If an object is present multiplc tines, nanc them according to t unique characteristic (colors, size, position, unique character1 The farmat should be as follows: [{box_2d: [ymin, xmin, ymax, "label°: <label for the object>)] nornalized to 8-18oo. Thc valu box_2d must onty be 1ntegers 11 11 1 inage_response = client,models.gencratc_content( nodel=MODEL_ID, contents=[ types.Part.fron_bytes { data=inage_bytes, mine_type= inage/jpeg prompt 1, config = types.Gcncratccontentconfig( tenmperature=0.5, thinking_config=types ThinkingConf ig(thinking_budget=9) print(image_response. text)

以下顾示模型傅回的方增。

S

如需完整的可缺行程式码,請参Roboticscookbook、“固像理解;真面也提供其他视景 工作管例,例如分割和物件值测。 轨 GeniniRobotics-ER1.5 可生成定蹄的黏序列,有助於引募概器人移勤。 有助於引募機器人移動 这個航例要求将红盖移動到收纳盒的款题,包括起贴和一系列中耀點 from google inport genai Initialize the GenAI cllent and specify the mode[ atAaud-gt-Ja-saT1oqou-qugua6, = 0I 1300 client = genai .client() Load your Inage and set up your prompt mith opent'path/to/image-with-objects.jpg'. "rb') as f: image_bytes = f.read() points_data = [] = 1duoud Place a point on the red pen, then 15 points for the trajector moving the red pen to the top of the organizer on the Left. The points should be Labeled by order of the trajectory. from [start point at left hand) to <n> (final point) The answer should follou the son fornat: ["point*: <point>, label: <labell>), -..1. The points are in [y, x] format normalized to 8-18o9. 1 11 11 Inage_response = client .models-generate_content( ‘300=apou contents=[ types.Part.fron_bytes{ data=inage_bytes, mime_type='Inage/1peg*, 1, ↓tiaoud 1, config = types.GenerateContentConfigf temperature=.5, print (image_response. text) 回丽是一组座標,说明红重庶遵循的路径轨臻,才能完成將红重移到收纳盒顶端的任释:

{point": [5ee, 60el, "label′: 1}. {“point": [350, 550]. "label′: 4}, {“point": [2e0, 460], "label: 7}, {"point": [140, 370]. "label′: *10"}. {"po1nt": [110, 320]. "label′:*12"}, {"point": [100, 305], "label′:*14"}, {St.=1eqe1.[00E00t]uTod.}

自動化调度管理 Genin1Robotics-ER1.5可敦行更高的空翻推理,根摊斯络理解推動作或找出显佳位 置 為蓝電预留空 这個舱例说明Gem1n1Robot1cs-ER如何推描空间。提示窗要求模型找出需要移動的物件, 以便益其他项目展出空圈。 Teua5 1uodut #16o6 uouj fron google.genai inport types Initialize the GenAI clfent and specify the mode[ atAaud-gt-Ja-saT1oqou-Tuua6, = 0I 1300 client = genai .client() Load your Inage and set up your pronpt with open('path/to/image-with-objects.jpg'. "rb') as f: image_bytes = f.read()

• 1duod

Point to the object that I need to renove to make roon for m The ansver should follow the json fornat: [{"point*: spoint> "label*: <label1>), --.]. The points are 1n [y, x] fornat no inage_response = client,models.generate_content( nodel=MODEL_ID, contents=[ types. Part fron_bytes{ data=inage_bytes, mime_type=′ Inage/jpeg* ), prompt 1, config = types.GenerateContentConfig( temperature=.5, thinking_config=types ThinkingConfig(thinking_budget=0) print(image_response. text) 回覆内容包含可回答使用者同题的物體2D座標,在本例中,核物證鹰移動,為筆電聘出空 同。 (“ppint": [67z, 301], “label′: *The object that I need to renove to m

• EIViajer

弹備午餐 模型也能提供多步累工作的操作现明,亚指出每图步器的相關物件。这图例瞬示模型如何规劃

一系列步黑来打包午餐袋。

tua5 1uodut a6oo6 souj fron googlc.gcnai inport types #Initialize the Gen/I cllent and specify the mode[ matna.d-g't-Ja-sa11oqa.u-Tutuab. = 01 1300n client = genai .client() # Load your Inage and set up your pronpt with openl 'path/to/image-af-lunch.jpg'. 'rh') as f: image_bytes = f.read()

• * = 1duoud

Explain hov to pack the lunch box and Lunch bag. Point to ea objcct that you refer to. Each point should bc in thc format ["point*: [y, x], *Label": }]. where the coordinates are narnalized between 9-1e8e. Inage_response = cllent.models generate_content( ‘α 1300=1apou ]=s types.Part.fron_bytes{ data=inage_bytes, mine_typc=inage/ipeg, 1, 1duod 1, config = types.GenerateContentConf1g( temperature=B.5, thinking_config=types, Thinkingconfig(thinking_budget=θ) print(image_response,text) 遗项提示的回覆是一组逐步操作现明,内容為如何根输入的圆片打包午餐袋。 输入图片

横型输出 Based on the inage, here 1s a plan to pack the Lunch box and Lunch bag

• #Pack the fruit into the lunch box,** Place the [apple](apple), [

1 -

• Add the spoon to the Lunch box,* Put the [blue spoon] (bluc spoo)

• +Close the Lunch box,** Secure the Lid on the [blue Lunch box (bl

• #place the lunch box inside the lunch bag.*x Put the closed [blue

• Pack the renaining itens into the lunch bag.* Place the [bluc 5]

5 . Here is the List of objects and thelr locations: [{"point*: [899, 44a]. “label: "apple*}] [{"point": [814, 363], “label: "banana*)] ["point": [727, 470], "label*: "red grapes"}] [{"point": [706, 529], “label°: "bluc lunch box"}] ["point": [864, 517]. "Label*: "blue spoon*}] [{"point": [614, 7e5], “label*: "brown snack bar"}] 呼叫自訂機器人API 遗图例说明如何使用自訂概器人API编排工作。这個API專為取放作案設計。这项工作的 目標是拿起蓝色种木,然後放入橘色中

與本真的其他例频似,完整的可轨行程式码位於Robotics食错。

第一步是使用下列提示找出道雨项物品:

pronp1 = Locatc and point to the blue block and the orangc bowl. The returned should be an 1dentifying nane for the object detec' The ansver should follow the json fornat: [{“point*: <point: Thc points are in [y, x] format nornalized to 9-leee. 模型回康會包含精木和腕的标化座标: {“point":[389, 252], “label′:orange hawl), {"point": [727, 659]. "label*:blue block"} 本例使用下列模機器人API: def move[x, y。 high): print[f"maving to coordinates: (x), (y). (1s if high else 5)) def setoripperstate(opened): def returnTadrigin(): print("Returning to origin pose*) 下一步是呼叫一系列API函式,亚使用必要通辑款行慰作。下列提示包含機器人API的说 明,模型惠在调道项工作時使用。 pronpt = f"" You are a robotic arm with six degrees-of-freedom. You have the following functions available to you: def move(x, y, high) : to True neans the robot arn should be lifted above the scene for avoiding obstacles during motion. ‘high’ set to False means the arm should have the gripper placed on the surface for interactin objects. def setGripperState(opened): 0pens the gripper if opened set to true, otherwise closes the def returnToorigin(): Returns the robot to an initial state. Should be called as a c operation. The origin point for calculating the moves is at normalized point y={robot_origin_y), x={robot_origin_x). Use this as the new (e,e) calculating noves, allowing x and y to be negative. Perform a pick and place operation where you pick up the blue bloc normalized coordinates ({block_x). (block_y)) (relative coordinate {block_relative_x}. {block_relative_y}) and place it into the oran (relative coordinates: {bowl_relative_x), (bowl_relative_y)). Provide the sequence of function calls as a JsoN list of objects, 1 each ohject has a “function* key (the function name) and an *args" (a list of argunents for the function). Also, include your reasoning before the JsoN output. For exanple: Reasoning: To pick up the block, I wilt first move the arn to a hi position above the block, open the gripper, nove down to the block close the gripper, lift the arm, move to a high position above the nove down to the bowl, open the gripper, and then lift the arn bac a high position. 以下是模型根捕提示和模機器人API可能產生的输出内容。输出内容包括模型的思考遇程, 以及因此规劃的任務。此外,也會题示模型依序软行的楼器人函式呼叫输出内容。 Reasoning: Ta perforn the pick and place aperation, I will first move robotic arm to a high position above the blue block to ensure clearanc vill open the gripper in preparation for grasping. Next, I will lower the surface level to position the gripper around the block and close t to secure it. After picking up the block, I will lift the arn to clear surface and move tovards the orange bowl, maintaining a high position collisions. Once positioned above the bowl, I vill lower the arn to pl. block inside. I will then open the gripper to release the block. Final Lift the arm out of the bowL and call *returnToorigin to reset the rol state. ',anou, :,uogoumy. "args": [ 163. 427. true f , "args": [ true

"function*:*nove", "args": [ 163, 427. false

"args": 1 false }, "function*:"nove", "args": 1 4 2 163.

4.3

true 1 "functian*: *nove", "args": [

• 247,

5 1 98. true

"function':'nove", "args": [ 247, 98. false 1

"args": [ true 1 } f "functian*:“

• nove",

"args": I

• 247,

98, true 1 ), }, "function*: returnToorigin, "args": (] Executing Function Calls: noving to coordinates: 163, 427. 15 Jaddrf furuado noving to coordinates: 163, 427, 5 CLosing gripper moving to coordinates: 163, 427, 15 noving to coordinates: -247, 90. 5 Jaddtu6 furuado noving to cordinates: -247, 99, 15 Returning to origin pose 程式碼執行 Gen1n1Robot1cs-ER 1.5 可以建及行Python 程式码,以行需要勤动作的工 作,例如放大图片區域以查看更多细前。 这個例说明模型如何建端使用程式碼敦行:工具放大圆片的特定區域,然後飘行道项操 作来回答使用者的間题。 fron google.genai inpprt types client - genai.client() # Load your inage and set up yovr proapt mith open('path/to/image-of-object.1pg',*rb) as f: Image_bytes = f.read() " = duoud what is the air quality reading? Using the code execution fe. zoon in on the image to take a closer look. 1 response = client.nodels.generate_content( nodel=MODEL_ID, contents=[ 1mg, propt 1, config = types.GenerateContentConfigf temperature=0.5, tool s=[types Tool(code_execution=types ToolCodeExecution)] for part in response.candidates[θ] .content -parts: if part.text is not Nonc: print(part.text) if part.executable_code is not None: 1f part.code_execution_result 1s not None: 以下是這项工作的可能模型回覆。输出内容奢显示模型生成的程式理,用於放大圈片,以便更清 楚地阅搞壹面内容。此外,這项功能也會题示系统提整截剪方式的患考遇程。如雲完整的舱例 输出内容,辅警“Robot1cs cookbook。 # Delecl relevant objects and ovtput in json formal [[“box_2d":[379, 407, 481, 553], "Label°:*disp[ay")] fron PIL import Inage 5 i uao # ing = Image.open(*input_filc_9-jpeg") 6 Define the coordinates for cropping (normallzed to 9-Ies0) # Based on visual inspection, the display is roughLy centered around [- 8# Let's try to zoon in on the display area. 9 Bounding box for display: [ynin, xmin, ymax, xmax] In pixels. 1θ # Froe visual inspeclion of the original Image (size 99489 bytes, reso unknown but let’s assume a standard aspect ratio and roughly 1e0ex1668 12relative positioning). 13 # the display is Located at approximately y=380-488, x=418-540. 14 # Let's definc a lorger area around it for context. 059=x 02 es=x pue oss=/< o esc=/ mou1 douo o1 1ueA an (es 5,a7 ST thought The crop was too tight and only captured part of the display. Let's ad The previous crop vas *img.crop((4ee, 35e, 6o, 5ee))°. This means xmi The display is centered around x=48e, y=436. The previous crop captured the Left side of the display. Let's expand the x range to the right and y range to cover the whole c. New bounding box estinate: ymin=37e, xmin=468, ynax=588, xmax=606, Let's try again with a wider crop. 25 # Detect relevant objects and output in Json forma #[{“box_2d": [379, 487, 481, 553], "[abel°: disp[ay"] ] fron PIL import Inage ing = Image.open(input_file_a-jpeg) 30 Define the coordinates for cropoing (normalized to 9-Isso) 1 The reading of 14 fa1ls into the Moderate category. The prompt asks for the “alr quality reading, which can be interprete elither the numerical value or the categorical description. Given the n value is explicitly displayed, it's the prinary reading. Let's provide the nunerical value as the reading. The reading is 014. Based on the display on the air purifier, the air quality reading is * 最佳做法 如要提升機器人感用程式的效能和华雍度,請務必腺解如何有效與Gem1n1模型互勤。本節將 税明提示、虚理视覺资料和建横工作时的显佳做法和重要策略,協助您擅得最可靠的结果。

1.使用麗单明胶的韬言。

• 使用自然語言:Gemini模型可理解自然對括語言。幅以語意清楚的方式建模提示,

模凝人類自然下注指令的方式。

• 使用日常用籍:請使用常見的日常用語,避免使用技術或事術语。如果模型到特定

字铜的回惠不如预期,請答試使用更常見的同款字重新描醇。

2.最佳化視算输入内容。

• 放大核視群翻資訊:虚理小型或難以辨磁的物件时,請使用定界框困式,將感典涩的

物件摄立出来。接著,您可以将固片裁剪成所避图,然後將新圈片傅送給模型,进 行更群组的分析。

• 誉試調整光線和色彩:光综不佳和色彩封比度不足,可能盲影塞模型的感如能力。

3.将被問题细分為较小的步眼,逐一虑理每個酸小的步跟,引漂模型得出更精磁的结果,

4.透通共造提高掌確度。到於需要高度精確的任務,您可以多次使用相同提示查询模型,只

要将傅回的结是平均,即可得出“共」,这通常會更率確且可靠。 限制 使用GeminiRobotics-ER1.5開發時,請注意下列限制:

• 预觉状:

• 预宽状態:

模型目前處於预段。API和功能可能會變更,且未經过徹底測試,因此可能不適合用 於生產環境關鍵應用程式。

• 延運:

延课: 複雜的查詢、高解析度輸入内容或大量資料可能會導致處理時間增加。thinking_budget

• 幻梵:

與所有大型語言模型一樣,GeminiRobotics-ER1.5有時可能會「產生幻觉」或提 供不正確的資訊,尤其是針對模棱雨雨可的提示或超出分布圍的輸入内容。

• 取决於提示品質:

模型輸出内容的品質高度取决於输入提示的清晰度和具體程度。如果提示含糊不清或結構 不佳,可能曾導致結果不理想。

• 運算成本:

執行模型(尤其是使用影片输入内容或高thinking_budget時)會消耗運算資源產 生費用。详情請参阅「思考」真面。

• 輸入類型:

如要瞭解各模式的限制,請参阅下列主題。

• 圖片輸入内容

• 視訊輸入

• 音訊輸入

定價 如需價格和適用區域的細資訊,請参阅價格真面。 後續步

• 探索其他功能,亚續試不同的提示和輸入内容,發掘GeminiRobotics-ER

1.5的更多應用。如需更多例,請参阅「機器人食譜」。

• 如要瞭解GeminiRobotics模型如何以安全為優先考量而建構,請前往Google

DeepMind機器人安全真面。

• 如要瞭解GeminiRobotics模型最新消息,請前往GeminiRobotics登陵真面。

原始排版图

原始导出图超过单张 WebP 的尺寸上限，以下图片按从上到下的顺序连续保存。