I-Google DeepMind ishicilele ibhulogi izolo (Julayi 30), imemezela ukwethulwa kwemodeli ye-Gemini Robotics ER 2, okuyimodeli yayo enamandla kakhulu yokucabanga ngokomzimba (embodied reasoning) eqondiswe kumarobhothi.
Mayelana nendima yayo, le modeli ingabhekwa “njengobuchopho obuthuthukile” berobhothi, obuxhumanisa ukuxhumana nabantu, ukuqonda umhlaba obonakalayo nokuhlela imisebenzi enezinyathelo eziningi. Kamuva le modeli izodlulisela imisebenzi kumamodeli asezingeni eliphansi okubona- ulimi-isenzo (VLA), futhi ilawule ukuqedwa kwemisebenzi ehlukahlukene.
Ngokwemisebenzi, uma iqhathaniswa nenguqulo 1.6 yangaphambilini, imodeli ye-Gemini Robotics ER 2 manje ivumela irobhothi ukuthi lilandelele inqubekelaphambili yokusebenza kwalo ngokubuka ividiyo ngokuqhubekayo, lilungise lapho kuvela izinkinga, futhi libone ngokunembile isikhathi sokuqala isinyathelo esilandelayo. Le modeli ingasebenzisa ngokwendabuko i-Google Search (ukusesha kwe-Google) noma imisebenzi echazwe unjiniyela.
Mayelana nama-agent aphathekayo, ngesikhathi irobhothi lenza umsebenzi, imodeli ye-Gemini Robotics ER 2 ingacabanga ngesikhathi esifanayo ngezinyathelo ezilandelayo, inciphise ukuma okubangelwa indlela yamarobhothi evamile ethi “yima — cabanga — bese wenza”. Le modeli ixhumeka endaweni yokusakaza yezindlela ezimbili ye-Gemini Live API (i-Gemini Real-time API), eqondiswe emisebenzini yamarobhothi ezwela ukubambezeleka.
Uma iqhathaniswa ne-Gemini Robotics ER 1.6, i-ER 2 ingahlaziya ukusakazwa kwevidiyo okuqhubekayo. Ngalokhu, irobhothi lingalandelela inqubekelaphambili yomsebenzi walo, lilungise noma liphinde lizame lapho isenzo sihluleka, futhi linqume ukuthi liqale nini isinyathelo esilandelayo.
Esivivinyweni sokuhlukanisa inqubekelaphambili yomsebenzi, ukunemba kwe-Gemini Robotics ER 2 kube ngu-57.4%. Leli khono lingasiza irobhothi ukuthi lilungise izinyathelo ezihlulekile ngaphandle kokuqalisa kabusha yonke inqubo yokusebenza.
Esivivinyweni sokuthola izikhathi ezibalulekile (moment-finding), imodeli kufanele ithole ifreyimu yevidiyo enembile lapho kwenzeka khona umcimbi obalulekile, njengokunquma ukuthi kufanele iyeke nini ukuthela ikhofi enkomishini. Ukunemba kwe-ER 2 kube ngu-91.3%, kanti iphutha lesikhathi eliphelele eliyisilinganiso kube yimizuzwana engu-0.96.
UGoogle uthi le modeli ifinyelela ukunemba okusondele esigabeni samamodeli amakhulu ngezindleko eziphansi zokubala, isebenza ngesivinini esiphindwe ka-4, futhi ihlangabezana nezidingo zokubambezeleka ezingaphansi kwesekhondi ezidingekayo ukuze amarobhothi omhlaba wangempela asebenze ngokuphepha.
I-ER 2 iphinde isekele ukubambisana kwamarobhothi amaningi. Izinhlobo ezahlukene zamarobhothi zingasebenzisa ukuqonda okwabelwana ngakho kwezincazelo ukuze zixhumane, zinikezelane imisebenzi, futhi ziqede izinqubo eziyinkimbinkimbi okunzima ukuthi idivayisi eyodwa iziqede yodwa. UGoogle ubonise isibonelo sokubambisana phakathi kwe-Apollo 2 yakwa-Apptronik ne-FR3 Duo yakwa-Franka.
I-Gemini Robotics ER 2 isiyatholakala esidlangalaleni kubathuthukisi nge-Gemini API nange-Google AI Studio; ngesikhathi esifanayo, isiqalise ukubonwa kwangasese ku-Gemini Enterprise Agent Platform (i-Gemini Enterprise Agent Platform). UGoogle uphinde wakhipha ikhodi eyisibonelo yokumiswa kwemodeli nezibonelo zemiyalelo.