Kwenzekeni
I-Black Forest Labs (BFL) ikhiphe i-FLUX 3 Action, i-'World Action Model' enesisindo esivulekile sepharamitha eyizigidi eziyizinkulungwane ezingu-7 eyenzelwe amarobhothi. Imodeli ithatha ukubonwa kwekhamera, isimo serobhothi, nemiyalo yolimi lwemvelo ukuze ikhiqize izenzo zomzimba. I-BFL ibika izinga lempumelelo elingu-42.92% kubhentshimakhi ye-NVIDIA's RoboLab-120, idlula umholi wangaphambilini, i-NVIDIA's Cosmos3-Nano-Policy (amapharamitha angu-16B), ngamaphesenti angu-6.1. Ukukhishwa kufaka phakathi izisindo, ikhodi, nezindlela zokupheka zokushuna kahle, okuvumela onjiniyela ukuthi bavumelanise imodeli namarobhothi athile basebenzisa izinhlaka ezifana ne-LeRobot.
I-Black Forest Labs, eyaziwa ngezithombe zayo ze-FLUX namamodeli wevidiyo, isikhule yaba amarobhothi ngokukhishwa kwe-FLUX 3 Action. Le modeli yepharamitha yebhiliyoni engu-7 iyisisindo esivulekile 'Imodeli Yesenzo Somhlaba' (WAM) ecubungula okokufaka kwekhamera, izimo zesistimu, kanye nemiyalelo yombhalo ukuze kukhishwe uchungechunge lwezenzo zomzimba ezingu-32 kanye nokubikezela kokuvela kwesigcawu esibukwayo.
Ngokusho kwe-VentureBeat, i-BFL ibika ukuthi i-FLUX 3 Action izuza izinga lempumelelo eliphelele elingu-42.92% kubhentshimakhi ye-NVIDIA's RoboLab-120. Lesi sikolo sidlula imodeli yangaphambilini evuliwe, NVIDIA's Cosmos3-Nano-Policy, ngamaphesenti angu-6.1. Ngokuphawulekayo, i-FLUX 3 Action isebenzisa amapharamitha ayizigidi eziyizinkulungwane ezingu-7 kuphela uma iqhathaniswa nezigidigidi ezingu-16 ze-Cosmos, futhi i-BFL ithi igijima izikhathi ezingu-1.43 ngokushesha, isungula umngcele omusha we-Pareto ukuze usebenze kahle nokusebenza.
Ukukhishwa kufaka phakathi izisindo zemodeli, ikhodi, amarisiphu acunula kahle, nezibonelo eziphindaphindekayo. I-BFL ithi amaqembu angakwazi ukushuna kahle imodeli esebenzisa imibukiso eqoqwe kumarobhothi awo, isho ngokuqondile ukuhambisana nerobhothi le-SO-101 kanye nohlaka lwe-LeRobot lwe-Hugging Face. Le nkampani iphinde yakhombisa amakhono emodeli ezindaweni okungezona ezerobhothi, okuhlanganisa ama-drones endizayo nokudlala umdlalo wevidiyo we-Doom ngaphandle kokufa okuphakathi kwegeyimu, nakuba lokhu kubonakala njengokuhlola kwasekuqaleni.
I-BFL yahlola kuqala Isenzo se-FLUX 3 ngoJulayi lapho yethula umndeni obanzi we-FLUX 3, kodwa ekuqaleni yayivinjelwe ocwaningweni olukhethiwe nabalingani bezentengiselwano. Ukukhishwa kwamanje kuphawula ukushintshela ekufinyelelekeni okubanzi, nakuba amanani entengo ethile ye-API yezentengiso yemodeli Yesenzo engakamenyezelwa. Imodeli yakhela ocwaningweni lokuzigeleza kwe-BFL, oluhlose ukufunda izethulo eziwusizo zokunyakaza kanye nesizathu-nomphumela kusuka kudatha yevidiyo ngaphandle kokuthembela kumamodeli azimele afriziwe.
Imininingwane yomthombo: venturebeat.com ↗
Kungani kubalulekile
Lokhu kukhishwa kubalulekile ngoba kunikeza enye indlela ehlangene, enesisindo esivulekile kumamodeli amarobhothi amakhulu obunikazi noma avaliwe. Ngokuthola amaphuzu aphezulu ebhentshimakhi anamapharamitha ambalwa (7B vs 16B) kanye nencazelo esheshayo, i-BFL yehlisa umgoqo wokungena emaqenjini afuna ukusebenzisa izinqubomgomo zerobhothi ezithuthukisiwe ngaphandle kwezinsiza zekhompyutha ezinkulu. Imvelo enesisindo esivulekile ivumela ukulungiswa kwendawo kanye nobumfihlo bedatha, obubalulekile ezinhlelweni zokusebenza zezimboni lapho idatha yokuphathelene ingakwazi ukuphuma endaweni. Iphinde iqinisekise indlela ye-'World Action Model', lapho ukuqeqeshwa kwangaphambili kwevidiyo kwazisa ukulawula ngokomzimba, okungenzeka kunciphise isidingo sokuqoqwa kwedatha eqondene nerobhothi.
Ukukhishwa kwe-FLUX 3 Action kuwumphumela embonini yamarobhothi ngoba inikeza inketho esebenza kahle, enesisindo esivulekile ephonsela inselelo amamodeli amakhulu, asebenzisa izinsiza. Ngokuthola imiphumela ephezulu yokuma okulinganiselwe ngamapharamitha ambalwa, i-BFL ibonisa ukuthi ukusebenza kahle nokusebenza akuhlukani, okubalulekile ekusetshenzisweni konqenqema lapho izinsiza zekhompiyutha zikhawulelwe.
Konjiniyela, imvelo yesisindo esivulekile semodeli iwumehluko oyinhloko. Ivumela amaqembu ukuthi ahlole, aguqule, futhi alungise kahle imodeli ngaphakathi kwengqalasizinda yawo, iqinisekisa ukuthi ukuboniswa kwamarobhothi okuphathelene nedatha yokusebenza kuhlala kuyimfihlo. Lokhu kuyinzuzo enkulu ngaphezu kwezinsizakalo zamarobhothi esingethwe noma amamodeli wobunikazi afana ne-Google's Gemini Robotics On-Device 2, akhawulelwe kubahloli abakhethiwe.
Isakhiwo semodeli, esihlanganisa ukuqeqeshwa kwangaphambili kwevidiyo nokukhiqizwa kwesenzo, siqondana nethrendi ekhulayo yokusebenzisa amamodeli akhiqizayo ukuze kuqondwe ukuguquguquka komzimba. Le ndlela ingase yehlise inani ledatha eqondene nerobhothi edingekayo ekuqeqesheni, njengoba imodeli ingase isebenzise izethulo ezibanzi ezibonakalayo nezingokwenyama ezifundwe kuvidiyo. Lokhu kungasheshisa ukuthuthukiswa kwamasistimu erobhothi emisebenzi efana nokuhlunga, ukuhlanganisa, nokuzulazula.
Nokho, umthelela ongokoqobo uncike emigomeni yokugcina yelayisensi kanye nekhono lamaqembu angaphandle ukukhiqiza kabusha imiphumela ebikiwe. Amabhentshimakhi erobhothi avame ukuhlukaniswa, futhi ukuqhathanisa okuqondile phakathi kwezakhiwo zamamodeli ahlukene (ama-WAM, ama-VLA, namamodeli okucabanga ngesenzo) kunzima. Impumelelo ye-FLUX 3 Action ekugcineni izonqunywa ukusebenza kwayo emhlabeni wangempela, ezindaweni ezihlukene zamarobhothi kunezilinganiso zokulingisa nje.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
In AI, what are a model's "parameters"?
Ongakubuka ngokulandelayo
Onjiniyela kufanele baqaphe ukutholakala kwemigomo yokugcina yelayisense yezohwebo, njengoba ukukhishwa kwamanje kungokocwaningo kanye nokufinyelela kozakwethu. Ukukhiqizwa kabusha okuzimele kwemiphumela yebhentshimakhi ku-hardware ehlukahlukene kubalulekile, njengoba amabhentshimakhi wamarobhothi evamise ukuntula ukumiswa. Ukwengeza, bheka ukuhlanganiswa kwe-FLUX 3 Action kuma-API ezentengiso abanzi noma ukuthi i-BFL inwebisa isu layo eliwumgogodla ovulekile ukuze ifake izindawo zokugcina eziqondene namarobhothi.
Into ebaluleke kakhulu okufanele uyibuke ukushicilelwa kwemibandela yokugcina yelayisense yezohwebo ye-FLUX 3 Action. Nakuba i-BFL ikhiphe izisindo nekhodi, izimo ezithile zokusetshenziswa kwezohwebo, ukusatshalaliswa kabusha, nokuguqulwa akukacaciswanga ngokugcwele. Lokhu kuzonquma ukuthi imodeli ingamukelwa kabanzi kangakanani kuzilungiselelo zezimboni.
Ukuqinisekiswa okuzimele kwemiphumela yebhentshimakhi kubalulekile. Izinga lempumelelo le-BFL elibikiwe elingu-42.92% ku-RoboLab-120 lisekelwe ekuhloleni kwabo, futhi ibhodi yabaphambili ye-NVIDIA ibingakabuyekezwa ngale miphumela ngesikhathi sombiko. Ukukhiqizwa kabusha ngamaqembu ocwaningo ezinkampani zangaphandle kuzodingeka ukuze kuqinisekiswe ukusebenza kwemodeli kanye nokugcwaliseka kwayo.
Onjiniyela kufanele futhi baqaphe ukuhlanganiswa kwe-FLUX 3 Action kuzinhlaka ezikhona zamarobhothi nehadiwe. I-BFL igqamise ukuhambisana nerobhothi le-SO-101 kanye ne-LeRobot, kodwa ukusebenza kwemodeli kwezinye izinkundla ezivamile zamarobhothi, njengalezo ezivela ku-NVIDIA noma i-Ai2, kusazobonakala. Ukucushwa kalula kwezingxenyekazi zekhompuyutha ezihlukene kuzoba yisici esibalulekile sokwamukelwa kwayo.
Ekugcineni, indawo yokuncintisana kumarobhothi anesisindo esivulekile ithuthuka ngokushesha. Ngamamodeli afana ne-Ai2's MolmoAct 2 kanye ne-NVIDIA's GR00T N1.7 asevele etholakala, impumelelo ye-FLUX 3 Action izoncika ekhonweni layo lokunikeza izinzuzo ezihlukile mayelana nokusebenza, ukusebenza kahle, kanye nokusebenziseka kalula. Izinyanga ezimbalwa ezizayo cishe zizobona ukuqhathanisa okwengeziwe namabhentshimakhi njengoba umphakathi uhlola imodeli ezinhlelweni ezihlukahlukene.