I-Zibian Robotics namuhla ikhiphe futhi yavula i-HOST framework (Human-to-robot One-Shot Skill AcquisiTion, ukutholwa kwekhono lomuntu kuya kurobhothi ngesampula elilodwa), ethi “ivumela irobhothi ukuthi lifunde ikhono elisha ngokubuka nje ividiyo yomuntu ethatha amashumi emizuzwana, kuyilapho ligcina amakhono eselivele liwafundile.”
I-HOST isebenzisa indlela entsha nephambili: esikhundleni sokuhumushela izenzo zomuntu zibe izenzo zerobhothi bese izama ukuzilingisa, irobhothi liqala ngokucabanga ngomphumela wokwenza isenzo, bese lihlukanisa kulowo mphumela izenzo okufanele lizenze. Imiphumela yale ndlela entsha iyajabulisa: izinga lempumelelo yerobhothi ekufundeni ikhono kwividiyo yomuntu yemizuzwana engu-29 lifinyelela ku-62%, lidlula i-zero-shot baseline engu-45%. Uma kuqhathaniswa nohlelo lwe-Pi-0.5 + fine-tuning, i-HOST idinga idatha encane izikhathi ezingu-50, kanti ijubane lokufunda amakhono lenyuka izikhathi ezingu-500.
Umbono we-HOST ususelwa emfundisweni “yokufunda ngokubuka” yesayensi yokuqonda: lapho abantu bebhekene nomsebenzi abangawujwayele, abadingi ukuwufunda ngokuzijwayeza isikhathi eside noma ngokuzama zonke izinketho. Kunalokho, basebenzisa amakhono abanawo ukuze balingise engqondweni umphumela olindelekile wesenzo, bese benza isenzo esifanele. Bephefumulelwe yilokhu, abacwaningi be-Zibian baguqule indlela yokufunda ye-HOST kusukela “kumjikelezo wokulungisa ngesikhathi sokuqeqesha” yaba “ukutholwa kwekhono ngesikhathi sokucabanga”, ukuze irobhothi lifunde amakhono amasha ngokubuka umuntu enza umsebenzi.
Inkinga yokuqala irobhothi okufanele liyixazulule yile: lingahlanganisa kanjani inqubekelaphambili yalo ekwenzeni umsebenzi nenqubekelaphambili ekwividiyo?
Ngokwesibonelo: ngemva kokusebenza imizuzwana engu-10, umuntu okwi-video kungenzeka ukuthi useqedile ukusika imifino futhi useqalile ukuyithosa, kuyilapho irobhothi lisasika imifino. Uma kuhlanganiswa isikhathi kuphela, irobhothi lizolahlekelwa yinqubekelaphambili yomsebenzi.
Ngenxa yalokhu, i-HOST isebenzisa “ukuhlanganisa ngokwenqubekelaphambili yomsebenzi”. I-HOST iguqula ifreyimu ngayinye yevidiyo nesinyathelo ngasinye sokusebenza kwerobhothi kube ama-vector asendaweni efanayo yama-vector; bese isebenzisa i-algorithm ye-“dynamic time warping” ukuze ihlanganise ngokuzenzakalelayo “ifreyimu engu-X yevidiyo yomuntu nesinyathelo esingu-Y sokusebenza kwerobhothi”. Ngaleyo ndlela, njalo lapho irobhothi lenza isinyathelo esithile, libona ifreyimu ehambisana nenqubekelaphambili yomsebenzi.
Ngenxa yale ndlela, i-HOST yehlisa iphutha lesikhathi ekuhlanganiseni inqubekelaphambili yomsebenzi ngezinga elilodwa lobukhulu, futhi ikwazi ukuvumelanisa ngokunembile ukwenziwa kwerobhothi nomboniso wevidiyo.
Ingxenye ewumgogodla ye-HOST iyimodeli yenqubomgomo eyicascade enekhono lokubikezela okubukwayo. Isebenzisa i-dual-expert MoT architecture, efana nobuchopho obubili obunemisebenzi ecacile: “ubuchopho obubonayo” (uchwepheshe wevidiyo) buphatha izithombe zevidiyo, buthole inqubekelaphambili yomsebenzi futhi bubikezele isimo esizayo; “ubuchopho bezenzo” (uchwepheshe wezenzo) buguqula isimo esibikezelwe sibe yizenzo okufanele zenziwe futhi bulawule ukwenziwa kwazo.
Ngesikhathi beqeqesha imodeli, ithimba le-Zibian lahlukanisa inqubo yaba izigaba ezimbili: “ukuqeqeshwa kwangaphambili komzimba ofanayo” kanye “nokuqeqeshwa kwevidiyo yomuntu nerobhothi”. Esigabeni sokuqala, irobhothi lifunda kumavidiyo erobhothi ngokwalo lenza imisebenzi, lifunde ukubikezela isimo ngemva kokuqedwa komsebenzi futhi likhiqize izenzo ezihambisanayo. Esigabeni sokuqeqeshwa kwevidiyo yomuntu nerobhothi, irobhothi liqhubeka lifunda kumavidiyo omuntu abonisa umsebenzi, liguqule inqubo umuntu enza ngayo umsebenzi ibe umphumela womsebenzi ngokombono werobhothi, bese licabanga ngezenzo ezidingekayo ukuze liqede umsebenzi ngokusekelwe kumphumela okuhlosiwe. Lokhu kufeza ukudluliswa kwekhono kusukela embonisweni womuntu kuya kurobhothi.
Njengamanje, iphepha locwaningo nekhodi ye-HOST sekukhishwe konke ngaphansi kwelayisensi evulekile. Ngokuzayo, lapho amarobhothi enza imisebenzi yasekhaya, angase angasadingi ukuqoqwa kwedatha nokuqeqeshwa okukhethekile, kodwa afunde amakhono amasha ngokubuka nje umboniso ocacile womuntu.
Izinkomba
Ikhasi lephrojekthi
Isixhumanisi sephepha
Isixhumanisi se-Github
Isixhumanisi se-Huggingface