Ukudala iimoto eziziqhubayo ezinokwenza izigqibo ezinjengomntu kuye kwaba yinjongo yexesha elide yeenkampani ezifana neWaymo, iGM Cruise, i-Uber kunye nabanye. I-Intel Mobileye ibonelela ngeResponsibility-Sensitive Safety (RSS) imodeli yemathematika, echazwa yinkampani njenge "common sense" indlela ebonakaliswa ngokucwangcisa i-autopilot ukuba iziphathe ngendlela "elungileyo", njengokunika ezinye iimoto ilungelo lendlela. . Kwelinye icala, i-NVIDIA iphuhlisa ngokukhutheleyo uMmandla woKhuseleko lweNdawo, itekhnoloji yesigqibo esekwe kwinkqubo ebeka iliso kwiintshukumo ezingakhuselekanga zabasebenzisi bendlela abajikelezileyo ngokuhlalutya idatha evela kwizinzwa zezithuthi ngexesha lokwenyani. Ngoku iqela lezazinzulu ezivela kwiMassachusetts Institute of Technology (MIT) lijoyine olu phando kwaye licebise indlela entsha esekwe kusetyenziso lweemephu ezifana neGPS kunye nedatha ebonakalayo efunyenwe kwiikhamera ezifakwe kwimoto ukuze i-autopilot ikwazi ukuhamba kwindawo engaziwayo. iindlela ezifanayo nomntu.
Abantu banobuchule obukhethekileyo ekuqhubeni iimoto kwiindlela abangazange bahambe kuzo ngaphambili. Sisuka nje sithelekise oko sikubonayo okusingqongileyo noko sikubona kwizixhobo zethu zeGPS ukuze sibone apho sikhoyo nalapho simele siye khona. Iimoto eziziqhubayo, kwelinye icala, zikufumanisa kunzima kakhulu ukuhamba kwiindawo ezingaziwayo zendlela. Kwindawo entsha nganye, i-autopilot ifuna ukuhlalutya ngononophelo indlela entsha, kwaye iinkqubo zolawulo ezizenzekelayo zixhomekeke kwiimephu ezintsonkothileyo ze-3D ababoneleli abazilungiselela kwangaphambili.
Kwiphepha elivezwe kule veki kwiNkomfa yaMazwe ngaMazwe ngeRobhothi kunye ne-Automation, abaphandi be-MIT bachaza inkqubo yokuzimela yokuzimela "efunda" kwaye ikhumbula iipatheni zokuthatha izigqibo zomqhubi womntu njengoko zihamba ngeendlela kwindawo encinci yedolophu usebenzisa kuphela idatha evela kwividiyo iikhamera kunye nemephu elula efana ne-GPS. Umqhubi wenqwelo-moya oqeqeshiweyo unokuqhuba imoto engaqhubi kwindawo entsha ngokupheleleyo, elinganisa ukuqhuba komntu.
Njengomntu, i-autopilot iphinda ibone nakuphi na ukungangqinelani phakathi kwemephu yayo kunye neempawu zendlela. Oku kunceda inkqubo ibone ukuba indawo ekuyo endleleni, izixhobo zoluvo, okanye imephu ayichanekanga ukuze ikwazi ukulungisa indlela yesithuthi.
Ukuqala ukuqeqesha inkqubo, umqhubi womntu uqhube i-Toyota Prius ezenzekelayo exhotywe ngeekhamera ezininzi kunye nesiseko se-GPS yokukhangela inkqubo ukuqokelela idatha kwizitrato zedolophu zasekuhlaleni, kubandakanywa izakhiwo zendlela ezahlukeneyo kunye nemiqobo. Inkqubo ke yaqhuba ngempumelelo imoto ecaleni kwendlela ecwangciswe kwangaphambili kwenye indawo enehlathi ejonge ukuvavanya iimoto ezizimeleyo.
"Ngenkqubo yethu, akufuneki uqeqeshe kuyo yonke indlela kwangaphambili," utshilo umbhali wokufunda uAlexander Amini, umfundi ophumelele iMIT. "Ungakhuphela imephu entsha yemoto yakho ukuhamba kwiindlela engazange ibonwe ngaphambili."
"Injongo yethu kukudala ukuhamba okuzimeleyo okuhambelana nokuqhuba kwiindawo ezintsha," wongeza umbhali-mbhali uDaniela Rus, umlawuli weComputer Science kunye ne-Artificial Intelligence Laboratory (CSAIL). "Umzekelo, ukuba siqeqesha isithuthi esizimeleyo ukuba siqhube kwindawo yasedolophini njengezitrato zaseCambridge, inkqubo leyo kufuneka ikwazi ukuqhuba kakuhle ehlathini, nokuba ayikaze iyibone indawo enjalo ngaphambili."
Iinkqubo zokuhamba eziqhelekileyo ziqhuba idatha yenzwa ngokusebenzisa iimodyuli ezininzi ezilungiselelwe imisebenzi efana neendawo, imephu, ukufumanisa into, ukucwangciswa kwentshukumo kunye nokulawula. Kwaphela iminyaka, iqela likaDaniela liye laphuhlisa iindlela zokuhamba zokugqibela eziqhuba idatha yenzwa kunye nokulawula imoto ngaphandle kwemfuneko yeemodyuli ezikhethekileyo. Ukuza kuthi ga ngoku, nangona kunjalo, ezi modeli zisetyenziswe ngokungqongqo kuhambo olukhuselekileyo endleleni, ngaphandle kwenjongo yokwenyani. Kumsebenzi omtsha, abaphandi bahlambulule inkqubo yabo yokuphela kwenjongo ukuya kwindawo yokuya kwindawo engaziwayo ngaphambili. Ukwenza oku, oosonzululwazi baqeqeshe i-autopilot yabo ukuqikelela usasazo olupheleleyo olunokwenzeka kuyo yonke imiyalelo enokwenzeka yolawulo nangaliphi na ixesha ngelixa uqhuba.
Inkqubo isebenzisa imodeli yokufunda yomatshini ebizwa ngokuba yi-convolutional neural network (CNN), esetyenziswa ngokuqhelekileyo ukuqondwa komfanekiso. Ngexesha loqeqesho, inkqubo ijonga indlela yokuziphatha komqhubi womntu. I-CNN inxulumanisa ukujika kwevili lokuqhuba kunye nokujika kwendlela, ethi ijonge kwiikhamera nakwimephu yayo encinci. Ngenxa yoko, inkqubo ifunda eyona miyalelo yokuqhuba kwiimeko ezahlukeneyo zokuqhuba, njengeendlela ezithe tye, iindlela ezine okanye ii-T-junctions, iifolokhwe kunye nokujika.
"Ekuqaleni, kwi-T-intersection, kukho iindlela ezininzi ezahlukeneyo zemoto ezinokujika," kusho uRus. "Imodeli iqala ngokucinga ngayo yonke le mikhombandlela, kwaye njengoko i-CNN ifumana iinkcukacha ezininzi malunga nezinto ezenziwa ngabantu kwiimeko ezithile endleleni, iya kubona ukuba abanye abaqhubi bajika ngasekhohlo kwaye abanye bajika ngasekunene, kodwa akukho mntu uya ngqo. . Umphambili othe ngqo ukhutshelwe ngaphandle njengomkhombandlela onokwenzeka, kwaye imodeli iphetha ngelithi kwi-T-junctions inokuhamba kuphela ngasekhohlo okanye ngasekunene.
Ngelixa uqhuba, i-CNN iphinda ikhuphe iimpawu zendlela ezibonakalayo kwiikhamera, ivumela ukuba iqikelele utshintsho lwendlela olunokwenzeka. Umzekelo, ichonga uphawu olubomvu lokumisa okanye umgca ophukileyo ecaleni kwendlela njengemiqondiso yendlela ezayo. Ngomzuzu ngamnye, isebenzisa unikezelo oluqikelelweyo olunokwenzeka lwemiyalelo yolawulo ukukhetha owona myalelo uchanekileyo.
Kubalulekile ukuqaphela ukuba, ngokutsho kwabaphandi, i-autopilot yabo isebenzisa iimephu ekulula kakhulu ukuzigcina kunye nokusebenza. Iinkqubo zolawulo ezizimeleyo zisebenzisa iimephu zelidar, ezithatha malunga ne-4000 GB yedatha ukugcina isixeko saseSan Francisco. Kwindawo entsha nganye, imoto kufuneka isebenzise kwaye yenze iimephu ezintsha, ezifuna inani elikhulu lememori. Ngakolunye uhlangothi, imephu esetyenziswa yi-Autopilot entsha igubungela ihlabathi lonke ngelixa lihlala kuphela i-40 gigabytes yedatha.
Ngexesha lokuqhuba ngokuzimeleyo, inkqubo iphinda iqhathanise idatha yayo ebonakalayo kunye nedatha yeemephu kunye neeflegi nakuphi na ukungafani. Oku kunceda isithuthi esizimeleyo ukuba siqonde ngcono ukuba siphi na endleleni. Kwaye oku kuqinisekisa ukuba imoto ihlala kweyona ndlela ikhuselekileyo, nokuba ifumana ulwazi lwegalelo eliphikisanayo: ukuba, yithi, imoto ihamba ngendlela ethe tye ngaphandle kokujika, kwaye i-GPS ibonisa ukuba imoto kufuneka ijike ngasekunene, imoto iya kuthi. ukwazi ukuya ngqo okanye ukuyeka.
U-Amini uthi: βEhlabathini lokwenyani, izixhobo zokuva umntu ziyasilela. "Sifuna ukuqiniseka ukuba i-autopilot yethu iyakwazi ukumelana neentsilelo ezahlukeneyo zenzwa ngokudala inkqubo enokufumana nayiphi na imiqondiso yengxolo kwaye isahamba ngendlela ngokuchanekileyo."
umthombo: 3dnews.ru