Nzira yekutsanangura sei yekupedzisira-ku-kuguma yakazvimirira yekutyaira system?
Tsanangudzo inonyanyozivikanwa ndeyekuti "end-to-end" system isystem inopinza ruzivo rwe sensor yakasvibira uye inoburitsa zvakananga zvinosiyana zvine chekuita nebasa. Semuenzaniso, mukucherechedzwa kwemufananidzo, CNN inogona kunzi "kuguma-kusvika-kumagumo" kana ichienzaniswa neyechinyakare ficha + classifier nzira.
Mumabasa ekutyaira akazvimirira, data kubva kune akasiyana masensa (akadai semakamera, LiDAR, Radar, kana IMU ...) inoiswa, uye masaini ekudzora mota (senge throttle kana steering vhiri angle) anobuda zvakananga. Kufunga nezvekugadzirisa nyaya dzemhando dzakasiyana dzemotokari, iyo inobuda inogona zvakare kudzoreredzwa kune yekutyaira trajectory yemotokari.
Zvichienderana nenheyo iyi, modular end-to-end concepts dzakabudawo, senge UniAD, iyo inovandudza mashandiro nekuunza kutariswa kweakakodzera mabasa epakati, mukuwedzera kune yekupedzisira inobuda kudzora masaini kana waypoints. Zvisinei, kubva pane tsanangudzo yakamanikana yakadaro, musimboti wekupedzisira-kusvika-kumagumo unofanirwa kunge uri kusarasikirwa kwekutapurirana kwemashoko ekunzwa.
Ngatitangei tiongorore kupindirana pakati pekunzwa nePnC module mune dzisiri-kuguma-kusvika-kumagumo masisitimu. Kazhinji, tinoona zvinhu zvakachena (zvakadai semotokari, vanhu, nezvimwewo) uye kuongorora nekufanotaura zvinhu zvavo. Isu tinodzidzawo nezve nharaunda isingachinjiki (senge mugwagwa, miganhu yekumhanya, marobhoti, nezvimwewo). Dai tainge takanyatsotsanangurwa, taizoonawo zvipingamupinyi zvepasi rose. Muchidimbu, ruzivo rwunobuda nemafungiro aya anoumba modhi yekuratidzira yezviitiko zvakaoma zvekutyaira.
Nekudaro, kune zvimwe zviripachena zviratidziro, yazvino inoburitsa pachena haigone kutsanangura zvizere zvinhu zvinokanganisa maitiro ekutyaira munzvimbo, kana mabasa atinoda kutsanangura madiki, uye zvakaoma kuverenga mabasa ese anodiwa. Naizvozvo, masisitimu ekupedzisira-kusvika-kumagumo anopa (zvichida zvisina kujeka) kumiririrwa kwakazara netariro yekuzviitira uye nekusarasikirwa kuita paPnCs neruzivo urwu. Sekuona kwangu, masisitimu ese anogona kuzadzisa ichi chinodiwa anogona kunzi generalized end-to-end.
Kana dziri dzimwe nyaya, dzakadai sedzimwe optimizations yeane simba yekudyidzana mamiriro, ndinotenda kuti kwete chete-kusvika-kumagumo kunogona kugadzirisa aya matambudziko, uye kupera-kusvika-kumagumo kungave kusiri iyo yakanyanya mhinduro. Nzira dzechinyakare dzinogona kugadzirisa matambudziko aya, uye chokwadi, kana huwandu hwe data hwakakura zvakakwana, kupera-kusvika-kumagumo kunogona kupa mhinduro iri nani.
Kumwe kusanzwisisana pamusoro pekupedzisira-ku-kuguma kutyaira kwega
1. Kudzora masaini uye waypoints inofanirwa kubuda kuti ive yekupedzisira-kumagumo.
Kana iwe uchibvumirana neyakakura-kusvika-kumagumo pfungwa yakakurukurwa pamusoro, saka dambudziko iri riri nyore kunzwisisa. Kupera-kusvika-kumagumo kunofanira kusimbisa kusarasikirwa kwekufambiswa kwemashoko pane kuburitsa zvakananga huwandu hwebasa. Nzira yakamanikana yekuguma-kusvika-kumagumo ichakonzera dambudziko rakawanda risingakoshi uye inoda mhinduro dzakawanda dzakavanda kuti uve nechokwadi chekuchengeteka.
2.Kuguma-kusvika-kumagumo system inofanira kunge yakavakirwa pamhando huru kana maonero akachena.
Hapana hukama hunodiwa pakati pekupedzisira-kusvika-kumagumo kuzvimirira kutyaira, hombe-modhi yekuzvimirira kutyaira, uye kungoona chete kutyaira uchizvimiririra nekuti iwo pfungwa dzakazvimiririra zvachose; gadziriro yekuguma-ku-kuguma haisati ichisundwa nemienzaniso mikuru, uyewo haisati ichisundwa nechiono chakachena. ye.
3.Mukufamba kwenguva, zvinokwanisika here kune yakataurwa pamusoro-yekupedzisira-yekupedzisira system nenzira yakamanikana kuti iwane kuzvitonga kutyaira pamusoro peL3 level?
Kuita kwezvino kunzi pure end-to-end FSD iri kure nekukwana kuti isangane nekuvimbika uye kugadzikana kunodiwa padanho reL3. Kunyanya kutaura, kana hurongwa hwekuzvityaira huchida kutambirwa neveruzhinji, chikuru ndechekuti veruzhinji vabvume here kuti pane dzimwe nguva muchina uyu unotadza, uye vanhu vanogona kuzvigadzirisa zviri nyore. Izvi zvakanyanya kuoma kune yakachena-kusvika-kumagumo system.
Somuenzaniso, zvose zviri zviviri Waymo neCruise muNorth America zvakava netsaona dzakawanda. Nekudaro, tsaona yekupedzisira yaCruise yakakonzera kukuvara kaviri, kunyangwe tsaona dzakadaro dzisingadzivisiki uye dzinogamuchirwa kune vatyairi vevanhu. Zvakadaro, mushure menjodzi iyi, hurongwa uhwu hwakatadza kuona nzvimbo yaitikira tsaona nenzvimbo yevakakuvara nekudzikiswa kusvika pakuita pull-over mode, zvichiita kuti vakuvara vadhonzerwe kwenguva yakareba. Maitiro aya haagamuchirike kune chero mutyairi wevanhuwo zvake. Hazvizoitwe, uye mhedzisiro yacho ichave yakaipa kwazvo.
Uyezve, iyi inhare yekumuka iyo yatinofanira kunyatso funga nzira yekudzivirira mamiriro aya panguva yekuvandudza uye kushanda kweanozvimirira ekutyaira masisitimu.
4.Saka panguva ino, ndedzipi mhinduro dzinoshanda dzechizvarwa chinotevera chezvizhinji-zvakagadzirwa zvinobatsira kutyaira masisitimu?
Maererano nekunzwisisa kwangu ikozvino, kana uchishandisa iyo inonzi inonzi kuguma-kumagumo modhi mukutyaira, mushure mekubudisa trajectory, ichadzorera mhinduro inobva pamaitiro echinyakare. Neimwe nzira, kudzidza-kwakavakirwa kuronga uye yechinyakare trajectory kuronga algorithms inoburitsa akawanda trajectories panguva imwe chete uye wozosarudza imwe trajectory kuburikidza neanosarudza.
Iyi mhando yemhinduro yekuvhara uye sarudzo inomisa muganho wepamusoro wekuita kweiyi cascade system kana iyi sisitimu yekuvaka yakagamuchirwa. Kana iyi nzira ichiri yakavakirwa pakudzidza kwemhinduro dzakachena, kukundikana kusingatarisirwi kuchaitika uye chinangwa chekuve akachengeteka hachizowanikwe zvachose.
Kana tikafunga kugadzirisa patsva kana kusarudza kushandisa nzira dzechinyakare dzekuronga panzira inobuda, izvi zvakaenzana negwara rinogadzirwa nenzira inotungamirwa nekudzidza; saka, sei isu tisingagadzirise zvakananga uye nekutsvaga nzira iyi?
Zvechokwadi, vamwe vanhu vangataura kuti kugadzirisa kwakadaro kana dambudziko rekutsvaga harisi-convex, rine nzvimbo yakakura yehurumende, uye hazvibviri kumhanya munguva chaiyo-nguva pane in-motokari system. Ndinoteterera munhu wese kuti anyatsofunga nezvemubvunzo uyu: Mumakore gumi apfuura, hurongwa hwekuona hwakagamuchira kanokwana kazana kugovewa kwesimba rekombuta, asi zvakadini nePnC module yedu?
Kana isu zvakare tikabvumira iyo PnC module kushandisa hombe yekombuta simba, yakasanganiswa nekumwe kufambira mberi mune yepamusoro optimization algorithms mumakore achangopfuura, iyi mhedziso ichiri yechokwadi? Kune rudzi urwu rwedambudziko, tinofanira kufunga izvo zvakarurama kubva pamisimboti yekutanga.
5.Kuyananisa sei hukama pakati pe data-inotyairwa uye nzira dzechinyakare?
Kutamba chess muenzaniso wakada kufanana nekutyaira wega. Muna Kukadzi wegore rino, Deepmind yakaburitsa chinyorwa chinonzi "Grandmaster-Level Chess Pasina Kutsvaga", ichikurukura kana zvichikwanisika kushandisa chete inotyairwa nedatha uye kusiya MCTS kutsvaga muAlphaGo neAlphaZero. Zvakafanana nekutyaira kuzvimiririra, network imwe chete ndiyo inoshandiswa kuburitsa zvakananga zviito, nepo matanho ese anotevera asina kufuratirwa.
Chinyorwa chinopedzisa kuti, kunyangwe yakawanda data uye modhi paramita, zvine musoro mhedzisiro inogona kuwanikwa pasina kushandisa kutsvaga. Zvisinei, kune misiyano yakakura kana ichienzaniswa nemaitiro ekushandisa kutsvaga. Izvi zvinonyanya kubatsira kubata nemamwe akaomarara ekupedzisira.
Kune dzakaomarara mamiriro kana makona makesi anoda akawanda-nhanho mitambo, fananidzo iyi ichiri kuita kuti zviome kusiya zvachose optimization yechinyakare kana kutsvaga algorithms. Kushandisa zvine mutsindo zvakanakira akasiyana matekinoroji seAlphaZero ndiyo nzira yakanakisa yekuvandudza mashandiro.
6.Traditional nzira = mutemo-based kana zvimwe?
Ndaifanira kugadzirisa pfungwa iyi kakawanda ndichitaura nevanhu vakawanda. Vanhu vazhinji vanotenda kuti chero bedzi isiri iyo data-inofambiswa, haina mutemo-yakavakirwa. Semuenzaniso, muchess, kuyeuka mafomula uye chess marekodhi nerote ndeye mutemo-yakavakirwa, asi seAlphaGo neAlphaZero, inopa iyo modhi kugona kuve nemusoro kuburikidza nekugadzirisa uye kutsvaga. Handifungi kuti inogona kunzi mutemo-based.
Nekuda kweizvi, modhi yakakura pachayo haisipo parizvino, uye vaongorori vari kuyedza kupa muenzaniso unotungamirwa nekudzidza kuburikidza nenzira dzakadai seCoT. Zvisinei, kusiyana nemabasa anoda kucherechedzwa kwemifananidzo inofambiswa nedata uye zvikonzero zvisingatsananguriki, chiito chese chemunhu anotyaira chine simba rekutyaira rakajeka.
Pasi peiyo algorithm yekuvakisa dhizaini, iyo sarudzo trajectory inofanirwa kuchinjika uye kugadziridzwa zvakafanana pasi pekutungamirwa kwezvinangwa zvesainzi, pane kumanikidzira kupeta nekugadzirisa maparamita kugadzirisa nyaya dzakasiyana. Hurongwa hwakadaro husina marudzi ese emitemo yakaoma-coded isinganzwisisike.
Mhedziso
Muchidimbu, kupera-kusvika-kumagumo kunogona kunge kuri kuvimbisa kwehunyanzvi nzira, asi mashandisirwo anoitwa pfungwa iyi inoda kumwe kutsvagisa. Ini ndinofunga boka re data uye modhi paramita handiyo yega mhinduro yakarurama, uye kana tichida kupfuura vamwe, tinofanira kuramba tichishanda nesimba.
Nguva yekutumira: Kubvumbi-24-2024