1. Gabatarwa
Ƙirƙirar software na sarrafa motoci masu sarrafa kansu yana da wahala a asalinsa, yana buƙatar tsarin ya iya ɗaukar matsaloli marasa iyaka a ƙarƙashin ƙuntatawa na albarkatu. Wannan takarda tana ba da shawarar sabuwar hanyar kaucewa hadarin karo mai amsawa ta amfani da Cibiyoyin Jijiyoyi na Juyin Halitta (ENN). Ba kamar hanyoyin gargajiya da suka dogara da abubuwan da aka ƙayyade ko siffofi da aka ƙera da hannu ba, wannan hanyar tana ba da damar mota ta koyi kai tsaye daga bayanan firji (mai auna nisa guda ɗaya mai fuskantar gaba) don tafiya cikin yanayi masu motsi ba tare da karo ba. An yi horo da tabbatarwa a cikin kwaikwayo, yana nuna ikon hanyar don yin amfani da shi ga yanayin da ba a gani ba.
Matsala ta Asali: Shawo kan iyakokin tsare-tsaren kaucewa hadarin karo, waɗanda ba su da daidaito, a cikin yanayin duniya marasa tsinkaya.
2. Hanyar Aiki
Tsarin da aka ba da shawara ya haɗa cibiyoyin jijiyoyi don fahimta/sarrafawa tare da algorithms na juyin halitta don ingantawa.
2.1 Tsarin Tsarin
Motar kai (ego-vehicle) tana sanye da na'urar auna nisa mai kwaikwayo mai fuskantar gaba. Wannan na'urar firji tana ba da jerin karatun nisa $d = [d_1, d_2, ..., d_n]$ a kusurwoyi masu yawa na kwance, yana samar da sauƙaƙan fahimtar yanayin gaba na nan take. Wannan vector $d$ shine kawai shigarwar cibiyar jijiyoyi mai ci gaba.
Sakamakon cibiyar jijiyoyi shine siginar sarrafawa mai ci gaba don kusurwar tuƙi na mota $\theta_{steer}$. Manufar ita ce a koyi aikin taswira $f$ kamar yadda $\theta_{steer} = f(d)$, wanda ke haifar da tafiya ba tare da karo ba.
2.2 Cibiyar Jijiyoyi ta Juyin Halitta (ENN)
ENN yana nufin cibiyar jijiyoyi wadda ake inganta ma'auni da tsarinta (har zuwa wani mataki) ta amfani da algorithm na juyin halitta, maimakon baya-bayan baya na gargajiya. A cikin wannan mahallin, kowane wakili na mota yana sarrafa shi ta hanyar cibiyar jijiyoyi ta musamman. "Hankali" na wakili yana cikin sigogin cibiyar sa.
2.3 Algorithm na Juyin Halitta don Horarwa
Ana amfani da Algorithm na Juyin Halitta (GA) don haɓaka yawan jama'ar wakilan mota a cikin tsararraki.
- Yawan Jama'a: Saitin wakilan mota, kowannensu yana da cibiyar jijiyoyi ta musamman.
- Ƙimar Ƙarfi: Ana kimanta kowane wakili a cikin kwaikwayon. Ƙarfin $F$ yawanci ana bayyana shi azaman aiki na nisan da aka yi tafiya ba tare da karo ba, misali, $F = \sum_{t} v_t \cdot \Delta t$, inda $v_t$ shine saurin gudu a lokacin $t$ kuma $\Delta t$ shine matakin lokaci. Karon yana haifar da hukunci mai tsanani na ƙarfi ko ƙarewa.
- Zaɓi: Ana zaɓar wakilai masu maki mafi girma na ƙarfi a matsayin "iyaye."
- Haɗin Kai & Canji: Ana haɗa sigogin cibiyar jijiyoyi (ma'auni) na iyaye (haɗin kai) kuma ana canza su ba da gangan ba (canji) don ƙirƙirar "'ya'ya" don tsara na gaba.
- Maimaitawa: Wannan tsari yana maimaitawa, yana haifar da wakilai mafi kyau wajen kaucewa karu.
3. Tsarin Gwaji & Sakamako
Takardar ta tabbatar da hanyar ta hanyar gwaje-gwaje guda shida masu mahimmanci da aka gudanar a cikin kwaikwayo.
3.1 Gwaji na 1: Titin Tsaye Mai Sako-sako
Manufa: Gwada ikon koyo na asali a cikin yanayi mai sauƙi, mai tsayi (misali, titin fanko tare da bango).
Sakamako: Motoci sun yi nasarar koyon tafiya a kan titin ba tare da karo ba, suna nuna ikon ENN na ƙwarewar kaucewa cikas daga bayanan firji marasa yawa.
3.2 Gwaji na 2: Binciken Ƙudurin Na'urar Firji
Manufa: Bincika tasirin ƙudurin kusurwar mai auna nisa (adadin katako $n$) akan aikin koyo.
Sakamako: Aikin ya inganta tare da ƙuduri mafi girma (katako masu yawa), amma an lura da raguwar dawowa. Wannan yana nuna ciniki tsakanin cikakken bayani na fahimta da rikitarwar lissafi/koyo. An gano mafi ƙarancin ƙuduri mai yuwuwa.
3.3 Gwaji na 3: Koyo na Motoci Da Yawa
Manufa: Kimanta hanyar a cikin yanayi mai motsi tare da motoci masu zaman kansu da yawa.
Ƙaramin gwaji 3.3.1: Motar kai guda ɗaya ta koyi guje wa sauran motocin da ke motsawa ba da gangan ba.
Ƙaramin gwaji 3.3.2: Ƙungiyar motoci lokaci guda ta koyi kaucewa karo daga farko.
Sakamako: Hanyar ta yi nasara a cikin duka biyun. Yanayin koyo na wakilai da yawa, lokaci guda yana da mahimmanci musamman, yana nuna fitowar hanyoyin kaucewa masu zaman kansu, masu kama da haɗin gwiwa ba tare da ƙa'idodin sadarwa na zahiri ba.
3.4 Gwaji na 4-6: Gwajin Gabaɗaya
Manufa: Gwada ƙarfi da gabaɗaya na manufar da aka koya.
Gwaji na 4 (Sabon Mai Kwaikwayo): An canza manufar da aka horar a cikin mai kwaikwayo na asali zuwa CarMaker, mai inganci mai inganci, mai kwaikwayon motsin mota na kasuwanci. Motar ta ci gaba da kaucewa karo, yana tabbatar da zaman kai na mai kwaikwayo.
Gwaji na 5 (Sabon Firji): An maye gurbin mai auna nisa na gaba da kyamara. Tsarin ENN, yanzu yana sarrafa bayanan danye/pixel, ya yi nasarar koyon kaucewa karu, yana nuna zaman kai na yanayin firji.
Gwaji na 6 (Sabon Aiki): An ba motar aikin koyon kula da layi ban da kaucewa karo. ENN ta yi nasarar koyon wannan aikin haɗe, yana nuna gabaɗaya aikin.
Mahimman Binciken Gwaji
- Yawan Nasarar a Titin Tsaye: >95% bayan N tsararraki.
- Mafi kyawun Katakon Firji: An samo shi tsakanin 5-9 don yanayin da aka gwada.
- Nasarar Wakilai Da Yawa: Ƙungiyoyin har zuwa motoci 5 sun koyi kaucewa lokaci guda.
- Nasarar Gabaɗaya: An canza manufar cikin nasara a cikin manyan canje-canje guda 3 (mai kwaikwayo, firji, aiki).
4. Binciken Fasaha & Fahimta ta Asali
Fahimta ta Asali
Wannan takarda ba wani ƙarin ci gaba ne kawai a cikin tsarin tafiya ba; hujja ce mai ƙarfi don koyo na amsawa akan kamala na lissafi. Marubutan sun gaza kuskuren mutuwa a cikin tarin na'urori masu sarrafa kansu na gargajiya: dogaro mai yawa akan bututun fahimta masu rauni, waɗanda aka daidaita da hannu da masu tsarawa waɗanda suka gaza a cikin yanayi na gefe. Ta bar Algorithm na Juyin Halitta ya bincika sararin manufa kai tsaye daga firji-zuwa-aiwatarwa, suna ƙetare buƙatar ƙididdigar yanayi na zahiri, bin diddigin abu, da ingantawar hanya. Gwanin gaske yana cikin ƙarancin - mai auna nisa guda ɗaya da umarnin tuƙi. Tunatarwa ne mai ƙarfi cewa a cikin ƙuntatawa, yanayin amsawa mai sauri, manufa mai kyau da aka koya daga bayanai sau da yawa ya fi tsarin da ya isa makara.
Matsala ta Hankali
Dabarun bincike yana da tsabta kuma yana ci gaba da buri. Ya fara da "Sannu Duniya" na na'urori masu sarrafa kansu (kar a buga bango mai tsayi), yana gwada ma'auni mai mahimmanci (ƙudurin firji) a hankali, sannan ya yi tsalle cikin zurfi tare da hargitsi na wakilai da yawa. Abin da ya fi dacewa shine trilogy na gabaɗaya: musanya mai kwaikwayo, firji, da aiki. Wannan ba tabbatarwa kawai ba ne; nuni ne na ƙarfi mai tasowa. Manufar ba ta haddace taswira ko siffofi na musamman ba; tana koyon alaƙar sararin samaniya ta asali: "idan wani abu yana kusa a hanya X, juya zuwa hanya Y." Wannan ka'idar ta asali tana canzawa a cikin yankuna, kamar yadda siffofin gani da CNN ta koya a cikin ImageNet ke canzawa zuwa wasu ayyukan gani, kamar yadda aka tattauna a cikin adabi mai zurfi na koyo.
Ƙarfi & Kurakurai
Ƙarfi:
- Sauƙi Mai Kyau: Tsarin yana da kyau sosai, yana rage matsalar zuwa ainihinsa.
- Gabaɗaya da za a iya tabbatarwa: Gwajin gabaɗaya mai kafa uku shine babban darasi a cikin ingantaccen kimantawa, ya wuce sakamako na yanayi guda ɗaya.
- Yuwuwar Wakilai Da Yawa Masu Zaman Kansu: Gwajin koyo lokaci guda shine hangen nesa mai ban sha'awa zuwa cikin daidaitawar runduna mai iya aunawa, ba tare da sadarwa ba.
- Ramin Kwaikwayo: Duk tabbatarwa yana cikin kwaikwayo. Tsalle zuwa duniyar zahiri - tare da hayaniyar firji, jinkiri, da rikitarwar motsin mota - yana da girma. Gwajin CarMaker mataki ne mai kyau, amma ba duniyar gaske ba ce.
- Rashin Ingancin Samfurin GA: Algorithms na juyin halitta sanannen suna da yunwa (lokacin kwaikwayo) idan aka kwatanta da hanyoyin koyo mai zurfi na zamani (RL) kamar PPO ko SAC. Takardar za ta fi ƙarfi tare da ma'auni na kwatancen da wakili na RL na zamani.
- Ƙuntataccen Sararin Aiki: Sarrafa tuƙi kawai yana yin watsi da taya da birki, waɗanda ke da mahimmanci don kaucewa karo na gaske (misali, tsayawar gaggawa). Wannan yana sauƙaƙa matsalar da za a iya cewa ya yi yawa.
Fahimta Mai Aiki
Ga masu aiki a masana'antu:
- Yi Amfani da Wannan a matsayin Tushe, Ba Magani ba: Aiwatar da wannan hanyar ENN a matsayin ingantaccen, ƙananan Layer na tsaro na tsaro a cikin tarin ku na sarrafa kai. Lokacin da babban mai tsarawa ya gaza ko ba shi da tabbas, ba da iko ga wannan manufar mai amsawa.
- Gina Ramin Sim-to-Real tare da Bazuwar Yanki: Kar a horar da kawai a cikin mai kwaikwayo mai kamala. Yi amfani da ƙarfin GA don horarwa a cikin dubu na kwaikwayo masu bazuwa (haske daban-daban, rubutu, hayaniyar firji) don haɓaka ƙarfin manufa, dabarar da ƙungiyoyin bincike kamar OpenAI suka yi amfani da su.
- Haɗaɗɗe: Maye gurbin vanilla GA don binciken manufa tare da hanyar da ta fi dacewa da samfurin kamar Evolution Strategies (ES) ko amfani da GA don inganta ma'auni na algorithm na RL mai zurfi. Fannin ya matsa daga GA mai tsabta don sarrafawa.
- Faɗaɗa Kayan Firji: Haɗa mai auna nisa na gaba tare da firji mai gajeren zango, mai fadi (kamar kyamara mai ƙarancin ƙuduri a ko'ina) don ɗaukar zirga-zirgar ƙetare da barazanar baya, yana matsawa zuwa ambulaf ɗin tsaro na digiri 360.
5. Tsarin Bincike & Misalin Lamari
Tsarin don Kimanta Manufofin Robotic da aka Koya:
Wannan takarda tana ba da samfuri don ingantaccen kimantawa. Za mu iya zana tsari mai matakai huɗu:
- Gwajin Ƙwarewa ta Asali: Shin zai iya yin aikin asali a cikin yanayi mai sauƙi? (Titin tsaye).
- Binciken Hali na Ma'auni: Ta yaya zaɓin kayan aiki/algorithm ke tasiri aiki? (Ƙudurin firji).
- Gwajin Danniya na Muhalli: Ta yaya yake aiki a ƙarƙashin ƙaruwar rikitarwa da rashin tabbas? (Yanayi mai motsi, yanayi na wakilai da yawa).
- Binciken Gabaɗaya: Shin ƙwarewar da aka koya ta asali ce ko an haddace ta? Gwada a cikin masu kwaikwayo, firji, da ayyukan da suka danganci.
Misalin Lamari: Robotic na Kayan Aiki na Warehouse
Yanayi: Rukunin motocin motsi masu sarrafa kansu (AMRs) a cikin ɗakin ajiya mai motsi.
Aiwatar da Tsarin:
- Gwajin Asali: Horar da robot guda ɗaya (ta amfani da ENN) don tafiya cikin layukan fanko ba tare da buga tarkace ba.
- Binciken Hali: Gwada tare da LiDAR 2D vs. kyamara zurfin 3D. Nemo mafi kyawun farashi/aiki.
- Gwajin Danniya: Gabatar da wasu robots da ma'aikatan ɗan adam suna motsawa ba da gangan ba. Horar da ƙungiya lokaci guda.
- Binciken Gabaɗaya: Canja manufar da aka horar zuwa wani shimfidar ɗakin ajiya daban (sabon "taswira") ko aiki tare da bin takamaiman hanya (kula da layi) yayin guje wa cikas.
6. Ayyuka na Gaba & Hanyoyi
Ka'idojin da aka nuna suna da fa'ida mai faɗi fiye da motocin babbar hanya:
- Jiragen Marufi na Bayarwa na Ƙarshe: Kaucewa amsawa a cikin sararin samaniyar birane mai cunkoso don guje wa cikas mai motsi (misali, tsuntsaye, sauran jiragen marufi).
- Robotic na Noma: Tarakta masu sarrafa kansu ko masu girbi suna tafiya cikin filayen da ba su da tsari, suna guje wa ma'aikata, dabbobi, da ƙasa mara tsari.
- Kujerun Guragu Mai Hankali & Taimakon motsi: Samar da abin dogaro, ƙananan kaucewa karo a cikin wuraren cunkoso na cikin gida (asibitoci, filayen jiragen sama), haɓaka amincin mai amfani tare da ƙaramin shigarwa.
- Cobots na Masana'antu: Ba da damar haɗin gwiwar ɗan adam-robot mafi aminci ta ba da robots wani abu na asali, koyon reflex don guje wa hulɗa, ƙari ga na'urori na ƙarfi na gargajiya.
- Haɗawa tare da Samfuran Hasashe: Haɗa ENN mai amsawa tare da ƙirar duniya mai sauƙi. Layer mai amsawa yana ɗaukar barazana nan take, yayin da Layer na hasashe yana ba da damar tsari mai santsi, mafi hasashe.
- Bayyanawa & Tabbatarwa: Haɓaka hanyoyin don bincika cibiyar jijiyoyi ta juyin halitta. Wane "dokoki" mai sauƙi ya gano? Wannan yana da mahimmanci don takaddun shaida na tsaro a cikin masana'antu masu ƙa'ida kamar mota.
- Haɗakar Firji Mai Yanayi Da Yawa: Haɓaka manufofin da za su iya haɗa bayanai daga firji daban-daban (LiDAR, kyamara, radar) daga ƙasa sama, maimakon haɗawa a matakin fasali.
- Koyo na Rayuwa: Ba da damar manufar don daidaitawa kan layi zuwa sabbin, canje-canjen muhalli na dindindin (misali, sabon gini, yankin gini na dindindin) ba tare da cikakken sake horo ba, watakila ta hanyar ci gaba da tsarin juyin halitta.
7. Nassoshi
- Eraqi, H. M., Eldin, Y. E., & Moustafa, M. N. (Shekara). Kaucewa Hadarin Karon Mota Mai Sarrafa Kanta ta Amfani da Cibiyoyin Jijiyoyi na Juyin Halitta. [Sunan Jarida/Taron].
- Liu, S., et al. (2013). Bincike kan kaucewa karo don jiragen sama marasa matuki. Journal of Intelligent & Robotic Systems.
- Fu, C., et al. (2013). Bita kan tsarin kaucewa karo don motoci masu sarrafa kansu. IEEE Transactions on Intelligent Transportation Systems.
- Sipper, M. (2006). Evolutionary Computation: A Unified Approach. MIT Press.
- OpenAI. (2018). Koyo da Dexterous In-Hand Manipulation. Yana nuna amfani mai ci gaba na kwaikwayo da bazuwar yanki don hadaddun ayyukan robotic. [https://openai.com/research/learning-dexterous-in-hand-manipulation]
- Schulman, J., et al. (2017) Algorithms na Ingantaccen Manufa na Kusa. arXiv:1707.06347. Muhimmin algorithm na ƙarfafawa na zamani don kwatanta da hanyoyin juyin halitta.
- IPG Automotive. CarMaker - Buɗe Dandalin Gwaji don Tuki na Gwaji na Virtual. [https://ipg-automotive.com/products-services/simulation-software/carmaker/]