Airffic InstituteAERIAL PHYSICAL AI
Okuqukethwe ☰

Aerial Physical AI Institute

I-Vision AI: indlela idrone eliqonda ngayo lokho elikubonayo

Ukubona akwanele ukuze undize. Ubuchopho budinga ukuqonda lokho okubonayo.

Impendulo emfushane: I-Vision AI yamadrone ithuthuke ngezigaba ezine: ① ukuthola izinto (detection), ② ukuzilandela (tracking), ③ ukwazi indawo yakho (SLAM, VIO), ④ ukuqonda imiyalo ngamagama (ama-foundation model). Simi emnyango wesigaba ④.

Kubuyekezwe: · Imizuzu engu-3 yokufundaIndlela esacwaninga ngayo →

Sir Airffic GoOkuningi lapha kuvela emaphepheni ocwaningo, kodwa amagama alula. Hamba nami.

Izigaba ezine zokuqonda

  1. Ukuthola: thola abantu, izimoto nezintambo zikagesi esithombeni bese udweba amabhokisi azungeze.
  2. Ukulandela: landela ukuthi into etholiwe iya kuphi ngokulandelayo.
  3. Ukuzithola: thola ukuthi uhambe kangakanani futhi waya kuphi ngokunyakaza kwesithombe kanye ne-IMU (VIO: visual-inertial odometry).
  4. Ukuqonda nomyalo: qonda ukuthi “hlola ngemuva kwendlu enophahla olubomvu” bese uxhumanisa amagama nalokho okubonayo (VLM: i-AI ephatha izithombe nolimi ndawonye).

Ukundiza ngenkathi wenza imephu: i-SLAM ne-VIO

Ngasekupheleni kweminyaka yo-2000, amalabhorethri amanyuvesi aqala ukulinganisa indawo yedrone elincane ngekhamera eyodwa ne-IMU. Konke kwaqala lapho.

Izindlela ezifana ne-SVO, i-VINS-Mono ne-ORB-SLAM zashicilelwa zaba yizisekelo ezabiwe. Ucwaningo oluvulekile lwaba yizingxenye zemikhiqizo.

Amadrone okuhweba anamuhla angandiza ngaphansi kwamabhuloho nangaphakathi ngenxa yalo msebenzi oqoqekile.

Ubufakazi: [P01] [P02] [P08] [P09] [P21] [P15]

Usuka phezulu, izinto zibukeka zincane futhi zigobile

I-AI eqeqeshwe ngezithombe zasemhlabathini iyahlupheka ngokubuka okusuka phezulu: abantu baba ngamachashazi futhi izimoto zishintsha ubunjalo.

Amaqoqo edatha yasemoyeni afana ne-VisDrone ne-DOTA akhiwa ukuze kuqeqeshwe i-AI eyenzelwe isibhakabhaka.

Lowo onedatha yilowo ololonga ukuthi i-AI ihlakanipha kangakanani. Lokho akushintshile nasenkathini yama-foundation model.

Sir Airffic GoUmehluko we-AI uvame ukuncika kunani ledatha yasemoyeni onayo, ngaphezu kwama-algorithm. Khumbula lokhu esahlukweni samasu.

Ubufakazi: [P14] [P23] [P24]

Usuku i-AI yahlula ngalo ompetha abangabantu

Ngo-2023 idrone le-AI elibizwa ngokuthi yi-Swift, elakhiwa eNyuvesi yaseZurich, lahlula abagijimi bamadrone abangompetha bomhlaba. Lo msebenzi washicilelwa ku-Nature.

I-Swift yathola indawo yayo ngekhamera ne-IMU kuphela, futhi yafunda ukundiza nge-reinforcement learning — izijwayeza izikhathi eziningi ku-simulation.

Lokhu kwakusenkundleni engashintshi. Ukufinyelela lelo khono esibhakabhakeni sangempela, nesimo sezulu nezithiyo ezishintshayo, kusewuselelo olusavulekile.

Ukugijima kwamadrone ezingeni lompetha nge-deep reinforcement learning (UZH)

Ividiyo yedrone le-AI i-Swift eshicilelwe yi-Robotics and Perception Group yeNyuvesi yaseZurich. YouTube ↗ · Akukho okulayishwa ku-YouTube uze ucindezele u-play

Ubufakazi: [N124]

Igagasi elilandelayo: bona, cabanga futhi wenze ngamagama

Kusukela cishe ngo-2023, amamodeli amakhulu e-AI aphatha izithombe nolimi ndawonye (ama-foundation model) aqala ukusebenza njengobuchopho bamarobhothi.

Ucwaningo lwamadrone mayelana nokulandela imiyalo ekhulunywayo ngenkathi liqonda okubonwayo (aerial vision-language navigation) luyakhula.

Amamodeli amakhulu ayasinda, futhi ukuwasebenzisa ngokushesha phakathi kwedrone kusenzima. Yingakho ama-chip — isahluko esilandelayo — kufanele afundwe ndawonye nalokhu.

Ubufakazi: [N133]

Imibuzo evame ukubuzwa

Yimiphi imisebenzi esebenzisa ukubona kwekhompyutha kwamadrone?

Ukuthola imifantu emabhulohweni nasezintanjeni zikagesi, ukuhlola impilo yezitshalo, ukuthola abantu ekuseshweni nasekutakulweni, ukuqapha ukuphepha nethrafikhi, nokuhlola izindawo zokulethwa kwempahla.

Iyini i-visual-inertial odometry (VIO)?

Indlela ehlanganisa ukunyakaza kwesithombe nezilinganiso ze-IMU zokusheshisa nokujikeleza ukuze kubalwe ukuthi idrone lihambe kangakanani. Ubuchwepheshe obuyisisekelo bokundiza ngaphandle kwe-GPS.