Amamodeli we-AI aphezulu ayi-10 Anika Amandla Amarobhothi Omhlaba Wangempela ngo-2026

Igebe phakathi kwamakhono emodeli yolimi nokusetshenziswa kwerobhothi liye lehla kakhulu phakathi nezinyanga eziyi-18 ezedlule. Isigaba esisha samamodeli ayisisekelo – injongo-eyakhelwe hhayi ukukhiqiza umbhalo kodwa isenzo somzimba – manje sisebenza ngezingxenyekazi zekhompuyutha zangempela kuwo wonke amafekthri, izindawo zokugcina izimpahla, namalebhu ocwaningo. Lawa masistimu asebenzisa izinqubomgomo zamarobhothi, ama-VLA okuhlola kuqala ayimfihlo, amamodeli ocwaningo lwesisindo esivulekile, namamodeli omhlaba asetshenziswa ukukala idatha yokuqeqeshwa kwamarobhothi. Ezinye ziyahlolwa noma zithunyelwe nozakwethu bezimboni; ezinye ziwucwaningo noma amasistimu abhekene nonjiniyela. Nakhu ukuhlukaniswa kweziyishumi ezibaluleke kakhulu ngo-2026.

I-NVIDIA Isaac GR00T N-Series (N1.5 / N1.6 / N1.7)

I-NVIDIA ikhiphe i-GR00T N1 yoqobo e-GTC ngoMashi 2025 njengemodeli yesisekelo yokuqala evulekile, engenziwa ngokwezifiso ngokugcwele yokucabanga namakhono obuntu obujwayelekile. Uchungechunge lwe-N seluthuthuke ngokushesha. I-GR00T N1.5, eyamenyezelwa e-COMPUTEX ngoMeyi 2025, yethula i-VLM eqandisiwe, ukuthuthukiswa kokhozi 2.5, inhloso yokuqeqeshwa kwe-FLARE evumela ukufunda kumavidiyo e-ego yomuntu, kanye nepulani ye-GR00T-Dreams – eyanciphisa ukukhiqizwa kwedatha yokwenziwa kusukela ezinyangeni kuya cishe emahoreni angu-36.

I-GR00T N1.6 yalandelwa ngoDisemba 15, 2025, enomgogodla omusha wangaphakathi we-NVIDIA Cosmos-2B VLM osekela ukulungiswa okuguquguqukayo, i-DiT engu-2× enkulu (izendlalelo ezingu-32 uma iqhathaniswa no-16 ku-N1.5), izingxenye zesenzo esihlobene nombuso ukuze zinyakaze kalula, kanye nezinkulungwane ezimbalwa zamahora engeziwe edatha ye-teleoperation ye-teleGIM1 ye-TelegiM Unit biman A. G1. Yaqinisekiswa emisebenzini yangempela ye-bimanual kanye ne-locommanipulation kuyo yonke leyo mifaniso.

Ukukhishwa kwakamuva kakhulu, i-GR00T N1.7 Early Access (April 17, 2026), iyipharamitha engu-3B evulekile, enelayisensi yezohwebo ye-VLA eyakhelwe kumgogodla we-Cosmos-Reason2-2B ene-32-layer DiT yokulawula imoto yezinga eliphansi – i-Action Cascade ye-dual-system Architecture. Intuthuko yayo emaphakathi yi-EgoScale: ukuqeqeshwa kusengaphambili ngamahora angu-20,854 wevidiyo yomuntu egxile emikhakheni yemisebenzi engu-20+, okudlula kakhulu amahora okusebenza ngamarobhothi asetshenziswa ezinguqulweni zangaphambili. I-NVIDIA ihlonze lokho ekuchazayo njengomthetho wokuqala ngqa wokukala wobuchule berobhothi – ukusuka emahoreni ayi-1,000 kuye kwangama-20,000 wedatha yomuntu egxile kakhulu kunesilinganiso esiphindwe kabili sokuqedwa komsebenzi. I-N1.7 Early Access iyatholakala ku-HuggingFace ne-GitHub enelayisensi ye-Apache 2.0, ngosekelo olugcwele lokukhiqiza oluboshelwe ekukhishweni kokutholakala okuvamile. Abamukeli bokuqala kulo lonke uchungechunge lwe-GR00T N bahlanganisa i-AeiRobot, Foxlink, NEURA Robotics, kanye ne-Lightwheel.

I-Google DeepMind Gemini Robotics 1.5

I-Gemini Robotics iyimodeli ye-vision-language-action (VLA) ethuthukisiwe eyakhelwe ku-Gemini 2.0, enezenzo zomzimba ezingezwe njengendlela entsha yokukhiphayo yokulawula ngokuqondile amarobhothi. Yethulwe ngoMashi 2025 eceleni kweGemini Robotics-ER (Embodied Reasoning). Isibuyekezo sango-September 2025, i-Gemini Robotics 1.5, yethule amakhono e-ejenti — iguqule imininingwane ebonwayo nemiyalo ibe imiyalo yezimoto kuyilapho yenza inqubo yokucabanga yemodeli ibe sobala, isiza amarobhothi ukuthi ahlole futhi aqedele imisebenzi eyinkimbinkimbi enezinyathelo eziningi ngokucace kakhudlwana.

Ukufinyelela kuhlala kutholakala kozakwethu abakhethiwe okuhlanganisa i-Agile Robots, i-Agility Robotics, i-Boston Dynamics, namathuluzi Enchanted, futhi akutholakali esidlangalaleni. Umndeni obanzi uyaqhubeka nokuvela: I-Gemini Robotics-ER 1.6, ekhishwe ngo-Ephreli 14, 2026, ithuthukisa ukucabanga kwendawo kanye nokuqonda kokubuka okuningi – okuhlanganisa nekhono elisha lokufunda ithuluzi elithuthukiswe ngokubambisana ne-Boston Dynamics lokufunda amageji ayinkimbinkimbi nezibuko zamehlo. I-Gemini Robotics-ER 1.6 iyatholakala konjiniyela nge-Gemini API ne-Google AI Studio.

I-Physical Intelligence π0 / π0.5 / π0.7

π0 iphakamisa ukwakheka okuhambisanayo okugelezayo okwakhiwe phezu kwemodeli yolimi lombono oluqeqeshwe kusengaphambili ukuze kuzuze ulwazi lwe-semantic yezinga le-inthanethi, oluqeqeshwe kuzo zonke izinkundla zamarobhothi ezinobuqili ezihlanganisa amarobhothi anengalo eyodwa, amarobhothi anengalo ekabili, nama-manipulators eselula. I-Physical Intelligence ivule umthombo π0 ngoFebhuwari 2025.

I-π0.5 yashicilelwa ngo-April 22, 2025, ngezisindo ze-openpi ezikhishwe kamuva ngo-2025. Kunokuba iqondise ubuciko obuthuthukisiwe, ukugxila kwayo ukuhlanganisa umhlaba wonke ovulekile: imodeli isebenzisa ukuqeqeshwa ngokubambisana kuyo yonke imisebenzi ehlukahlukene, amarobhothi amaningi, ukubikezela kwe-semantic kwezinga eliphezulu, kanye nedatha yewebhu ukuhlanza amakhishi angajwayelekile abonwa emakhishini okuqeqesha kanye namakamelo okulala. Inguqulo eyalandela yasebenzisa indlela ye-RECAP (RL With Experience & Corrections via Advantage-conditioned Policies) – ukuqeqeshwa ngokubonisa, ukuqeqesha ngokusebenzisa izilungiso, kanye nokuthuthukiswa kokuhlangenwe nakho okuzimele – okubike ukuthi i-Physical Intelligence iphindaphindeka kabili imisebenzi efana nokufaka isihlungi emshinini we-espresso, ukugoqa ikhadi elingakaze libonwe kanye nokuhlanganisa ibhokisi lokuwasha ngaphambilini.

Ukukhishwa kocwaningo lwasesidlangalaleni kwakamuva kakhulu ngu-π0.7, okushicilelwe ngomhla ka-April 16, 2026. Kuwuhlelo lwesigaba socwaningo olugxile ekuhlanganiseni okujwayelekile: ukuhlanganisa amakhono afundiwe kusuka kuzimo ezihlukene ukuze kuxazululwe imisebenzi imodeli engazange iqeqeshwe ngokusobala. I-Physical Intelligence iyichaza njengemodeli elawulekayo enamakhono avelayo – isinyathelo sangaphambi kwesikhathi kodwa esiphusile esiya ebuchosheni berobhothi obunenjongo evamile. Iphepha lisebenzisa ulimi lokubiyela ngokucophelela kulo lonke, futhi awukho umugqa wesikhathi wokuthunyelwa kwezentengiso oshiwo.

Umfanekiso we-AI Helix

Ikhishwe ngoFebhuwari 20, 2025, i-Helix iyi-VLA yokuqala ukukhipha izinga eliphezulu, ukulawula okuqhubekayo kwawo wonke umzimba ongaphezulu we-humanoid, okuhlanganisa izihlakala, i-torso, ikhanda, neminwe ngayinye. Isebenzisa umklamo wohlelo olukabili: Isistimu 2 iyipharamitha ye-inthanethi ye-VLM engu-7B eqeqeshelwe kusengaphambili esebenza ku-7–9 Hz ukuze kuqondwe indawo kanye nokuqonda ulimi; Isistimu 1 iyipharamitha engu-80M ye-cross-attention encoder-decoder transformer esebenza ngo-200 Hz, ehumusha izethulo ze-S2 ze-semantic zibe izenzo zerobhothi ezinembile. Imodeli yaqeqeshwa cishe emahoreni angu-500 edatha yamarobhothi amaningi, esebenza ngezingcingo eziningi, ngokulebula okuzenzakalelayo kweziyalezo nge-VLM esetshenziswa ngemuva. Zonke izinto zokuqeqeshwa azifakwa ekuhlolweni ukuze kuvinjelwe ukungcoliswa.

I-Helix isebenzisa ngokuphelele ama-GPU ashumekiwe asebenzisa amandla aphansi, iyenze ifaneleke ocwaningweni lokusatshalaliswa kohwebo kanye nezinhlelo zokusebenza zesikhathi esizayo ze-humanoid. Isebenzisa isethi eyodwa yezisindo zenethiwekhi ye-neural kukho konke ukuziphatha – ukucosha nokubeka izinto, kusetshenziswa amadrowa neziqandisi, nokusebenzisana kwamarobhothi – ngaphandle kwanoma yikuphi ukuhlela kahle okuqondene nomsebenzi othize. Kuye kwaboniswa emisebenzini yokukhohlisa yasendlini kanye nephakheji yokuhlela, futhi ingasebenza ngesikhathi esisodwa kumarobhothi amabili ngesakhiwo sokuqondisa esihlukanisa imigomo yonke ibe imisebenzi engaphansi yerobhothi ngalinye.

I-OpenVLA

I-OpenVLA iyipharamitha engu-7B yomthombo ovulekile we-VLA oqeqeshwe eqoqweni elihlukahlukene lemibukiso yamarobhothi yomhlaba wangempela engu-970,000. Yakha phezu kwemodeli yolimi lwe-Llama 2 ehlanganiswe nesifaki khodi esibonakalayo esihlanganisa izici eziqeqeshwe kusengaphambili ezivela ku-DINOv2 ne-SigLIP. Naphezu kokuba incane ngo-7×, i-OpenVLA idlula i-RT-2-X (amapharamitha angu-55B) evaliwe ngamaphesenti angu-16.5 ngezinga eliphelele lempumelelo yomsebenzi kuyo yonke imisebenzi engama-29 kanye nokufanekisa amarobhothi amaningi.

Iphepha langoFebhuwari 2025 lethule iresiphi ye-OFT (Optimized Fine-Tuning), ehlanganisa ukukhishwa kwekhodi okuhambisanayo, ukuhlukaniswa kwesenzo, ukumelwa kwesenzo esiqhubekayo, kanye nenjongo yokuhlehla ye-L1. I-OFT iletha isivinini esisheshayo esingu-25–50x futhi ifinyelela izinga lempumelelo elimaphakathi elingu-97.1% kubhentshimakhi yokulingisa ye-LIBERO, esebenza kahle kakhulu ku-π0, Octo, kanye Nenqubomgomo Yokusabalalisa. Inguqulo ethuthukisiwe, i-OFT+, ingeza isimo se-FiLM ukuze kusekelwe ulimi olungcono futhi inike amandla ukulawulwa kwe-bimanual ye-high-frequency bimanual kurobhothi ye-ALOHA. I-OpenVLA isekela ukulungiswa kahle kwe-LoRA nokubalwa komthamo wokusetshenziswa okuvimbelwe yinsiza, futhi ukusonga komphakathi kwe-ROS 2 kukhona ukuze kuhlanganiswe nezinhlelo zokusebenza zamarobhothi.

Okthoba

I-Octo yinqubomgomo yerobhothi elivela kumthombo ovulekile elivela ku-UC Berkeley, etholakala ngosayizi ababili: I-Octo-Small (amapharamitha angu-27M) kanye ne-Octo-Base (amapharamitha angu-93M). Zombili zisebenzisa i-transformer backbone ene-diffusion decoding, eqeqeshwe kusengaphambili ngeziqephu zamarobhothi ezingu-800,000 kusukela kudathasethi yedathasethi ye-Open X-Embodiment kuwo wonke amasethi wedatha angu-25. Imodeli isekela kokubili imiyalelo yolimi lwemvelo nesimo sesithombe somgomo, futhi ivumela ukubhekwa okuguquguqukayo nezikhala zesenzo ezihlanganisa izinzwa ezintsha nezethulo zesenzo ngaphandle kwezinguquko zezakhiwo.

I-Octo yakhelwe ngokukhethekile ukusekela ukulungiswa okusebenzayo ekusetheni amarobhothi amasha. Ekuhloleni okusemthethweni, umsebenzi ngamunye usebenzisa cishe imibukiso yesizinda esiqondiwe esiyi-100, futhi u-Octo wenza kahle kakhulu ngokuqeqeshwa kusukela ekuqaleni ngesilinganiso esingu-52% kuzo zonke izinhlelo zokuhlola eziyisithupha ezihlanganisa izikhungo ezihlanganisa i-CMU, i-Stanford, ne-UC Berkeley. Isebenza ngokulinganayo ne-RT-2-X (amapharamitha angu-55B) kuzilungiselelo ze-zero-shot kuyilapho ingama-oda obukhulu obuncane. I-Octo ngokuyinhloko iyithuluzi lokucwaninga nelonjiniyela, futhi iyindawo yokuqala eqinile engasindi yamalebhu adinga ukuphindaphinda ngokushesha emisebenzini emisha yokukhohlisa enekhompuyutha elinganiselwe.

I-AGIBOT BFM ne-GCM

Ngo-Ephreli 2026, i-AGIBOT ese-Shanghai yamemezela amamodeli amabili ayisisekelo njengengxenye ye-“One Robotic Body, Three Intelligences” yokwakhiwa kwesiteki esigcwele. I-Behavioral Foundation Model (BFM) ibekwe eduze kokulingisa nokudluliselwa kokuziphatha – eklanyelwe ukuthola indlela entsha yokuziphatha enyakazayo ngempumelelo emibonisweni. I-Generative Control Foundation Model (GCFM) ibekwe eduze kokukhiqiza ukunyakaza kwamarobhothi okuqaphela umongo kusuka kokokufaka kwe-multimodal okuhlanganisa umbhalo, umsindo, nevidiyo.

I-AGIBOT ibeka i-AGIBOT WORLD 2026 njengengxenye yesisekelo sedatha yesitaki sayo samarobhothi esibanzi – umthombo ovulekile, isethi yedatha yomhlaba wangempela yezinga lokukhiqiza ehlanganisa izindawo zentengiso, amakhaya, nezimo zansuku zonke. Inkampani imemezele u-2026 ngokuthi “Unyaka Wokuqala Wokuthunyelwa” eNgqungqutheleni Yozakwethu Yango-April 2026 futhi yamemezela ukukhishwa kwerobhothi layo le-10,000 ngoMashi 2026.

I-Gemini Robotics Kudivayisi

I-Gemini Robotics On-Device iyimodeli ye-VLA yamarobhothi e-bi-arm enziwe ukuthi asebenze endaweni irobhothi ngokwalo elinombono wokubambezeleka okuphansi, ngaphandle kokudinga uxhumano lwenethiwekhi yedatha. Ikhishwe ngoJuni 2025, iyimodeli yokuqala ye-VLA i-Google DeepMind eyenzile yatholakala ukuze ilungiswe kahle. Yakha phezu kokwenziwa komsebenzi kanye namakhono obuhlakani emodeli ye-Gemini Robotics esekelwe emafini, ethuthukiselwe ukusebenza kudivayisi lapho ukubambezeleka noma izithiyo zokuxhuma zisebenza. Imodeli yaqeqeshwa ngokuyinhloko kumarobhothi e-ALOHA futhi isiguqulelwe ku-Bi-arm Franka FR3 kanye ne-Apptronik's Apollo humanoid. Ijwayelana nemisebenzi emisha ngemiboniso embalwa efika ku-50 kuye kweyi-100. Ukutholakala okwamanje kubahloli abakhethiwe abathembekile, hhayi ukukhishwa okujwayelekile komphakathi.

I-NVIDIA Cosmos World Foundation Models

I-Cosmos ayiyona imodeli yenqubomgomo yerobhothi ngomqondo ojwayelekile — iyimodeli yomhlaba ekhiqizayo ekhiqiza idatha yokwenziwa yomzila ukuze kukale amapayipi okuqeqesha amanye amamodeli akulolu hlu. Ipulani ye-GR00T-Dreams isebenzisa i-Cosmos ukuze ikhiqize inani elikhulu ledatha ye-synthetic trajectory evela esithombeni esisodwa kanye nokufundiswa kolimi, okuvumela amarobhothi ukuthi afunde imisebenzi emisha ezindaweni ezingajwayelekile ngaphandle kokudinga idatha ethile yokusebenza ngocingo. Lokhu kusekele ngokuqondile ukuthuthukiswa kwe-GR00T N1.5. I-Cosmos Predict 2, inguqulo esetshenziswe ku-GR00T-Dreams, iyatholakala ku-HuggingFace ngezithuthukisi zokusebenza zekhwalithi ephezulu yomhlaba kanye nokubona izinto ezingekho. Izinkampani ezihlanganisa i-Skild AI ne-FieldAI zisebenzisa izingxenye zokulingisa ze-Cosmos ne-Isaac ukuze zenze idatha yokwenziwa kwamarobhothi yokuqeqeshwa futhi ziqinisekise ukuziphatha kwamarobhothi ngokulingisa ngaphambi kokusetshenziswa komhlaba wangempela.

I-SmolVLA (HuggingFace LeRobot)

Ikhishwe ngoJuni 3, 2025, i-SmolVLA iyi-VLA ehlangene ye-HuggingFace engu-450M eyakhelwe ngaphakathi kohlaka lwe-LeRobot futhi yaqeqeshwa ngokuphelele ngedatha yomthombo ovulekile enikelwe ngumphakathi. Isebenzisa umgogodla wolimi lombono we-SmolVLM-2 ohlanganiswe nochwepheshe besenzo se-transformer esihambisana nokugeleza – ukukhipha izenzo eziqhubekayo esikhundleni samathokheni angabonakali, ukumelela okufanayo kwesenzo esisetshenziswa yi-π0 kanye ne-GR00T N1. Iqeqeshwe kusengaphambili kumafreyimu ayizigidi ezingu-10 akhethwe kusukela kumadathasethi omphakathi angu-487 amakwe ngaphansi kokuthi “lerobot” ku-HuggingFace, ehlanganisa izindawo ezihlukahlukene kusukela kumalebhu kuya kumagumbi okuhlala.

I-SmolVLA isebenzisa ihadiwe yabathengi ehlanganisa ama-GPU ekilasi le-RTX elilodwa nama-MacBooks. Amabhentshimakhi wokushuna kahle asemthethweni abonisa cishe amahora ama-4 ku-A100 eyodwa ezinyathelweni zokuqeqesha ezingu-20,000. Ekuhlolweni kwerobhothi langempela kusetshenziswa izingalo ze-SO100 kanye ne-SO101, ifinyelela cishe u-78.3% wesilinganiso sempumelelo esimaphakathi ngemva kokulungisa kahle umsebenzi othile. Ihambisana noma yenza kahle kakhulu kunamamodeli amakhulu afana ne-ACT ku-LIBERO kanye ne-Meta-World yokulingisa amabhentshimakhi, futhi isekela ukucatshangelwa okuvumelanayo kokuphendula okusheshayo okungu-30% kanye nokuphuma komsebenzi okungu-2×. I-SmolVLA iyindawo yokungena efinyeleleka kakhulu ku-VLA ecosystem yamaqembu anekhompiyutha elinganiselwe.