UnifoLM-WLA-1.0 Uyasebenza

I-UnifoLM-WLA-1.0 iyimodeli enkulu ye-multimodal embodied engu-6B eyethulwe yi-Unitree Robotics. Iqeqeshwe kusetshenziswa cishe amahora angu-2500 edatha eqoqwe kumarobhothi angempela, ihlanganisa ukubona, ulimi kanye nezenzo ibe yimodeli eyodwa, futhi imodeli eyodwa ilawula imisebenzi engu-64 yokusebenza phezu kwetafula neyokunyakaza komzimba wonke. Kukhishwe izisindo zemodeli yokubona nokucabanga engu-UnifoLM-ER-1 engaba ngu-4B kanye nemodeli yokumodela izenzo i-ER-Flow; lena yokugcina ihlanganisa ukubikezela ukuguquka kwesikhathi esizayo nokufunda izenzo ezihlukanisiwe. Uhlelo oluphelele nekhodi yokuqeqeshwa kwangemva kokuqeqeshwa kuzokhishwa kancane kancane.

Ukubona nokucabanga okugxile emzimbeni: Ihlonza izindawo zezinto, ubudlelwano bendawo kanye nezindawo ezingasetshenzwa, iphinde iphendule imibuzo efana nokuthi “Iphi into okuqondiswe kuyo, futhi buyini ubudlelwano phakathi kwezinto?”, inikeze isisekelo sokubona sokwenza izinqumo zokusebenza.
Ukwenza okuhlanganisiwe kokubona, ulimi nezenzo: Iphendula imiyalelo yomsebenzi ngokuqondile ibe yizenzo ezingenziwa irobhothi, imodeli eyodwa ilawule imisebenzi engu-54 ephezu kwetafula nemisebenzi engu-10 yokusebenza komzimba wonke, ngaphandle kokuqeqesha isu elihlukile ngomsebenzi ngamunye.
Ukubikezela imiphumela yokusebenzisana: Ngaphambi kokwenza isenzo, ibikezela izinguquko zesimo ezidalwa ukusebenza, ngaleyo ndlela inciphise imizamo engasebenzi.
Ukusebenza okuhambisanayo komzimba wonke: Ihlanganisa ukulawula ukusebenza kwezingalo nokunyakaza kwemilenze, futhi ingenza imisebenzi eyinkimbinkimbi edinga ukunyakaza nokusebenza ndawonye, njengokwendlala umbhede, ukugoqa izingubo nokuthulula udoti.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI omthombo ovulekile
Isakhiwo esihlanganisiwe se-VLA: Imodeli yolimi nokubona iyisisekelo; izithombe, imiyalelo yolimi nezenzo kumodeli eyodwa, futhi kumodelwe ndawonye. Imodeli enamapharamitha angu-6B iqeqeshwe ekugcineni kuya ekugcineni kusetshenziswa cishe amahora angu-2500 edatha yerobhothi langempela, kugwenywe ukunqwabelana kwamaphutha okubangelwa ukuhlukanisa ukubona, ukuhlela nokulawula kube amamojula ahlukene.
Ukumelwa kwezenzo ezihlukanisiwe: I-RVQ ihlukanisa indawo yezenzo ehlanganisiwe, okuhlanganisa ukuma kwendawo yokugcina, amalunga esandla esinobuchule namalunga emilenze, ibe ama-token ezenzo. Lokhu kuguqula inkinga yokulawula okuqhubekayo ibe inkinga yokubikezela ama-token, okuyinto amamodeli olimi enza kahle kuyo, futhi imininingwane yokubona, umbhalo nezenzo icutshungulwa ngokulandelana okufanayo.
Ukubikezela ukuguquka kwesikhathi esizayo (ungqimba lwe-ER-Flow): Kusetshenziswa indlela yama-token afihlekile ukuze kubikezelwe ngesikhathi esisodwa ama-token ezenzo nama-token ezindawo ezizoguquka esikhathini esizayo. Lokhu kwenza imodeli iqale icabange ngezinguquko zesimo ezidalwa ukusebenza “ngaphambi kokuthatha isinyathelo”, bese ikhiqiza ukulandelana kwezenzo, kuhlanganiswe ukubona nokwenza kube umjikelezo ovaliwe.
Ukwakhiwa kwemodeli okunezigaba: I-ER-1 iphethe ukubona nokucabanga okuyisisekelo, kanti i-ER-Flow ingeza ukumodela izenzo phezu kwayo. Zombili zakha isisekelo sokubona nongqimba oluphakathi lohlelo lwe-WLA, bese uhlelo oluphelele lwe-6B luhlanganisa ukuqeqeshwa kwangemuva ukuze luthole ikhono lokusebenza komzimba wonke.

Thola ikhodi: Vakashela ikhasi eliyinhloko lephrojekthi ku-unigen-x.github.io/unifolm-wla.github.io noma indawo yokugcina ye-GitHub ethi unitreerobotics/unifolm-wla, bese ulandela i-README ukuze u-clone ikhodi ufake nezindawo ezidingekayo.
Landa izisindo: Landa ku-HuggingFace izisindo ezikhishiwe zamamodeli i-UnifoLM-ER-1 ne-UnifoLM-ER-Flow usebenzisa umyalo we-huggingface-cli download.
Qalisa isibonelo sokucabanga: Qala ngokulayisha i-ER-1 ngeskripthi sokucabanga esinikezwe yiphrojekthi, ufake isithombe samanje serobhothi nencazelo yomsebenzi, bese uhlola ukuthi imiphumela yokubona nokucabanga, njengokuthola indawo okuqondiswe kuyo nokuqonda ubudlelwano bendawo, isebenza kahle yini.
Hlanganisa ukubikezelwa kwezenzo: Layisha i-ER-Flow, ufake isithombe samanje + umyalelo womsebenzi + izenzo ezingakhethwa. Imodeli izokhipha ama-token ezenzo ahlukanisiwe nokubikezela kwezindawo ezizoguquka esikhathini esizayo; kufanele kusetshenziswe indlela esemthethweni yokukhipha ikhodi ye-RVQ ukuze kubuyiselwe amasiginali okulawula aqhubekayo.
Yivumelanise nokulawula irobhothi langempela: Xhuma imiyalelo yezenzo ekhishwe ikhodi ezindaweni zokuxhumana zokulawula ukunyakaza kwamarobhothi afana ne-Unitree G1 / H1, bese uqinisekisa ngezinga elincane ekulingiseni noma erobhothini langempela ngaphambi kokuyisebenzisa, ngoba izisindo ngokwazo azinalo uhlelo oluphelele lokulawula irobhothi.

Imodeli eyodwa, imisebenzi eminingi: Ngamapharamitha angu-6B, ilawula imisebenzi engu-64 iyonke, okuhlanganisa imisebenzi engu-54 ephezu kwetafula nemisebenzi engu-10 yokusebenza komzimba wonke, ngaphandle kokuqeqesha isu elihlukile ngomsebenzi ngamunye, okwenza ibe namandla okujwayela ahamba phambili.
Idatha yerobhothi langempela eqinile: Iqeqeshwe cishe amahora angu-2500 ngedatha yokusebenza yerobhothi langempela, hhayi idatha yokulingisa kuphela, ngakho imiphumela yezenzo isondelene kakhulu nezici zangempela zamandla nokunyakaza kwerobhothi.
Isisekelo sokubona esisebenza kahle: Imodeli eyisisekelo i-ER-1 isekelwe ku-Qwen3-VL-4B futhi iqiniswa ngamasampula angaphezu kwezigidi ezingu-5. Ibe ngowokuqala ekuhlolweni kwamamodeli omthombo ovulekile okungu-7, okuhlanganisa i-RefSpatial-Bench.
Iqala “ibikezele” bese yenza: I-ER-Flow ibikezela izinguquko zesimo zesikhathi esizayo ezidalwa ukusebenza isebenzisa ama-token afihlekile, icabange ngemiphumela yesenzo ngaphambi kokusenza, ngaleyo ndlela inciphise imizamo engasebenzi nobungozi bokushayisana.
Ukulawula umzimba wonke okuhlanganisiwe: I-RVQ ifaka isandla, ukuma kwendawo yokugcina namalunga emilenze kuma-token ezenzo afanayo, ukuze ukusebenza kwezingalo nokunyakaza kwemilenze kwenziwe izinqumo ezihambisanayo kumodeli eyodwa.

Iwebhusayithi yephrojekthi: https://unigen-x.github.io/unifolm-wla.github.io/
Indawo yokugcina ye-GitHub: https://github.com/unitreerobotics/unifolm-wla
Indawo yamamodeli ye-HuggingFace: https://huggingface.co/unitreerobotics/UnifoLM-ER-Flow
https://huggingface.co/unitreerobotics/UnifoLM-ER-1

Izinsizakalo zasekhaya: Yenza imisebenzi yasekhaya yansuku zonke njengokugoqa izingubo, ukwandlala umbhede nokuthulula udoti; lezi yizimo ezisemqoka ezikhonjiswe kakhulu emcimbini wokwethulwa.
Ukukhiqiza kwezimboni: Yenza imisebenzi ephindaphindwayo emigqeni yokukhiqiza, njengokucosha izingxenye, ukuletha amathuluzi nokuhlanganisa okulula.
Ukuphathwa kwempahla nokugcinwa: Yenza imisebenzi yasezinqolobaneni edinga ukuhlanganiswa kokunyakaza nokusebenza, njengokuhlukanisa impahla, ukuhambisa amaphasela nokuhlela amashalofu.
Ucwaningo nemfundo: Njengesisekelo se-VLA sokuphela kuya ekupheleni esivuleke ngokuphelele, isetshenziswa amanyuvesi namalabhorethri ukucwaninga ama-algorithm obuhlakani obugxile emzimbeni nokuqeqesha idatha yokucabanga okugxile emzimbeni.
Ukuqondisa nezinsizakalo zentengiso: Yenza imisebenzi yezinsizakalo exhumana nabantu, njengokuqondisa nokudlulisa izinto, ezindaweni ezifana namahholo ombukiso nezitolo ezinkulu.