LingBot-Touch Uyasebenza

I-LingBot-Touch iyimodeli eyisisekelo eqeqeshwe kusengaphambili yokuthinta, ekhishwe umphakathi yi-Ant LingBo, esebenza njengomhumushi phakathi kwezinzwa zokuthinta namamodeli ezenzo zamarobhothi. I-LingBot-Touch ihlanganisa amasiginali ezinzwa ezihlukahlukene, ezifana nokuthinta kazibuthe, ukuthinta okubonakalayo kanye ne-piezoresistive, iwafake ngendlela efanayo kuma-Token okuthinta, bese ikhipha izici ezingaqondwa imodeli yesenzo, njengendawo yokuxhumana, ubukhulu nam方向 yamandla. Lokhu kuxazulula inkinga yokuthi ukufaka amasiginali aluhlaza ngokuqondile kunganciphisa izinga lempumelelo. I-LingBot-Touch iqeqeshwe kusetshenziswa indlela ye-JEPA, isebenzisa i-Transformer backbone eyodwa yokuthinta, inamapharamitha angaba yizigidi ezingu-5 kuphela, futhi isiqinisekiswe emisebenzini engu-26 yemishini yangempela. Ingadlulisela ulwazi ngaphandle kokufunda kabusha iye ezinzwa engakaze ibonwe, igcwalise isixhumanisi esibalulekile phakathi kokuthi umbono uphatha ngaphambi kokuxhumana nokuthi ukuthinta kuphatha ngemva kokuxhumana.

Ukuhlanganisa amasiginali ezihlukahlukene: Ihlanganisa amasiginali aluhlaza ezinzwa ezintathu, okungukuthinta kazibuthe, ukuthinta okubonakalayo kanye ne-piezoresistive, iwahumushele kuma-Token okuthinta ajwayelekile, ukuze imodeli yesenzo iqonde idatha yokuthinta yamafomethi ahlukahlukene.
Ukuzwa isimo sokuxhumana: Ihlaziya indawo yokuxhumana, ubukhulu bamandla kanye noshintsho lwendlela ngokuhamba kwesikhathi, ukuze irobhothi lazi kahle ukuthi “iminwe ithinte kuphi, isebenzise amandla angakanani, nokuthi kukhona yini ukushelela”.
Ukudluliselwa phakathi kwezingxenyekazi zekhompuyutha: Isebenzisa indawo yokuxhumanisa eyodwa kanye nesakhiwo se-backbone eyabiwe, ngakho ingadluliselwa ezinzwa nasekuhlelweni kwezandla ezihlakaniphile ezingakaze zibonwe ngaphandle kokuqeqeshwa kabusha.
Ukuxhuma nokusebenzisa ngokushesha: Njengesifaki khodi sezici esiqandisiwe, ingaxhunywa ngokuqondile kumamodeli wezenzo adumile afana ne-π0.5, i-Diffusion Policy kanye ne-T-Rex, futhi yengeze amapharamitha angaba ngu-0.13%–0.16% kuphela.
Ukucubungula okusheshayo nokusebenza kahle: Inguqulo encane kunazo zonke, i-HaT-Tiny, inamapharamitha ezingaba izigidi ezingu-5.46 kuphela futhi isekela ukucubungula kwesikhathi sangempela okuvama kakhulu, okuhlangabezana nezidingo zokusebenza ngesikhathi sangempela.
Ukwandisa izinga lempumelelo yokusebenza: Emisebenzini engu-26 yemishini yangempela, izinga lempumelelo elijwayelekile lenyuke kakhulu ngamaphuzu angaphezu kuka-20, ikakhulukazi emisebenzini yokuxhumana enemininingwane efana nokudonsa uziphu, ukujija isivalo sebhodlela kanye nokufaka istraw.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokushintshisana ngamaphrojekthi e-AI omthombo ovulekile
Isixhumi esibonakalayo sejiyomethri esijwayelekile kanye nendlela yekheli lama-Token: Ukuze kubhekwane nomehluko omkhulu enanini nasekuhlelweni kwezindawo zokuthinta zezinzwa ezihlukahlukene, imodeli iqala ngokwenza wonke amaphuzu okuthinta abe sendaweni eyodwa ejwayelekile yezixhumanisi zeminwe. I-Token ngayinye inikezwa “ikheli” elakhiwe isikhathi, inombolo yomunwe kanye nendawo enezinhlangothi ezintathu, bese lifakwa kumodeli nge-rotary positional encoding (RoPE). Lokhu kususa ukuncika enanini elinqunyiwe nasekuhlelweni kwezindawo zokuthinta, kwakhe okokufaka okunwebekayo okusekelwe emakhelini.
Isakhiwo esixubile esinezindlela ezihlukene zokungena kanye ne-backbone eyodwa: Imodeli yakhela uhlobo ngalunye lwezinzwa ezintathu i-Tokenizer ekhethekile elula, eguqula ukufundwa okuluhlaza kube ama-Token futhi igaye umehluko wehadiwe phakathi kwezinzwa. Ngemva kwalokho wonke ama-Token ahlangana ku-backbone ye-Transformer yokuthinta eyabiwe, ukuze imithetho ehlukahlukene ye-physics yokuxhumana ifundwe ngokuhlanganyela endaweni efihlekile eyodwa.
Ukubikezela indawo yezici okusekelwe ku-JEPA: Njengoba ukufundwa kwamasiginali okuthinta isikhathi esiningi kungu-“0” futhi imodeli ingase ikhethe indlela emfushane yokufunda isizinda, ngesikhathi sokuqeqesha kufihlwa ngokungahleliwe u-90% wezindawo zokuthinta futhi ukufihlwa kuqhubeke isikhathi eside. Imodeli isebenzisa kuphela amasiginali angu-10% abonakalayo asele ukubikezela ukumelwa okufihlekile kwengxenye efihliwe endaweni yezici, okuyiphoqa ukuthi ifunde ulwazi lwangempela lokuxhumana.
Ukuphazamisa izindawo zokuthinta ukuvimbela ukukhumbula ukuhlelwa: Njengoba ukuqeqeshwa kwangaphambili kuhlanganisa ukuhlelwa kwezinzwa ezintathu kuphela, imodeli ingase ibambe ngekhanda izindawo zokuthinta. Ngakho ngesikhathi sokuqeqesha kushintshwa ngokungahleliwe ukuminyana kokusampula futhi indawo yephuzu ngalinye lokuthinta iphazanyiswe kancane, kwakhiwe ngokwenziwa ukuhlelwa kwesine ukuze kuqinisekiswe ukuthi imodeli ifunda ukumelwa kwe-physics yokuxhumana okudlulisekayo.

Ukuthola imodeli: Landa umbiko wobuchwepheshe, ikhodi kanye nezisindo zemodeli ezivela ku-GitHub, ku-Hugging Face noma emphakathini we-Moda.
Ukuxhuma izinzwa: Faka amasiginali aluhlaza ezinzwa zokuthinta kazibuthe, zokuthinta okubonakalayo noma ze-piezoresistive zesandla esihlakaniphile esikhona kumodeli, ngaphandle kokushintsha ihadiwe noma ukulungisa izinzwa uqobo.
Ukufaka amakhodi ezici: I-LingBot-Touch ifaka amakhodi amasiginali aluhlaza ibe ama-Token okuthinta ajwayelekile, ikhiphe ukumelwa okuqondakalayo kwendawo yokuxhumana, amandla kanye nendlela.
Isu lokuqeqesha ngemva kwalokho: Ngesikhathi uqeqesha imodeli yesenzo, efana ne-π0.5 noma i-Diffusion Policy, sebenzisa izici eziqandisiwe ze-LingBot-Touch njengokufaka okwengeziwe ekuqeqeshweni kwangemuva; amapharamitha emodeli awadingi ukushintshwa.
Ukusetshenziswa nokucubungula: Faka isu eliqeqeshiwe emshinini wangempela. I-LingBot-Touch isebenza ngesikhathi sangempela njengesifaki khodi esiqandisiwe, futhi yengeza amapharamitha angaba ngu-0.13%–0.16% kuphela.

Idatha ehlukahlukene ehlanganisiwe: Ihlanganisa amasiginali ezinzwa ezinamafomethi ahlukene kakhulu, ezifana nokuthinta kazibuthe, ukuthinta okubonakalayo kanye ne-piezoresistive, iwafake kuma-Token ajwayelekile, ixazulule inkinga yemboni yokungahambisani kwedatha yokuthinta nokuthi ukushintsha isandla kudinga ukuqala phansi.
Ukudluliselwa kwehadiwe ngaphandle kokufunda izibonelo: Ngisho noma inzwa ingakaze ivele kudatha yokuqeqeshwa kwangaphambili, njenge-XHand, futhi kungashintshiwe nopharamitha owodwa, izinga lempumelelo lisuka ku-48% liye ku-74.17%, okwenza “ulwazi lwehadiwe eyodwa lube yisiqalo serobhothi elilandelayo”.
Ukuxhuma nokusebenzisa ngokushesha, nokungenelela okuncane: Njengesifaki khodi esiqandisiwe, ingaxhunywa ngokuqondile kumamodeli wezenzo adumile afana ne-π0.5 ne-Diffusion Policy, futhi yengeze amapharamitha angu-0.13%–0.16% kuphela, ngaphandle kokuphazamisa isakhiwo semodeli sokuqala.
Ilula kakhulu futhi isebenza kahle: Inguqulo encane kunazo zonke, i-HaT-Tiny, inamapharamitha ezingaba izigidi ezingu-5.46 kuphela, isekela ukucubungula kwesikhathi sangempela okuvama kakhulu futhi ihlangabezane nezidingo zokuphendula ngezinga lama-millisecond.
Ukunemba okuphakeme kakhulu: Iphutha lokunquma isikhungo sokuxhumana lingamamilimitha angu-0.12–0.16 kuphela, lingcono kakhulu kune-FTP-1 enamamilimitha angu-2.19–5.89, futhi le yakamuva idinga i-adaptha ehlukile yesandla ngasinye.
Izindleko zokulebula ziphansi kakhulu: Idatha yokuqeqeshwa kwangaphambili idinga kuphela ukulebula uhlobo lomsebenzi, ngaphandle kokulebula ubukhulu bamandla. Izingxenye ezingu-2/3 zedatha ziqoqwe kusetshenziswa amagilavu, okwehlisa kakhulu umgoqo wokuthola idatha yokuthinta.

Iwebhusayithi yephrojekthi: https://technology.robbyant.com/lingbot-touch
Indawo yokugcina ye-GitHub: https://github.com/Robbyant/lingbot-touch
Iqoqo lamamodeli le-HuggingFace: https://huggingface.co/collections/robbyant/lingbot-touch
Iphepha lobuchwepheshe: https://arxiv.org/pdf/2611.12806

Ukuhlunga ezitolo: Ezindaweni ezifana namakhemisi nezitolo ezidayisa izinto ezilula, ibamba izimpahla ezinobukhulu nobumo obuhlukahlukene, isebenzise ukuthinta kweminwe ukunquma amandla okubamba nokuthi kukhona yini ukushelela, yehlise izinga lokulimala kwezimpahla.
Ukuhlunga kwezimpahla: Lapho ihlunga amaphakheji athambile nezinto eziphuka kalula, ezifana nezimpahla ezisezikhwameni noketshezi olusemabhodleleni, isebenzisa ukuthinta ukuzwa ukuguquka kwesimo sephakheji nokunyakaza kwangaphakathi, ukuze ibambe futhi ibeke ngokuzinzile.
Ukuhlanganisa ezimbonini: Yenza imisebenzi yokuhlanganisa enembile efana nokufaka nokukhipha, ukujija nokubeka endaweni efanele, njengokujija izivalo zamabhodlela nokufaka ukhiye ukuze kuvulwe, iqhubeke nokuzwa izinguquko zokumelana ngesikhathi sokuxhumana lapho umbono wekhamera uvinjiwe.
Ukusebenza kokudla neziphuzo: Yenza imisebenzi enemininingwane efana nokufaka istraw, ukukhama umuthi wokuxubha amazinyo kanye nokuvala izikhwama, izwe ukushintsha okukhulu kwamandla ngesikhathi ifilimu ibhobozwa, igweme ukwehluleka okubangelwa amandla amaningi.