Open-RAIL Uyasebenza

I-Open-RAIL iyisisekelo sobunjiniyela esijwayelekile sobuhlakani obugxile emzimbeni, esivulwe yi-China Mobile emhlabeni wonke, futhi iyisistimu yezinzwa ejwayelekile yokuqala embonini exhumanisa amamodeli e-VLA/WAM nemizimba yamarobhothi. Ezingeni lohlaka, i-Open-RAIL ixhumanisa umjikelezo ophelele wokucabanga—ukwenza—ukubuyiselwa kwedatha—nokuthuthukiswa kwemodeli. Amaphayiphi asebenza ngokuhambisana athuthukisa ubushelelezi bokunyakaza ngokuphindwe izikhathi eziyikhulu, kanti ikhodi yokuyifaka ku-end, edge noma cloud ayidingi ukushintshwa. Ukuxhuma irobhothi elisha kwehliswa kusukela emavikini kuya emahoreni, futhi imodeli entsha idinga imigqa yekhodi engu-50—100 kuphela. Njengamanje isivele isekela amarobhothi ahlukahlukene angu-4 namamodeli amakhulu angu-10; irobhothi elihambayo i-Xingchu selisebenza njalo e-Hangzhou.

Ukucabanga nokwenza: Ihlukanisa ukubuka, ukucabanga nokulawula kube ngamaphayiphi amathathu asebenza ngokuhambisana, ihlanganiswe namasu amabili okushelelisa, ukuze ukuphambuka okujwayelekile kokushesha kwamalunga kwehle kusuka ku-10+ rad/s² kuye ku-0.1 rad/s², okwenza ukunyakaza kuzinze futhi kube bushelelezi.
Ukufakwa ku-end, edge ne-cloud: Ukucabanga kubekwa ngokufanayo ohlangothini lwe-Server, kanti irobhothi lenza kuphela umsebenzi wokwenza. Ngokushintsha ikheli lokuxhumana endaweni eyodwa, kungashintshwa phakathi kwe-end, edge ne-cloud njengoba kudingeka; amandla okubala awasakhawuleli ukukhethwa kwemodeli.
Ukusekela amarobhothi ahlukahlukene: Ngesendlalelo esilula sokukhipha ihadiwe kanye nesivumelwano esifanayo sokuxhuma amamodeli, izixhumi zokulawula nezimo ziyalinganiswa. Ukushintsha irobhothi noma imodeli akudingi ukuqala phansi.
Umjikelezo wedatha: Ukuqoqwa kwedatha kwakhelwe kukho konke ukucabanga nokwenza. Izithombe ezivela emibukweni eminingi, izimo zamalunga nemiphumela yokucabanga kugcinwa ngokuzenzakalelayo ngefomethi evamile elungele ukusetshenziswa, futhi kungafakwa ngqo ekuqeqesheni okulandelayo.
Ukuhlanganiswa kokucabanga nokulawulwa kude: Lapho imodeli iphambuka, umuntu angangenelela noma nini ukuze ayilungise. Imizila yokulungisa nemizila yoqobo igcinwa ngokuhambisana nezitembu zesikhathi eziqondaniswe ngokuqinile; ukulungisa ngakunye komuntu kuba isifundo esiqondisiwe sekhwalithi ephezulu.
Umjikelezo “wokuthumela—ukuqoqa—ukuhlola”: Ukucabanga kwemodeli, ukwenziwa irobhothi, ukubuyiselwa kwedatha nokuphindaphindwa kwemodeli kuhlanganiswa kube umjikelezo ojwayelekile, kugcwalise igebe embonini futhi kwenze imiphumela ikopishwe ngobuningi kwamanye amarobhothi namamodeli.

Landela i-WeChat bese uphendula “开源” ukuze ujoyine iqembu lezingxoxo zamaphrojekthi e-AI avulekile
Ukwakhiwa kokucabanga ngamaphayiphi ahambisanayo: Ngenxa yomehluko ophindwe izikhathi ezingu-30—50 phakathi kokubambezeleka kokucabanga kwamamodeli e-VLA/WAM nomjikelezo wokulawula irobhothi, i-Open-RAIL ihlukanisa ukubuka, ukucabanga nokulawula kube ngamaphayiphi amathathu asebenza ngokuhambisana ohlangothini lwerobhothi: intambo yokubuka iqhubeka nokuqoqa isimo sendawo namalunga, intambo yokucabanga iqhubeka ibala amabhulokhi ezenzo, kanti intambo yokulawula iqhubeka ithumela imiyalo ku-controller esezingeni eliphansi. Lezi zintathu azilindelani, ngaleyo ndlela kususwe ukumiswa kokunyakaza okubangelwa ukulinda ngokulandelana.
Ukushelelisa izenzo okusezingeni ezimbili ngesikhathi sangempela: Okukhishwa yimodeli ngamaphuzu ezenzo ahlukene; ukuwenza ngqo kungadala ukusheshisa nokunciphisa isivinini okusheshayo. I-Open-RAIL isebenzisa isu lokushelelisa elisezingeni ezimbili, ngaphakathi kwebhulokhi naphakathi kwamabhulokhi, elibalwa ngesikhathi sangempela ngaphandle kokungaxhunyiwe. Akudingeki kushintshwe izisindo zemodeli noma kwenziwe ukuqeqeshwa okwengeziwe; ukuphambuka okujwayelekile kokushesha kwamalunga kwehla kusuka ku-10+ rad/s² kuye ku-0.1 rad/s², kunciphise umthelela osheshayo kuhadiwe.
Ukuhlukaniswa kwe-Server-Client phakathi kwe-end, edge ne-cloud: Uhlelo lubeka ukucabanga ku-Server, kanti umzimba werobhothi (Client) wenza kuphela izenzo futhi ubika izimo. I-Server ingafakwa emzimbeni werobhothi, emandleni okubala e-edge noma ku-cloud. Ikhodi iyafana ngokuphelele; ngokushintsha ikheli lokuxhumana kuphela, kungashintshwa phakathi kwezinhlobo ezintathu zokufakwa ngaphandle kokushintsha ikhodi, ukuze ubukhulu bemodeli bungasakhawulwa amandla okubala e-embedded erobhothini.
Isendlalelo sokukhipha ihadiwe nesivumelwano esifanayo sokuxhuma: I-Open-RAIL ikhipha isendlalelo esilula sokukhipha ihadiwe (HAL), esihlanganisa ngendlela ejwayelekile imiyalo yokulawula, ukufundwa kwezimo nokwenziwa kwezenzo. Irobhothi elisha lidinga kuphela ukusebenzisa i-adapter esezingeni eliphansi bese libhaliswa; ingqondo esezingeni eliphezulu isetshenziswa kabusha ngokuphelele. Ohlangothini lwemodeli, kuchazwa isixhumi esifanayo sokucabanga; imodeli entsha ingaxhunywa ngokusebenzisa lesi sixhumi esijwayelekile, ngokuvamile idinga imigqa yekhodi engu-50—100 kuphela, kuncishiswe izindleko zokuxhumanisa imizimba namamodeli ahlukahlukene.

Qinisekisa indawo: Lungisa indawo ye-Linux ene-Python ≥ 3.10. Iseva idinga i-NVIDIA GPU + CUDA + PyTorch; i-client ifaka i-SDK yomkhiqizi noma i-ROS ngokwezidingo zerobhothi. Lezi zinhlangothi ezimbili zixhumana nge-ZMQ kunethiwekhi yendawo efanayo.
Thola ikhodi: Clone-a i-repository kusuka ku-Gitee, GitHub noma emphakathini i-Huanxin (git clone https://gitee.com/cmcc-tao/open-rail.git).
Faka izidingo: Ngemva kokudala indawo ye-conda, sebenzisa pip install -e .; kuzobhaliswa ngokuzenzakalelayo izindawo ezimbili zokungena zomugqa womyalo, i-vla-server ne-vla-web-client.
Qalisa iseva: Qalisa python run_server.py --model_type <uhlobo lwemodeli> --model_path <indlela yezisindo> ukuze ulayishe imodeli ekhethiwe ye-VLA/WAM futhi uveze indawo yokucabanga nge-ZMQ.
Qalisa i-Web client: Qalisa python run_web_client.py (ngokuzenzakalela ifunda conf/default_conf.yaml futhi ilalele ichweba 9000); sebenzisa --conf ukuze ucacise ukucushwa ngokwezifiso.
Hlola uxhumano: Vula http://localhost:9000 esipheqululini. Uma ubona i-Web interface, isithombe sekhamera nesimo serobhothi silungile, ukuqaliswa kuphumelele; khetha umsebenzi ukuze ushayele irobhothi.
Qala nge-simulation (ngaphandle kwerobhothi langempela): Ku-conf/robots_conf.py, khomba i-dataset_path ye-Mock backend kudathasethi ye-LeRobot, bese usebenzisa --model_type mock ukuhlola lonke uchungechunge.
Ukwenza ngerobhothi langempela: Shintsha robots.type ku-conf/*.yaml ibe yi-adapter oyifunayo (agibot_g1 / ti5_t170c / navi_wa2), bese ulungisa izihloko zekhamera ne-action_layout ngokwemibhalo; qalisa ngemiyalo emibili efanayo.
Ukushintsha phakathi kwe-end, edge ne-cloud: Akudingeki kushintshwe ikhodi; shintsha kuphela inkambu yekheli lenethiwekhi leseva ekucushweni ukuze ukucabanga kusebenze emzimbeni werobhothi, ku-edge noma ku-cloud.
Thola idatha futhi uhlole: Ukuqalisa ngakunye kukhiqiza ngokuzenzakalela idatha yokuqeqesha ye-Parquet esesitayeleni se-LeRobot namalogi okuhlola e-eval/ (JSON/CSV), alungele ukubuyiselwa ekuqeqesheni; imodi exubile “yokucabanga + ukulungisa ngokulawulwa kude” ihlelelwe ukukhishwa ngo-Okthoba.

Ukunyakaza kuzinzile futhi kubushelelezi: Iphayiphi elingavumelaniswa lezintambo ezintathu kanye nokushelelisa okusezingeni ezimbili kususa ngokwezinga lohlaka umehluko ophindwe izikhathi ezingu-30—50 phakathi kwemvamisa yokucabanga neyokulawula. Ukuphambuka okujwayelekile kokushesha kwamalunga kwehla kusuka ku-10+ rad/s² kuye ku-0.1 rad/s², ngaphandle kokushintsha imodeli noma ukukhawulela ukwakheka kwemodeli.
Ukuxhuma okukodwa kusetshenziswa yonke indawo: Isendlalelo esilula sokukhipha ihadiwe nesivumelwano esifanayo sokuxhuma kwehlisa ukufakwa kwerobhothi elisha kusuka emavikini kuya emahoreni, kanti imodeli entsha idinga imigqa yekhodi engu-50—100 kuphela; ingqondo esezingeni eliphezulu isetshenziswa futhi ngokuphelele.
Amandla okubala awasesona isithiyo: Ngesakhiwo esihlukanisa i-server ne-client, ukucabanga kubekwa ku-Server. Ukushintsha ikheli lokuxhumana elilodwa kuvumela ukushintsha phakathi kwe-end, edge ne-cloud ngaphandle kokushintsha ikhodi; amadivayisi e-embedded nawo angasebenzisa amamodeli amakhulu.
Ukusebenza kusho ukuqoqwa nokuhlolwa ngokuzenzakalela: Ukuqoqwa kwedatha namalogi okuhlola kwakhelwe kukho konke ukucabanga, kukhiqize ngokuzenzakalela idatha esesitayeleni se-LeRobot nezinkomba zesikhathi eziphindaphindekayo, kuhlanganise umjikelezo “wokuthumela—ukuqoqa—ukuhlola”.
Ukulungisa komuntu kuba isampula lokufundisa: Kumodi exubile yokulawulwa kude, umuntu angamisa, angenelele, alungise bese eqhubeka noma nini. Imizila yokulungisa nemizila yoqobo igcinwa ngokuhambisana nezitembu zesikhathi eziqondaniswe ngokuqinile, ibe idatha yokuqeqesha eqondisiwe yekhwalithi ephezulu.
Iqinisekiswe ezimweni zangempela: Irobhothi elihambayo i-Xingchu selisebenza njalo epaki lase-Hangzhou, likhiqiza izinkomishi zekhofi ezingaphezu kwekhulu ngosuku, nenqubo ephelele yenkomishi ngayinye ethatha imizuzwana engu-90. Emisebenzini emide, imodeli yenza zonke izinqumo ngokuzenzakalela, okufakazela ukubaluleka kwalesi sisekelo.

Iwebhusayithi yephrojekthi: https://cmcc-tao.github.io/open-rail/

Amarobhothi esevisi yezentengiso: Njenge-“Xingchu ehambayo” esivele isetshenziswa, enza imisebenzi emide efana nokulungisa ikhofi nokushisisa ukudla ku-microwave. Ikhiqiza amakhulu ezinkomishi zekhofi ngosuku, inqubo ephelele yenkomishi ngayinye ithatha imizuzwana engu-90, futhi imiyalo eminingi iqhutshwa ngokuhambisana kuyilapho imodeli yenza zonke izinqumo ngokuzenzakalela.
Ukuhlola nokugada kwezimboni: Eziteshini zikagesi, emagumbini amaseva nasezindaweni zezimboni, yenza imisebenzi ephindaphindwayo efana nokuhlola imishini, ukufunda amamitha nokubika izexwayiso. Ukufakwa ku-end, edge ne-cloud kuvumela ukucabanga kwamamodeli amakhulu ukuthi kwenziwe ku-edge noma ku-cloud ngaphandle kokudinga amandla amakhulu okubala emzimbeni werobhothi.
Usizo kubantu abadala nezokwelapha: Ezimweni ezifana nokuphakela itiye namanzi, ukuletha izinto nokusiza ekululameni, ukwenziwa okushelelayo kuqinisekisa ukuphepha nobumnene lapho irobhothi lisebenzisana nabantu eduze.
Ucwaningo nokufundisa: Amanyuvesi namaqembu aqalayo angaqala nge-Mock simulation backend ngaphandle kwehadiwe ukuze ahlole ngokushesha ukusebenza kwamamodeli e-VLA. Ukuxhuma imodeli entsha kudinga imigqa yekhodi engu-50—100 kuphela, ukuze abacwaningi bagxile ekusunguleni amamodeli ngaphandle kokwakha amapayipi obunjiniyela.