U-Robbyant Ukhipha I-LingBot-VLA 2.0: Imodeli Ye-Open-Source 6B Vision-Language-Action (VLA) ye-Cross-Embodiment Robot Manipulation

0 0 4 minutes read

U-Robbyant Ukhipha I-LingBot-VLA 2.0: Imodeli Ye-Open-Source 6B Vision-Language-Action (VLA) ye-Cross-Embodiment Robot Manipulation

URobbyant we-Ant Group usephumile I-LingBot-VLA 2.0imodeli yesisekelo ye-Vision-Language-Action (VLA) yamarobhothi. Ukukhishwa kufaka umbiko wezobuchwepheshe, i-codebase ye-Apache-2.0, kanye nendawo yokuhlola engu-6B. Ithimba locwaningo liqondise igebe elaziwa kakhulu: Amamodeli e-VLA avame ukusebenza kumalebhu kodwa akhubeke ekusetshenzisweni. I-LingBot-VLA 2.0 ithuthukisa inguqulo yangaphambili ngezimbazo ezintathu ezisebenzayo. Lokhu ukwenza okuvamile, indawo yesenzo enwetshiwe, kanye nokumodela okubikezelwayo kokuguquguqukayo.

Iyini i-LingBot-VLA 2.0?

I-LingBot-VLA 2.0 iyinqubomgomo yamarobhothi ajwayelekile eyakhelwe phezu komgogodla wolimi lombono. Iguqula izithombe zekhamera kanye nomyalelo wolimi kube izenzo zamarobhothi. Imodeli esesidlangalaleni yi-lingbot-vla-v2-6b, indawo yokuhlola engu-6B 'yokujula komdabu'. Isebenzisa Qwen3-VL-4B-Yala njengomgogodla we-VLM. Amamodeli amabili othisha, Ukujula kwe-LingBot futhi Ividiyo ye-DINOyengamela ukuqeqeshwa ngokusebenzisa i-distillation.

Ikholi eyodwa yokucabanga ithatha cishe 130 ms ku-NVIDIA GeForce RTX 4090D. Leso silinganiso sisebenzisa izinyathelo eziyi-10 zokukhipha umsindo. Ingcweti yesenzo isebenzisa idizayini ye-Mixture-of-Experts (MoE) yokukala.

Ipayipi ledatha: amahora angama-60,000 ekucushweni okungu-20

Ukujwayela kuqala ngedatha. Ithimba locwaningo lilinganisa cishe Amahora angu-60,000 yedatha yokuqeqeshwa kwangaphambili. Lokhu kuhlanganisa Amahora angu-50,000 we-robot trajectories kanye Amahora angu-10,000 wamavidiyo abantu abazicabangela wona. Idatha yerobhothi iyadlula 20 ukucushwa kwamarobhothikusuka kumarigi engalo eyodwa kuya kuma-humanoid agcwele. Ichibi elingahluziwe likhulu: cishe amahora angama-90,000 amarobhothi namahora angama-20,000 wokuzicabangela. Ipayipi eliklanywe kabusha lihlunga amasampula anomsindo kuze kufike kusethi yekhwalithi ephezulu.

Ukuhlunga kucacile futhi kuyalinganiseka. Ithimba locwaningo lihlanganisa i-jerk ye-oda lesithathu kanye nesivinini nokusheshisa izikolo ze-Z nge-embodiment ngayinye. Iziqephu ezinokushelela okungavamile noma amasiginali amile angaphezu kuka-95% ziyehliswa. Amavidiyo ahlolwa ngokumelene nezimo eziphinde zadlalwa kusetshenziswa i-URDF yerobhothi ngalinye. Izichasiselo zisusa ukufiphala, ukuvaleka, ozimele abawisiwe, nokungaboni ngaso linye kokubukwa okuningi. Iziqeshana ze-Egocentric zidlula isihlungi se-VLM, bese kuba i-SLAM ye-egocentric kanye nokwakhiwa kabusha kwe-MANO ngesandla.

Isichasiselo sizenzakalela ngemodeli yolimi lombono. Qwen3.6-27B amasegimenti ividiyo ngayinye ibe imisebenzi engaphansi ehambisanayo yesikhashana. Umsebenzi omncane ngamunye uthola isenzo se-athomu kusuka kusilulumagama esivaliwe sezigaba eziyi-18. Lolo lwamagama luphethe izenzo zakudala ezingu-15 kanye nezokuthutha, ukungenzi lutho, nokunye. Ku-corpus, ukunyakaza nokuhamba kubusa ngamaza.

Ukumelwa kwesenzo esihlanganisiwe

Amarobhothi ahlukene aveza amalunga ahlukene, ngakho i-LingBot-VLA 2.0 iyawahlanganisa. Isebenzisa a Ivekhtha ye-canonical engu-55-dimensional kuzo zombili izifunda nezenzo. Isakhiwo silungisiwe kukho konke ukufana kudathasethi.

Isakhi	Ubukhulu
Isikhundla sokuhlangana kwengalo	14
Ukuma komphumela	14
Isikhundla se-Gripper	2
Ukuma okuhlangene kwesandla	12
Ukuma okhalweni	4
Indawo yekhanda	2
Isignali yokuhamba	3
Kugciniwe	4

Isikhundla ngasinye somphumela wengalo sisebenzisa izixhumanisi ze-XYZ kanye ne-quaternion yokuzungezisa, okunikeza izilinganiso ezingu-7 ngengalo ngayinye. Amarobhothi angenayo ingxenye yomzimba avele aphequlule izilinganiso ezihambisanayo. Lokhu kuvumela imodeli eyodwa ukuthi ilawule izingalo, izandla, ama-grippers, okhalo, amakhanda, nezisekelo zamaselula.

Isazi sesenzo se-MoE

Ingcweti yesenzo ishintsha inethiwekhi yayo yokudlulisela phambili nge-srse I-MoE izendlalelo. Isendlalelo ngasinye se-MoE sigcina uchwepheshe oyedwa okwabelwana ngawo kanye nochwepheshe abaningana bomzila. Ochwepheshe bemizila ephezulu ye-K kuphela abasebenza ngethokheni ngayinye, ngakho-ke ukubala okusebenzayo kuhlala kunomkhawulo. Uchwepheshe ngamunye u-a I-SwiGLU MLP enobubanzi obuncane obuphakathi.

Umzila ulandela i-sigmoid-based, i-axiliary-loss-free isu eliphefumulelwe i-DeepSeek-V3. Ukuchema kochwepheshe ngamunye kulungisa ukungalingani komthwalo ngaphandle kokwengeza ukulahlekelwa kokulinganisela komthwalo. Ukuzethemba komzila kusavela kuzikolo zangempela zemodeli, zokuhambisana ezingachemile. Ngaphansi kwamapharamitha asebenzayo afanisiwe, imodeli ye-MoE ifinyelela ekulahlekelweni okuphansi kokuqeqeshwa kunesisekelo esiminyene. Iphinde ifinyelele iphutha eliphansi lesenzo sokuqinisekisa emisebenzini ye-GM-100.

I-dual-query distillation yama-predictive dynamics

Ukwenza kwangempela kudinga ukulindela, hhayi nje ukusabela kuhlaka lwamanje. I-LingBot-VLA 2.0 yengeza imibuzo emibili efundekayo kumathokheni abonakalayo nawemibhalo. I-Qt iqondise ekuhlolweni kwamanje, futhi i-Qt+T iqondise ekubonweni okuzayo. Umkhathizwe T ulingana nosayizi wesiqephu sesenzo.

Othisha ababili abaqondisa le mibuzo. Ukujula kwe-LingBot inikeza izinkomba zejiyomethri ezicacile ngokubikezela ukujula. Ividiyo ye-DINO inikeza okubalulekile kwe-semantic okusekelwe okwesikhashana. I-DINO-Video yakhelwe phezu kwe DINOv3 umgogodla onokunaka kwesikhashana okuhlakaniphile kwe-block-wise kanye ne-3D-RoPE. Iqeqeshelwa iziqeshana zevidiyo eziyi-5M ezihlanganisa i-inthanethi, i-egocentric, nedatha yerobhothi. Ekuhlolweni kwe-LARYBench, i-DINO-Video ihamba phambili ngamamethrikhi amathathu kwamane.

Imiphumela yebhentshimakhi

U-Robbyant uhlola imodeli kusilungiselelo se-generalist ku- GM-100 (Great March 100) ibhentshimakhi ye-bimanual. Umgomo owodwa uqeqeshwa ngokuhlanganyela emisebenzini eyisishiyagalolunye emvelweni ngamunye. Imiphumela ibikwa njengenqubekelaphambili yamaphuzu / izinga lempumelelo.

Inkundla	GR00T N1.7	π0.5	I-LingBot-VLA-1.0	I-LingBot-VLA-2.0
I-AgileX Cobot Magic	36.3 / 17.8	59.1 / 32.2	58.2 / 30.0	66.2 / 34.4
I-Galaxea R1Pro	16.4 / 5.6	27.4 / 8.9	32.7 / 15.6	34.6 / 15.6

Ekukhohlisweni kweselula okude komkhathizwe, imodeli ihlolwa ngaphansi kwezilungiselelo ezimbili. Isizinda esingaphakathi (ID) sisebenzisa ukusabalalisa kokuqeqeshwa, kuyilapho i-OOD perturbs ima kanye nezinto.

Ukufanekisa	Umsebenzi	Ukusetha	I-LingBot-VLA-2.0	π0.5
I-Astribot S1	Ukuhlunga kwesiqandisi	Isizinda	77.1 / 60.0	65.3 / 46.7
I-Astribot S1	Ukuhlunga kwesiqandisi	OOD	37.0 / 13.3	30.3 / 6.7
I-Cobot Magic-ARX X5	Ukuhlanza isitofu	Isizinda	84.3 / 66.7	79.9 / 60.0
I-Cobot Magic-ARX X5	Ukuhlanza isitofu	OOD	67.5 / 40.0	62.5 / 33.3

Izinzuzo zinkulu emisebenzini edinga ukugxiliswa kwento okunembile. Ku-Agilex Retrieve keychain, impumelelo isuka ku-60.0 iye ku-100.0 iqhathaniswa nenguqulo 1.0. Eminye imisebenzi isakhombisa igebe phakathi kwenqubekelaphambili nempumelelo. Lelo gebe likhomba ukwehluleka ekubekweni okunembile kokugcina noma isinyathelo sokukhululwa.

Ukuqalisa

Imikhumbi yenqolobane ifaka, ilande, futhi ithumele imibhalo. Isibonelo esingezansi silanda izisindo ezikhishwe.

# Environment: Python 3.12, PyTorch 2.8.0, flash-attn 2.8.3
python3 scripts/download_hf_model.py --repo_id robbyant/lingbot-vla-v2-6b --local_dir lingbot-vla

Ukuthunyelwa kwerobhothi langempela kusebenzisa iseva yenqubomgomo ngokuqonda okuhlanganisiwe.

export QWEN3VL_PATH=path_to_Qwen3-VL-4B-Instruct
python -m deploy.lingbot_vla_v2_policy 
  --model_path path_to_posttraining_ckpt 
  --use_compile 
  --use_length 25 
  --port port

Ukusetshenziswa kwangemuva kokuqeqeshwa LeRobot i-v2.1 noma i-v3.0 idathasethi. Isibonelo esinikeziwe sishuna kahle I-RoboTwin 2.0 imisebenzi engama-50. Umzila ungasebenzisa ukulahlekelwa kokusiza okuhlakaniphile okulandelanayo no-z-loss, noma ukusetha okungalahleki. I-config iphinde iveze i-Muon optimizer, ne-AdamW njengokuzenzakalelayo.

Sebenzisa amacala anezibonelo

Amamephu esikhala esenzo anwetshiwe kuzimo zokusebenzisa ezibambekayo.

Ukuguqulwa kweselula kwekhishi: I-Astribot S1 ihlunga izithelo neziphuzo zibe esiqandisini. Lokhu kudinga ukunyakaza kwesisekelo, ukuvulwa komnyango, nokubekwa kwento ndawonye.
Ukuhlanza ubuso: I-Cobot Magic-ARX X5 isula igwebu esitofini ngesipontshi. Lokhu kuhlanganisa ukubamba, ukusula, nokubeka kabusha ithuluzi.
Ukupakisha nokuhlelwa kwe-Bimanual: Imisebenzi ye-GM-100 ihlanganisa ukupakisha amaqanda, ukupakisha amathuluzi, nokuhlelwa kwamabhulokhi.
Ukulawulwa kwesandla okune-Dexterous: I-Unitree G1, i-Fourier GR-2, ne-AgiBot A2 zisebenzisa izandla eziyi-12-DoF, hhayi ama-gripper.

I-Interactive Dynamic Explainer

'+v+'

';} umsebenzi renderEmb(){ var e=EMB[+sel.value]eya=e[2]umzimba=e[4]>0, isEgo=e[1]==='Egocentric'; var act={arm:true,eef:!isEgo,grip:(ee.indexOf('Gripper')>-1),isandla:(ee.indexOf('Hand')>-1), okhalweni:umzimba,ikhanda:umzimba,isixuku:umzimba,res:amanga}; uma(isEgo){act.grip=false;act.hand=true;act.waist=false;act.head=false;act.mob=false;} $$('#strip .grp').forEach(function(g){g.classList.toggle('act',!!act)[g.dataset.k]);}); var eeTxt=ee==='Isandla/Isandla'?'Isandla/Isandla':ee; $('#embstats').innerHTML=st('Type',e[1])+st('End-effector',eeTxt)+st('Arm DoF',e[3])+ st('Body DoF', isib[4])+st('Isamba se-DoF', isib[5],1)+st('Imvamisa yenqubomgomo',e[6]+'Hz'); } sel.onchange=renderEmb; sel.value=11; renderEmb(); /* ===== ABASEBENZISI ABABANE ===== */ amabha omsebenzi(host,rows,hotIdx){host.innerHTML=''; rows.forEach(function(r,i){ var div=document.createElement('div');div.className=”row”+(i===hotIdx?' hi':''); div.innerHTML='

'+r[0]+'

'+r[1].kulungisiwe(1)+'

'; host.appendChild(div); var f=div.querySelector('.gcwalisa'); requestAnimationFrame(umsebenzi(){f.style.width=Math.max(1.5,r[1])+'%';}); }); } function segBtns(host,items,active,cb){host.innerHTML=''; items.forEach(function(it,i){ var b=document.createElement('button');b.className=”chip”+(i===active?' on':'');b.textContent=it; b.onclick=function(){cb(i);};host.appendChild(b); }); } /* ===== PANEL 2 : ABLATION ===== */ var TASKS=[‘Barcode’,’Scoop Rice’,’Ketchup’,’Microwave Bowl’,’Avg’]; var ABL={ 'Ithagethi yesenzo':{okungcono kakhulu:1,inothi:'Izenzo ezihlanganyelwe ziguqula ukuhlehla komhlaba kube ukuhlehla kokunyakaza kwendawo. Lokho kwehlisa ukuhluka futhi kuphakamisa impumelelo emaphakathi ukusuka ku-33.7 kuye ku-55.0.', amasethi:[[‘Absolute (abs)’,[13.3,22.7,55.0,43.8,33.7]],[‘Relative (rel)’,[58.7,42.7,41.7,76.8,55.0]]]}, 'Isikhala sesenzo':{okungcono kakhulu:1,inothi:'EEF kanye nezenzo ezihlanganyelwe zilingana ngokwesilinganiso (56.0 vs 55.0). Isikhala esingcono kakhulu sincike ku-task physics, hhayi izibalo zokusabalalisa zodwa.', setha:[[‘EEF’,[24.0,60.0,81.7,58.3,56.0]],[‘Joint’,[58.7,42.7,41.7,76.8,55.0]]]}, 'Ukujwayela':{okungcono kakhulu:2,inothi:'I-MeanStd igcina ububanzi obuguquguqukayo obukhulu kunawo wonke futhi ilondoloza kangcono ukunyakaza okunomsila omude, okufinyelela impumelelo eyisilinganiso engu-55.0.', amasethi:[[‘MinMax’,[48.0,26.7,61.7,53.6,47.5]],[‘Q01-Q99’,[42.7,44.7,63.3,39.1,47.4]],[‘MeanStd’,[58.7,42.7,41.7,76.8,55.0]]]}, 'Ukulahlekelwa':{okungcono kakhulu:1,inothi:'L2 ilingana nesifunda esiminyana kakhulu sezinto eziqondiwe ezincane. Ithuthukisa impumelelo emaphakathi ngaphezu kwe-L1 isuka ku-46.4 iye ku-55.0.', amasethi:[[‘L1’,[45.0,28.7,61.7,50.0,46.4]],[‘L2’,[58.7,42.7,41.7,76.8,55.0]]} }; var ablKeys=Object.keys(ABL),curStudy=ablKeys[0],curSet=1; umsebenzi renderAbl(){ var s=ABL[curStudy]; segBtns($('#ablStudy'),ablKeys,ablKeys.indexOf(curStudy),function(i){curStudy=ablKeys[i];curSet=ABL[curStudy].okungcono kakhulu;renderAbl();}); $('#ablSetLab').textContent=”Isilungiselelo · “+curStudy; segBtns($('#ablSet'),s.sets.map(function(x){return x[0];}),curSet,function(i){curSet=i;renderAbl();}); var vals=s.sets[curSet][1]; amabha($('#ablBars'),TASKS.map(function(t,i){return [t,vals[i]];}),4); var selected=(curSet===s.best)?' Lokhu ukucushwa kwemikhumbi ye-LingBot-VLA 2.0 nge.':'; $('#ablNote').innerHTML=s.note+khethiwe; shintsha usayizi (); } renderAbl(); /* ===== IPHEPHA LESI-3 : IBENCHMARK ===== */ var MODELS=[‘GR00T N1.7′,’π0.5′,’LingBot-VLA-1.0′,’LingBot-VLA-2.0’]; var BM={ 'AgileX Cobot Magic':{prog:[36.3,59.1,58.2,66.2]succ:[17.8,32.2,30.0,34.4]}, 'Galaxea R1Pro':{prog:[16.4,27.4,32.7,34.6]succ:[5.6,8.9,15.6,15.6]}}; var bmKeys=Object.keys(BM),curPlat=bmKeys[0],curMet=”prog”; umsebenzi renderBm(){ segBtns($('#bmPlat'),bmKeys,bmKeys.indexOf(curPlat),function(i){curPlat=bmKeys[i];renderBm();}); segBtns($('#bmMet'),[‘Progress score’,’Success rate’],curMet==='prog'?0:1,umsebenzi(i){curMet=i?'succ':'prog';renderBm();}); $('#bmCap').textContent=(curMet==='prog'?'Imiphumela yentuthuko':'Izinga lempumelelo')+' (%)'; Val = BM[curPlat][curMet]; amabha($('#bmBars'),MODELS.map(function(m,i){return [m,vals[i]];}),3); shintsha usayizi (); } renderBm(); /* ===== RESIZE NGOKUZENZAKALELAYO ===== */ shintsha usayizi womsebenzi(){ var h=R.offsetHeight+40; zama{parent.postMessage({type:'lbv2-resize',height:h},'*');}catch(e){} } window.addEventListener('load', resize); uma(window.ResizeObserver){new ResizeObserver(resize).observe(R);} setTimeout(shintsha usayizi,300); })();

Source link

nimda 3 hours ago

0 0 4 minutes read