Chii chaitika
Vatsvaguri vanokurudzira PointRL, iyo inogoneka yekusimbisa-yekudzidza sisitimu yekudzidzisa mamodheru emutauro wechiratidzo kuona nzvimbo dzakanangwa dzine mapoinzi anoronga. Pepanhau rinoshuma kuti PointRL yakasimudza Qwen3.5-4B's PointArena kurongeka kubva pa56.11% kusvika 65.58%, nezvakawaniswa zvakashumwawo pamabhenji matatu ekunze.
Iro bepa, rakatumirwa kuarXiv musi waNyamavhuvhu 26, 2026, rinogadzirisa dambudziko rakananga mumhando dzemutauro wechiratidzo: kushandura rairo kuita imwe kana anopfuura poindi inoronga iyo nyika pane zvakatarisirwa zvinotariswa. Vanyori vanoti mapoinzi coordinates ari kuwedzera kushandiswa se compact, inogoneka interface ye graphical-mushandisi-interface yekudyidzana, marobhoti manipulation, uye mamwe anodyidzana ekuona masisitimu. Dambudziko nderekuti dunhu rakanangwa rinogona kuve neakawanda anorongedzerwa, saka imwe chete yakagadziriswa poindi haisiriyo yega mhinduro chaiyo.
PointRL inobata zvirevo zvekuona zviripo sehumbowo hwekutarisa kufanotaura. Zvinoenderana neiyo abstract, inoshandura mabhokisi anosungirirwa, segmentation masks, uye muenzaniso mavara kuita anongedza mirairo. Inochengeta ruzivo rwenzvimbo dzakatariswa, izvo zviitiko ndezvezvakakumbirwa seti, uye zvinomanikidza pane iyo seti pachayo kunze kweiyo modhi yekukurumidza. A deterministic checker anobva ashandisa humbowo hwakavanzika hwukohwe zvinobuda zvemodhiyo pane kuburitsa mhinduro inomiririra zvakananga murairo.
Mubairo wakashumwa unoongorora zvinhu zvakati wandei zvekufanotaura: zvingave zvichipatsanurwa, kana mapoinzi ari echokwadi, angave mamiriro akakumbirwa akafukidzwa, kana nhamba yekufanotaura ichienderana nekuraira, uye kuti kufanotaura kusingaite here kana kushaikwa. Izvi zvinoitirwa kugadzirisa zvese zviri zviviri-chinangwa uye akawanda-mienzaniso zvikumbiro. Kunobva hakupi bepa rakazara algorithmic tsananguro, hurongwa hwekudzidziswa, modhi yekutarisa, dhatabheti kuumbwa, kana ablation mhinduro, saka izvo zvinyorwa hazvigone kuongororwa kubva pane zvakapihwa.
PaPointArena, vanyori vanorondedzera kuti PointRL yakavandudza kukwana kwese kweQwen3.5-4B kubva ku56.11% kusvika ku65.58%. Iyo abstract inotaurawo kuti ongororo paRoboSpatial, BLINK, uye Ref-Adv yakaratidza zvakawanikwa vachishandisa musana wakafanana, uyo unoturikirwa nevanyori sehumbowo hwekuti inovimbika yedanho-nhanho mhinduro inogona kuvandudza nzvimbo munzvimbo idzi. Kwakabva hakupi zvibodzwa zvemunhu, zviverengero zvemienzaniso, fungidziro yekusavimbika, kuenzanisa nedzimwe nzira dzekudzidzisa, kana ruzivo rwekuti ongororo dzekunze dzakaitwa nevaongorori vakazvimiririra.
Zvakatorwa pamwechete, tsananguro yakapihwa inoburitsa PointRL senzira yekudzidzisa iyo mhinduro inotorwa kubva kune humbowo hwetsanangudzo hunochengetwa kunze kwekukurumidza. Kutevedzana kwayo kwakashumwa kwekuongorora kunotanga nePointArena kuenzanisa kweQwen3.5-4B uyezve inosanganisira zvakafanana-musana mhinduro paRoboSpatial, BLINK, uye Ref-Adv. Iyo account iripo inotaridza mhando dzezvirevo zvakashandiswa, zvikamu zvekufungidzira zvakapihwa, uye shanduko yemusoro wenyaya, asi haipi iyo iri pasi pealgorithm, chirongwa chekudzidziswa, chekutarisa, dhatabheti kuumbwa, ablations, ega ega zvibodzwa zvekunze, sampuli kuverenga, fungidziro yekusavimbika, kana ruzivo rwekuzvimirira. Izvo zvakasiiwa zvinotsanangura muganhu wezvinogona kugumiswa kubva pane zvakapihwa.
Nei zvichikosha
Point-level grounding ndeye compact interface yemasisitimu anofanirwa kubatanidza mutauro nenzvimbo mumifananidzo, kusanganisira graphical-mushandisi-interface kupindirana uye marobhoti manipulation. A deterministic verifier inogona kuita kuti mhinduro yekudzidzisa inyatsojeka pane kubata chero coordination imwe chete semhinduro chaiyo.
Mapoinzi mapoinzi anogona kuve nyore kune yakadzika system kuti iite pane rondedzero refu yemavara yenzvimbo yechinhu. Muzvigadziro zvakadanwa nebepa, modhi ingangoda kuratidza kuti bhatani riri papi, pakamisikidzwa chinhu serobhoti, kana kuti matunhu api anogutsa kuraira kwekuona. Sarudzo yepoindi inovimbika inogona nokudaro kuderedza danho rekushandura pakati pekufunga nemutauro wekuona uye chiito, kunyangwe bepa risingaratidzi sisitimu yakatumirwa kana basa renyika.
Mupiro wepakati webepa mabatirwo edata rezvinyorwa seumboo hunogoneka kwete kungoita sechinyorwa chekunangwa. Mabhokisi ekusungirira uye masks anotsanangura nzvimbo umo mapoinzi akawanda anogona kugamuchirwa, nepo marebhura nemaseti ekumanikidza anobatsira kusiyanisa mhinduro dzakakwana kubva kune chidimbu kana chakapetwa. Iyo dhizaini inogadzirisa kushaya simba kweiyo coordination-based maneja: kupa mubairo imwe yakasarudzwa coordination inogona kuranga imwe kurongeka iyo zvakare yakarurama.
Kuwedzera kwakataurwa paPointArena kwakakura mumashoko akakwana, zvichifambisa chokwadi chakataurwa kubva pa56.11% kuenda pa65.58%. Kana ikabata pasi pekuyedzwa kwakakura, mhedzisiro inogona kubatsira kune vaongorori vanogadzira mhando dzinofanirwa kudzika mutauro munzvimbo chaidzo dzinoonekwa. Izvo zvinonzi zvakawanwa paRoboSpatial, BLINK, uye Ref-Adv zvine basa nekuti vanoti gadziriso iyi inogona kusaganhurirwa kubhenji rinoshandiswa pakudzidzisa kana ongororo yekutanga.
Izvo zvinorehwa zvinoramba zvine zvirevo nekuti sosi imwechete arXiv preprint uye iwo mameseji akapihwa anongori abstract. Mhedzisiro iyi humbowo hwegwara rekutsvagisa, kwete yakazvimiririra simbiso yekuti PointRL inovandudza kuvimbika mune inoshanda interfaces kana marobhoti. Iyo abstract haitauri kutadza kuomarara, kunonoka, mutengo weanotsidzira, kunzwisiswa kugadziriso yemufananidzo, kana maitiro emaitiro kana zvinonangwa zvakavharika, zvisinganzwisisike, zvidiki zvisingaite, kana ndisipo.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Mibairo yakashumwa inobva kune preprint abstract uye haigadziri kuvimbika kwepasirese kana kugadzirira kutumirwa. Zvinokosha zvisingazivikanwe zvinosanganisira mhedzisiro izere yebhenji, kudzidziswa uye kuongororwa mamiriro, kusiyanisa kwenhamba, kuita mumamodheru nemabasa, uye kuti nzira kana mhando dzakadzidziswa dzichaburitswa.
Chekutanga chakakosha inyaya yekuongorora kwebepa rakazara. Vaverengi vanofanira kutsvaga chaiwo PointArena, RoboSpatial, BLINK, uye Ref-Adv protocol; nhamba uye rudzi rwemabasa; mhando dzekutanga; uye kana izvo zvinokurudzira zvakafanana, data, uye zvibodzwa mitemo zvakashandiswa pakuenzanisa. Iwo madonzo ndiwo achaona kana shanduko yechokwadi yakashumwa ichiratidzira kuvandudzwa kwepasi kana kukwana kudiki kune inosimbisa uye chimiro chebhenji.
Kudzokorodza kwakazvimirira kwaizobatsira kuona kana zvakawanikwa zvichifambiswa kumhuri dzemhando, zvinyorwa zvinyorwa, mitauro, madhomeni emifananidzo, uye masitayipi ekuraira. Iwo abstract mazita Qwen3.5-4B semuenzaniso unogamuchira kuvandudzwa kwePointArena, asi haitauri kana mamodheru akakura kana akasiyana akadzidziswa mitauro yekuona anobatsirwa zvakafanana. Izvo zvakare hazviratidze kuti ingani data rekudzidzisa kana komputa PointRL inoda hukama neyakajairwa-tuning.
Kuendesa kunoshanda kunoda kuyedzwa kupfuura static benchmark chokwadi. Poindi inoshanda mukati menzvimbo yakanangwa inogona kunge ichiri isina kuchengetedzeka kurobhoti, isingashandisike munzvimbo ine vanhu vakawanda, kana kusakwana kuita basa rinoda kudzika, nguva, kana kuzivikanwa kwechinhu. Ongororo dzenguva yemberi dzinofanira saka kuongorora kuregera, kusavimbika, kupotsa uye kudzokorora zviitiko, kugovera mashifiti, uye mhedzisiro yemapoinzi asiri iwo mumasisitimu anodyidzana.
Mamiriro ekuburitswa kwepepa ndechimwe chisingazivikanwi. Sosi yacho inozivisa rekodhi rearXiv uye zvinongedzo kune bepa, asi zvinyorwa zvinopihwa hazvitaure kana kodhi, verifier maitirwo, zvirevo zvakashandurwa, kana nzvimbo dzekutarisa dzakadzidziswa dziripo. Kusvikira izvo zvinyorwa uye zvizere zvawanikwa zvaongororwa, PointRL inonzwisiswa zvakanyanya senzira inovimbisa yakashumwa yeinogoneka point-level grounding, ine kukosha kwayo kwepasirese kuchiri kusimbiswa.