Dzokera kuNhau
InnovationAI Understanding muchidimbu

$R^3$ inodzidzisa marobhoti kuti ashandise kufunga kwemutauro wechisikigo panguva yekushandura

arXiv preprint inosuma $R^3$, nzira yemashure-yekudzidzira inoshandisa mahara-fomu yemutauro kufunga kutungamira marobhoti manipulation policy. Vanyori vanoshuma zvakawana pamabhenji maviri anodzorwa, uku vachisiya kuita kwepasirese uye saizi yezviwanikwa zvisina kutaurwa.

5 min readRead the primary source
Primary-source image accompanying $R^3$ trains robots to use natural-language reasoning during manipulation
Primary-source documentKwakanyorwa
Muparidzi
arxiv.org
Source link
arxiv.orghttps://arxiv.org/abs/2608.26053
Source type
Gwaro rekutanga - chiziviso chepamutemo, bepa, faira, kana peji rebato rekutanga ratinoverenga zvakananga.
ContextNzwisisa izvi mumasekonzi makumi matanhatu

Tanga pano

Matemu akakosha

Vision-Language Model (VLM)
Iyo multimodal modhi iyo yakabatana inogadzirisa zvinoonekwa uye zvinyorwa zvemashoko.
Kusimbisa Kudzidza
Kudzidziswa nemasaini masaini apo mumiririri anodzidza zviito zvinowedzera kudzoka kwenguva refu.
Test-Nguva Compute
Yekuwedzera inference computation inoshandiswa panguva yekugadzira mhinduro kuvandudza kunaka kana kufunga.
Zviedze iwe pachakoAI Agents Quiz

Chii chaitika

Vatsvaguri vakaunza $R ^ 3$, resipi yemashure-yekudzidzira yakagadzirirwa kugadzira mhando dzemutauro wechiratidzo chikonzero mumutauro wechisikigo uku ichitungamira yakaderera-level marobhoti manipulation. Iyo nzira inotanga kudzidzisa pane-yakagadzirwa nenyanzvi kutsvaga kutsvaga, yobva yashandisa imwe-nhanho, rubriki-yakavakirwa kusimbaradza kudzidza kune offline chiito data. Vanyori vanoshuma yakavandudzwa kuongorora uye kugadzirwa paMutauro Tafura uye kutevedzera gimanual girosari kurongedza, uye vanoti yakapfuura kuraira-chete kutevedzera-yekudzidza yekutanga pane ese mabhenji.

Kunobva kunyorwa kwearXiv kwakatumirwa muna Aug. 26, 2026, yakanzi "$R ^ 3$: Kudzidzisa Robhoti Kufunga Mutauro Wechisikigo kuburikidza neKusimbisa Kudzidza." Inodzidza kana mamodheru emutauro wechiono achigona kudzidziswa kugadzira kufunga kwemutauro wechisikigo kunotungamira marongero akaderera. Iro bepa rinogadzira dambudziko rakatenderedza mabasa akareba, uko robhoti inofanirwa kuteedzera kufambira mberi, chikonzero chehukama pakati pezvinhu, kufanotaura mhedzisiro uye kupora kubva mukukanganisa.

Resipi yakarongwa ine matanho maviri akataurwa. Chekutanga, vaongorori vepakati-vanodzidzisa modhi yemutauro wechiratidzo pamatanho ekufunga anogadzirwa nenyanzvi, nechinangwa chekutanga maitiro ekufunga anodiwa. Ivo vanobva vavandudza muongorori nedanho rimwe chete, rubric-yakavakirwa kusimbisa kudzidza vachishandisa offline chiito data. Kunobva kunotsanangura izvi senzira yekudzima-iyo-sherufu-yemutauro-mamodheru kuita marobhoti mabatiro, pane senge yemhando nyowani yekuvaka.

Iro bepa rinosiyanisa maitiro aro kubva kune marobhoti ekufunga nzira ayo anonyanya kushandisa maitiro akarongwa seyebetsero yekutarisa. Zvinoenderana neabstract, $R ^ 3$ pachinzvimbo inodzidzisa yemahara-fomu mutauro kufunga kupa nhungamiro panguva yekuyedza. Tsanangudzo yevanyori ndeyekuti kufunga kwemutauro kunogona kushanda seyeyedzo-nguva yekombuta nzira: modhi inogona kushandisa imwe computation pakuparadza basa, kutevedza zvimhingamipinyi kana kutarisira mhedzisiro yeramangwana usati watungamira danho repasi.

Ongororo iyi inoshandisa ma testbed maviri anodzorwa anonzi mune kwakabva: Mutauro Tafura uye akateedzera mavirimanual grocery packing. Vanyori vanoshuma kuti $R ^ 3 $ inonatsiridza kuongorora uye kugadzirwa kwese pamabasa asingaonekwe uye inokunda zvakanyanya kuraira-chete kutevedzera-kudzidza mabhenji pane ese mabhenji. Iyo abstract haipe mhedzisiro yenhamba, tsanangura iyo chaiyo basa rakakamurwa kana kutsanangura iyo yekutanga masisitimu, saka iwo madhairekitori haagone kuongororwa akazvimirira kubva kune yakapihwa kwakapihwa.

Kwakabva mashoko: arxiv.org ↗

Nei zvichikosha

Iro bepa rinotarisa kusimba kunoshanda mukureba-horizon robhoti kutonga: kupedzisa basa kungangoda kuronda kufambira mberi, kunzwisisa hukama hwechinhu, kutarisira mhedzisiro uye kupora kubva mukukanganisa. Chirevo chayo chepakati ndechekuti mutauro wemahara unogona kuita seyekuwedzera-yenguva computation yekutungamira yakaderera-level policy. Kana iyo yakashumwa pateni ikaramba ichipfuura zviedzo zvinodzorwa, nzira yacho inogona kupa nzira yekuvandudza mamodheru aripo emutauro wekuona pasina kudiwa modhi nyowani yekuvaka. Humbowo hunoramba hwakaganhurirwa kumabhenji evanyori uye abstract-level zvichemo.

Kubata marobhoti kwakaoma muchikamu nekuti budiriro haisi nguva dzose mhinduro yedanho rimwe kune rairo. Robhoti inogona kuda kuziva zvakatoitika, izvo zvinhu zvinomanikidza kufamba kunotevera uye izvo zvikanganiso zvingaite kuti zviito zvikundikane. Kwakabva kunopa kufunga kwemutauro sechinhu chinobvira chepakati pakati pekuraira kwepamusoro-soro uye neruzha rwepasi-chikamu chekudzora. Ndiyo nhungamiro yekutsvagisa yakakosha nekuti inonangana nekuronga uye kudzoreredza, kwete kungoita mazwi emirairo.

Chikumbiro ichi chakakoshawo pakuedza kushandisa mamodheru emutauro wekuona. Tsime rinoti resipi inoshanda neasiri-pasherufu maVLM mushure mekuwedzera kudzidziswa, iyo, kana ikagadzirwazve, inogona kuita kuti nzira ive nyore kuyedza mumhuri dzakasiyana dzemhando. Nekudaro, iyo abstract hairatidze mamodheru ekutanga, chiyero chekudzidzira, zvinodiwa zvehardware kana kuti masisitimu anowanikwa aripo kuti ashandiswe neveruzhinji. Kusiiwa ikoko kunodzikamisa mhedziso pamusoro pekuita uye mutengo.

Izvo zvakashumwa zvinowanikwa zvine chekuita nekuferefeta uye kujekeswa kumabasa asingaonekwe, ese ari maviri akakosha kumarobhoti asingakwanisi kurongwa nemaoko kune yega gadziriro yezvinhu. Zvakadaro, humbowo hunobva kune maviri akadzorwa marongero, imwe yacho yakateedzerwa zvakajeka. Kuita pabhenji hakuratidze kuti robhoti richazvibata zvakachengeteka kana kuvimbika mudzimba, matura kana dzimwe nharaunda dzinoshanduka. Iyo sosi haina chirevo chekutumirwa munzvimbo chaiyo-yepasirese.

Mhedzisiro yebepa nekudaro inofanirwa kunzwisiswa sechirevo chesimba kubva kune preprint, kwete sehumbowo hwekuti kufunga kwemutauro wechisikigo kunogadzirisa kunyengedza kwenguva refu. Nzvimbo yakapihwa haitaure kana zvakawanwa zvichibva mukutaura kwemutauro pachawo, nyanzvi inoteedza, maitiro ekusimbisa-kudzidza kana kusanganiswa kwawo. Izvo zvakare haitauri kutadza modes, kusimba kune zvikanganiso zvekuona, kana kuenzanisa nemaitiro ese akakodzera ekutanga. Iyi mibvunzo yakakosha yekuongorora iyo nzira yeruzhinji uye inoshanda kukosha.

Interactive Mechanism

Interactive Mechanism: Iyo Inonyatsoshanda

Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Interactive Concept Check+10 Points
AI Agents Quiz

An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?

Zvekutarisa zvinotevera

Pepa rakazara rinodiwa kuti riongorore kuvandudzwa kwakashumwa, kusanganisira zvibodzwa zvakakwana, nheyo, kuverenga kwebasa, kusiyanisa kwenhamba uye ablations. Zvinokosha zvisingazivikanwe zvinosanganisira kana nzira yacho inotamiswa kubva mukufananidza kuenda kumarobhoti emuviri, ingani computation uye latency kutaura kwemutauro kunowedzera, uye kana kufunga kwakagadzirwa kwakavimbika kana masensa kana zviito zvine ruzha. Basa rekutevera rinofanirwa kuyedza mabasa akakura ekunyengedza, nharaunda dzisingaonekwe uye hardware yemuviri, uye inofanirwa kuenzanisa $R ^ 3 $ nedzimwe nzira dzekufunga dzemarobhoti pasi pemamiriro anoenderana.

Nzvimbo yechokwadi yekusimbisa ndiyo yakazara bepa humbowo hwehuwandu. Vaverengi vanofanira kutsvaga mwero wekubudirira wakakwana, nhamba nekusiyana kwemabasa, fungidziro yekusavimbika, miedzo inodzokororwa uye tsananguro chaiyo ye "kupfuura zvakanyanya." Iyo abstract inopa kutungamira kwemhedzisiro asi hapana manhamba, saka ukuru hwekuvandudza hahuzivikanwe parizvino.

Muedzo wakakosha ndewekuti maitiro akashumwa anofambiswa kubva kune maviri akadzorwa mabhenji kuenda kune yemuviri Hardware. Mazita ekwaanobva akateedzera kurongedza kwegirosari asi haatauri kuti nzira yacho yakaongororwa parobhoti chairo. Ongororo dzekutevera dzinofanirwa kuyera mashandiro pasi peruzha rwe sensor, kusakwana kwechinhu chinowanikwa munzvimbo, zvipingamupinyi zvisingatarisirwe uye kutadza kwechiito, nekuti iwo mamiriro ari pakati pekudzoreredza uye kumanikidza-kutevera kugona iyo nzira inotarisirwa kuvandudza.

Kunyatsoshanda uye kuvimbika kunodawo kuongororwa. Kufunga kwemahara-fomu kungangoda mamwe matanho ekufungidzira panguva yekuyedzwa, zvichigona kuwedzera latency kana kushandisa komputa panguva yekunyengedza. Kunobva hakuyere mutengo uyu, uye haatsanangure kuti sisitimu inobata sei zvisirizvo, zvisingaenderani kana kufunga zvisina basa. Ongororo dzinofanirwa kuzivisa kupera kwebasa pamwe chete nelatency, compute zvinodiwa uye hukama pakati pemhando yekufunga kwekutaura uye mhedzisiro yemuviri.

Pakupedzisira, kuenzanisa kunofanira kuwedzera kupfuura kuraira-chete kudzidza kutevedzera. Iro bepa rinoti nzira yaro inosiyana nemaitiro ekutanga anoshandisa nzira dzakarongwa seyekubatsira tarisiro, asi iyo abstract hairatidze nzira dzakasimba dzekukwikwidza kana chirevo chakafanana kuenzanisa. Basa rinobatsira rekutevera raizopatsanura mipiro yevanyanzvi vanoteedzera, rubric-based reinforcement kudzidza uye yemahara-fomu mutauro, uku uchiyedza zvinhu zvisingaonekwe, kureba kwemabasa horizons uye nharaunda dzisina kugadzirwa kutenderedza nzira.

Related guides & Quizzes

AI AgentsAI Models InotsanangurwaKudzidziswa kweAIRamangwana reAIEdza zvaunoziva - edza yemahara AI quizTarisa kumusoro izwi reAI mune yedu glossaryTevedza iyo AI modhi yekuburitsa tracker
Wakawana izvi zvinobatsira?