Physical AI na buɗaɗɗen tushe: LeRobot na buɗe ƙofar robotics ga kowa

TL;DR. A taron Economist Enterprise a London, Pepijn Kooijmans ya nuna yadda LeRobot, VLA models na buɗaɗɗen tushe da kayan aiki na $200 ke buɗe ƙofar robotics ga kowa.

Mahimman Bayanai

Daga motion planning zuwa imitation learning: sauyin tsari. A gabatarwarsa ta nazari, Open for business: how open source is accelerating physical AI (A buɗe don kasuwanci: yadda buɗaɗɗen tushe ke hanzarta Physical AI), Pepijn Kooijmans ya bayyana tafiyarsa daga injiniyan…
A cikin tsarin VLA mai hanyoyi da yawa. A jawabinsa, Kooijmans ya bayyana cikakken pipeline na neural network na end-to-end da ke tafiyar da policies na Physical AI na buɗaɗɗen tushe na zamani, yana nuna yadda ake fassara bayanan gani masu girma zuwa…
Kwatanta fasaha: automation na gargajiya da Physical AI na buɗaɗɗen tushe. Tushe: Nazarin bayanan gabatarwa, gwaje-gwajen fasaha da nunin kai tsaye a taron Economist Enterprise Physical AI and Robotics Summit (Convene 200 Aldersgate, London, 13 ga Oktoba 2026).
Abin da wannan ke nufi a kowane fanni. Tasowar Physical AI a buɗe wanda al'umma ke tafiyarwa yana sauya lissafin kasuwanci a masana'antu da dama.

Tsawon shekaru sittin, industrial robotics na aiki ne a ƙarƙashin wani tsari mai tsauri: injiniya ɗan adam yana tsara kowace milimita ta motsin inji a fili, ta amfani da harsunan kamfani na musamman, teach pendants da kinematic solvers masu tsauri. Idan abin da ake aiki a kai ya kauce da milimita biyar ko hasken wuri ya canza, robot ɗin yana samun matsala ya tsaya. A taron Economist Enterprise Physical AI and Robotics Summit a London ranar 13 ga Oktoba 2026, Pepijn Kooijmans, shugaban LeRobot a Hugging Face, ya gabatar da wani tsari na dabam da ke rusa wannan gado: “Ba za a ƙara tsara robot na gaba aiki bayan aiki ba. Za su koya.”

Daga motion planning zuwa imitation learning: sauyin tsari #

A gabatarwarsa ta nazari, Open for business: how open source is accelerating physical AI (A buɗe don kasuwanci: yadda buɗaɗɗen tushe ke hanzarta Physical AI), Pepijn Kooijmans ya bayyana tafiyarsa daga injiniyan mechatronics na gargajiya zuwa deep learning. Tsawon shekaru da dama, an koyar da masana robotics su raba sarrafa abubuwa zuwa ƙananan matsaloli daban-daban: object segmentation, kimanta matsayin 6D, sauya coordinate frames, tsara hanya tare da guje wa karo, da sarrafa motar haɗin gwiwa ta PID.

Duk da cewa wannan rarrabuwar nazari tana aiki sosai a masana'antun motoci masu tsauraran fixtures, inda sassa ke isowa cikin jigs na ƙarfe masu daidaito har zuwa micron, tana rushewa a tsofaffin masana'antu, ɗakunan gwaje-gwaje, rumbunan ajiya da gidaje. A duniyar zahiri, abubuwa suna canza siffa, saman abubuwa na haskaka haske ta hanyar da ba a iya hasashe, kuma abokan aiki mutane suna kawo cikas masu motsi.

Babban ra'ayin da Hugging Face ta gabatar shi ne ba za a iya ƙayyade ƙwarewar jiki ta hanyar nazari da dokokin da aka rubuta da hannu ba; dole ne a same ta ta hanyar gogewa. Kamar yadda manyan language models (LLMs) suka koyi nahawu, tunani da ilimin duniya ta hanyar hasashen token na gaba a cikin ɗimbin rubutu, wakilan Physical AI suna samun ƙwarewar motsi ta hanyar koyon rarrabuwar ayyuka bisa ga abin da aka gani da abin da na'urori suka ji.

Wannan sauyi daga tsara kowane motsi zuwa koyarwa ta hanyar nunin ɗan adam shi ne alamar shigar robotics zamanin foundation models.

A cikin tsarin VLA mai hanyoyi da yawa #

A jawabinsa, Kooijmans ya bayyana cikakken pipeline na neural network na end-to-end da ke tafiyar da policies na Physical AI na buɗaɗɗen tushe na zamani, yana nuna yadda ake fassara bayanan gani masu girma zuwa sarrafa abubuwa na zahiri cikin santsi.

Tsarin Vision-Language-Action na end-to-end na Hugging Face LeRobot da ke nuna visual tokens na SigLIP, joint state vector, umarnin harshe, cross-attention na Gemma VLM da action expert na flow matching da ke fitar da ayyukan haɗin gwiwa 30 masu jituwa .class="img-fluid clearfix"

Tsarin yana aiki ne a matsayin sensorimotor pipeline mai hanyoyi da yawa da aka haɗa sosai:

1. Karɓar hotuna daga kyamarori da yawa #

Robot ɗin yana kallon wurin aikinsa ta hanyar kyamarori da yawa da aka daidaita lokacinsu (yawanci kyamara ɗaya a sama da ke ba da yanayin wuri, tare da kyamarorin da aka ɗora a wuyan hannun hagu da na dama don bin diddigin taɓawa daki-daki).

Ana tura waɗannan hotunan pixel ɗanye zuwa wani vision transformer SigLIP (parameters 400M) da aka daskarar, wanda ke mayar da hotunan bidiyo zuwa visual tokens masu cike da bayani ba tare da buƙatar sabunta gradients masu nauyi ta cikin vision backbone ba yayin fine-tuning na policy.

2. Kinematic state vector da sharaɗin aiki #

A lokaci guda, tsarin yana auna yanayin cikin robot ɗin na proprioception: vector mai girma 16 da ke ɗauke da kusurwoyin haɗin gwiwa 7, gudu da faɗin buɗewar grippers na hannaye biyu.

Mai amfani yana ba da manufar aiki ta wani umarni a harshen ɗan adam (misali, "fold the t-shirt", "pick the M8 hex bolt and place it in the bracket" ko "route the cable through the conduit"), wanda ake raba shi zuwa text embeddings na yau da kullum.

3. Cibiyar cross-attention mai hanyoyi da yawa (Gemma VLM) #

Ana saka visual tokens, proprioceptive state vector da text tokens cikin Gemma Vision-Language Model (parameters 2B). Maimakon kawai ya samar da text tokens, Gemma VLM yana yin cross-attention mai zurfi a duk hanyoyi uku, yana haɗa babbar manufar umarnin harshe da fasalin wurin abin da ake gani da yanayin injin hannun a wannan lokacin.

Abin da wannan haɗaɗɗen foundation model ke fitarwa shi ne wakilci mai cike da task features na VLM wanda ke ɗauke da abin da ya kamata a yi dangane da inda abubuwa suke.

4. Action Expert: flow matching da diffusion policies #

Ana tura features na VLM kai tsaye zuwa Action Expert (parameters 300M) da ke aiki bisa ƙa'idojin flow matching. Ba kamar hanyoyin imitation learning na farko da suka dogara da sauƙaƙan mean squared error (MSE) regression ba (wanda ke sa model ya fitar da matsakaicin nunin da suka saɓa wa juna, har robot ya zama mai jinkiri ko ya tsaya cak), flow matching yana kwaikwayon rarrabuwar hanyoyin ayyukan ɗan adam masu fuskoki da yawa.

Ta amfani da rectified flow differential equations da ake warwarewa a matakai goma na lissafi, Action Expert yana tsarkake Gaussian noise ya zama hanya mai jituwa ta ayyuka 30 na gaba a mitar sarrafawa ta 50Hz.

5. Action chunking tare da temporal ensembling #

Ta hanyar hasashen matakan lokaci 30 a jere (hanyar da ake kira Action Chunking), policy ɗin yana kawar da girgiza mai sauri da jinkiri. Robot ɗin yana aiwatar da rukunin ayyuka na farko yayin da yake lissafa shiri na gaba a bayan fage, yana samun motsi mai santsi, mai sassauci kuma mai ci gaba kusan kamar na halitta.

Kwatanta fasaha: automation na gargajiya da Physical AI na buɗaɗɗen tushe #

Ma'auni / fannin tsari	Industrial robotics na gargajiya (ROS / PLC / teach pendant)	Physical AI na kamfani a rufe (humanoid stacks na masu sayarwa)	Physical AI na buɗaɗɗen tushe (stack na Hugging Face LeRobot)
Tsarin ci gaba	Procedural; hanyoyin haɗin gwiwa a fili da finite state machines	Deep learning na end-to-end; weights a rufe kamar baƙin akwati; cloud APIs a kulle	VLA na end-to-end; weights a buɗe, bayanai a buɗe da PyTorch training loops a buɗe
Shingen kayan aiki na shiga	£30,000 zuwa £100,000+ kowane hannun masana'antu (KUKA, FANUC, ABB)	£150,000 zuwa £350,000+ kowane humanoid na kamfani	Daga ƙasa da $200 zuwa $1,500 da ƙirar buɗaɗɗen kayan aiki da ake bugawa da 3D (SO-100, Koch)
Jure sauyin yanayi	Yana gazawa nan take idan wurin wani sashe ya kauce da milimita	Daidaitawa mai girma, amma ba a iya duba ko gyara halayensa a gida	Daidaitawar gani mai girma; ana iya fine-tuning na policies a gida a edge nodes na benen masana'anta
Tsarin bayanai da telemetry	Tsare-tsaren binary na kamfani da PLC registers da ke kulle ga mai sayarwa	Telemetry na cloud na kamfani da aka ɓoye	Bayanan Apache Parquet iri ɗaya, MP4 video streams da Hugging Face Hub
Gama ayyuka gaba ɗaya	Babu; kowane sabon siffar sashe na buƙatar sake tsarawa da hannu gaba ɗaya	Iƙirarin ayyuka masu faɗi, amma jinkiri wajen daidaitawa da kayan aiki na musamman	Daidaitawa cikin sauri da nunin mutane 50 zuwa 100 da aka naɗa ta leader-follower rigs
Kayan aikin inference	PLC na musamman ko industrial IPC da ke gudanar da real-time Linux	Manyan GPU clusters a cikin robot da ke cin 300W zuwa 800W	Inference na gudana a kan edge silicon na masu amfani (RTX 4070 guda ko Apple Silicon M-series)

Tushe: Nazarin bayanan gabatarwa, gwaje-gwajen fasaha da nunin kai tsaye a taron Economist Enterprise Physical AI and Robotics Summit (Convene 200 Aldersgate, London, 13 ga Oktoba 2026).

Abin da wannan ke nufi a kowane fanni #

Tasowar Physical AI a buɗe wanda al'umma ke tafiyarwa yana sauya lissafin kasuwanci a masana'antu da dama.

Ƙananan da matsakaitan kamfanoni (SMEs) #

SMEs ne ke da sama da 80% na ƙarfin masana'antu na Turai da UK, amma a tarihi ƙasa da 5% ne suka iya biyan kuɗin automation. Kafa sel na industrial automation yana buƙatar ɗaukar ƙwararrun systems integrators, biyan kuɗin lasisin software na lambobi shida, da dakatar da benen masana'anta lokacin girkawa.

Da tsarin LeRobot, SME da ke ƙera brackets na musamman masu daidaito zai iya buga wurin aiki na SO-100 mai hannaye biyu da 3D ƙasa da £500, ya saka masa USB webcams na £40, kuma ya naɗa nuni hamsin na ma'aikaci yana saka sassa a cikin CNC milling machine ta amfani da leader arm da aka buga da 3D. Cikin sa'o'i uku na fine-tuning a kan kwamfutar tebur ta yau da kullum, policy na buɗaɗɗen tushe yana iya gudanar da aikin kula da injin da kansa, kuma kuɗin ya dawo cikin makonni.

Bincike na ilimi da manyan makarantu #

Har zuwa kwanan nan, sassan robotics na jami'o'i suna fuskantar babban rabuwa: ɗakunan gwaje-gwaje masu kuɗi na iya sayen dandamalin Boston Dynamics ko Willow Garage, yayin da ƙananan cibiyoyi suka taƙaitu ga simulation kawai a Gazebo ko MuJoCo.

Physical AI na buɗaɗɗen tushe ya mayar da gwaje-gwajen zahiri abin da kowa zai iya yi. Ɗaliban digiri na gaba a duk duniya na iya gina dandamali iri ɗaya, raba checkpoints na policies da aka horar a Hugging Face Hub, kuma su maimaita sakamakon gwaje-gwaje da tsauraran ƙididdiga. Kooijmans ya jaddada cewa ma'ajiyar LeRobot ta girma daga ɗan ƙaramin rukunin masu ba da gudummawa na Turai zuwa dubban masana robotics masu aiki a duk duniya cikin watanni kaɗan, abin da ke nuna cewa buɗaɗɗiyar dama na hanzarta tabbatarwa ta gwaji fiye da rufaffun ɗakunan kamfanoni.

Haɗa kaya iri-iri masu ƙanƙantar adadi bisa kwangila #

A haɗa kayan lantarki da na'urorin likitanci, nau'ikan kayayyaki na canzawa kowane mako. Layukan automation da aka ƙayyade a lamba ba su da riba saboda kuɗin sake shirya kayan aiki da sake tsara lamba ya fi ribar samarwar.

Policies na vision-language-action suna warware wannan matsala ta hanyar ba da aiki a harshen ɗan adam. Ma'aikaci zai iya sauya tsarin haɗawa ta hanyar rubuta sabon umarni kawai ko nuna sabon haɗin sau biyu. Hanyoyin cross-attention na model suna amfani da wakilcin gani da aka riga aka horar don gane sababbin sassa ba tare da sake horar da dukkan neural network ba.

Rusa katangar bayanai: teleoperation, Parquet da world models #

Babban darasin da Kooijmans ya ba shugabannin kamfanoni shi ne babban cikas a Physical AI ba ƙarfin model ba ne yanzu; bambancin nunin ne.

A natural language processing, an horar da models a kan tiriliyoyin tokens da aka tattara daga buɗaɗɗen yanar gizo. A robotics na zahiri, babu wani “internet na sarrafa abubuwa na zahiri” da ya riga ya kasance. Kowane nunin gripper yana ɗaukar abu, juya bawul ko ninka zane dole ne a naɗa shi a duniyar zahiri ta hanyar teleoperation da na'urori masu auna.

Hugging Face LeRobot na tunkarar wannan ƙalubale ta manyan shirye-shirye uku:

Leader-follower teleoperation rigs: Ta hanyar ƙera hannayen “leader” masu araha da aka buga da 3D, waɗanda ke da magnetic encoders da ke kwaikwayon robot ɗin “follower” mai aiki ta hanyar inji, ma'aikata na iya sarrafa robot ɗin daga nesa cikin sauƙi nan take, suna naɗa kusurwoyin haɗin gwiwa, gudu da bidiyo daga kusurwoyi da yawa a lokaci guda.
Tsare-tsaren bayanai na Parquet iri ɗaya: Ana tattara nunin a cikin teburorin Apache Parquet iri ɗaya masu sauri da aka haɗa da MP4 video feeds a dunƙule. Wannan yana ba kayan aikin data engineering na yau da kullum (Polars, DuckDB, PyArrow) damar sarrafa terabytes na telemetry na robot cikin sauri.
Sim-to-real da faɗaɗawa da world models: Don wuce iyakokin teleoperation na zahiri, world models na buɗaɗɗen tushe suna koyon yanayin motsin zahiri ta hanyar generative. Bayan an horar da su a kan muhimman bayanan nunin duniyar zahiri, world models suna kwaikwayon miliyoyin yanayi na hasashe (haske daban-daban, ƙara tarkace, sauya gogayyar tebur) kai tsaye a cikin latent space, suna gwada policies sosai kafin a tura su duniyar zahiri.

Kammalawa #

Gabatarwar Hugging Face LeRobot a taron Economist Enterprise Physical AI and Robotics Summit ta fito da wani muhimmin mataki na sauyi a masana'antar robotics. Zamanin da robotics ke buƙatar kasafin kuɗi na miliyoyin fam, harsunan tsara lamba na kamfani da kejin masana'anta masu tsauri yana ba da wuri ga tsari a buɗe, na haɗin gwiwa kuma wanda bayanai ke jagoranta.

Ta haɗa foundation models na vision-language masu hanyoyi da yawa da action chunking na flow matching da kayan aiki a buɗe masu araha, Physical AI na buɗaɗɗen tushe yana mayar da robot daga automata masu rauni zuwa tsarin da ke daidaitawa kuma yake koyo. Ga shugabannin injiniya da manyan jami'an fasaha, umarnin dabaru a bayyane yake: fifikon gasa ba zai ƙara fitowa daga mallakar kayan aikin inji na kamfani ba, sai dai daga ƙwarewa a data pipelines, hanyoyin tattara nuni da foundation models da ke koyar da injuna su ƙware a duniyar zahiri.

Tambayoyin da Ake Yawan Yi #

Menene Hugging Face LeRobot? Hugging Face LeRobot wani framework ne na buɗaɗɗen tushe kuma shirin al'umma da aka tsara don mayar da robotics da Physical AI abin da kowa zai iya samu. Yana ba da aiwatarwar PyTorch na mafi ci gaban algorithms na imitation learning, tsare-tsaren bayanai iri ɗaya, foundation models na Vision-Language-Action (VLA) da aka riga aka horar, da ƙirar kayan aiki a buɗe (kamar SO-100) don rage shingayen kuɗi da na fasaha a robotics.

Ta yaya imitation learning ya bambanta da tsara robot na gargajiya? Tsara robot na gargajiya yana buƙatar injiniya ya rubuta lambar procedural a fili da ke ƙayyade ainihin Cartesian coordinates, mafita ta inverse kinematics da sauye-sauyen yanayi ta amfani da teach pendants ko ROS scripts. Imitation learning kuwa yana horar da neural networks su kwaikwayi halayen ɗan adam ta hanyar kallon nunin da aka naɗa na mutane suna yin aikin ta teleoperation, yana koyon policy da ke haɗa abin da na'urori suka gani kai tsaye da umarnin mota.

Menene aikin Gemma VLM a pipeline na Physical AI? A tsarin Physical AI na zamani, Gemma Vision-Language Model (VLM) shi ne injin tunani mai hanyoyi da yawa. Yana karɓar visual tokens daga kyamarori (da aka ɓoye ta SigLIP), joint state vector na robot, da umarni a harshen ɗan adam. Ta amfani da cross-attention, yana haɗa manufar ma'ana da fasalin gani don samar da wakilcin aiki na babban mataki da ke jagorantar action expert.

Menene flow matching a hasashen ayyukan robot? Flow matching wata dabarar generative modeling ce da ke da alaƙa da continuous normalizing flows da diffusion models. A robotics, tana mayar da sauƙaƙan Gaussian noise zuwa hanyoyin ayyuka masu sarƙaƙiya masu fuskoki da yawa ta hanyar warware ordinary differential equations a matakai da yawa. Wannan yana ba robot damar tunkarar yanayi masu ruɗani inda akwai motsi masu inganci da yawa, yana guje wa kurakuran matsakaici da ke sa sauƙaƙan regression models su gaza.

Shin hannayen robot da aka buga da 3D da ba su kai $200 ba za su iya yin ayyukan samarwa? Duk da cewa hannaye da ba su kai $200 ba kamar SO-100 ba su da ƙarfin ɗaukar nauyi, taurin tsari da maimaita daidaito ƙasa da milimita da ake buƙata wajen sassaƙa ƙarfe mai nauyi na aerospace, suna da cikakken ikon sarrafa abubuwa masu laushi, dubawa, sarrafa ruwa a ɗakin gwaje-gwaje da tattara ƙananan sassa. Mafi muhimmanci, policies da aka horar kuma aka tabbatar a kan waɗannan dandamali masu araha za a iya tura su kai tsaye zuwa hannayen collaborative robot na masana'antu da ɗan fine-tuning kaɗan.

Manazarta #

Economist Enterprise, (2026). Physical AI & Robotics Summit: fifikon aiki ta hanyar tsarin autonomous ⧉. Economist Enterprise.
Kooijmans, P., (2026). Open for business: yadda buɗaɗɗen tushe ke hanzarta Physical AI ⧉. Hugging Face.
Hugging Face, (2026). LeRobot: mafi ci gaban machine learning don robotics na duniyar zahiri a PyTorch ⧉. GitHub.
Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., da Song, S., (2024). Diffusion Policy: koyon visuomotor policy ta hanyar action diffusion ⧉. Columbia University da MIT.
Zhao, T. Z., Kumar, V., Levine, S., da Finn, C., (2023). Koyon sarrafa abubuwa da hannaye biyu daki-daki da kayan aiki masu araha ⧉. Stanford University.
Zhai, X., Mustafa, B., Kolesnikov, A., da Beyer, L., (2023). Sigmoid loss don language image pre-training (SigLIP) ⧉. Google DeepMind.
Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nicklas, M., da Le, M., (2023). Flow matching don generative modeling ⧉. Meta AI Research.

Bita ta ƙarshe 2026-10-13.

Bita ta ƙarshe 2026-10-27.

Sake buga wannan labari

Physical AI na buɗaɗɗen tushe: LeRobot na buɗe ƙofar robotics ga kowa — Sebastien Rousseau

A taron Economist Enterprise a London, Pepijn Kooijmans ya nuna yadda LeRobot, VLA models na buɗaɗɗen tushe da kayan aiki na $200 ke buɗe ƙofar robotics ga kowa.

An lasisin wannan labari a karkashin Creative Commons Attribution 4.0 International. Sake bugawa na bukatar nuna asalin URL na asali.

Physical AI na buɗaɗɗen tushe: LeRobot na buɗe ƙofar robotics ga kowa — Sebastien Rousseau

A taron Economist Enterprise a London, Pepijn Kooijmans ya nuna yadda LeRobot, VLA models na buɗaɗɗen tushe da kayan aiki na $200 ke buɗe ƙofar robotics ga kowa.

Originally published at https://sebastienrousseau.com/ha/2026-10-13-physical-ai-budadden-tushe-lerobot-bude/ by Sebastien Rousseau.
Licensed under CC-BY-4.0.

SEBASTIEN ROUSSEAU FOUNDER · ENGINEER