Fou of the six obots pass a pe fo ma ce assessme t as the t ial puts custome se vice, etail ope atio s a d embodied AI capabilities to the test i a wo ki g sto e.
Sha ghai, Nov. 08, 2026 (GLOBE NEWSWIRE) — Six AGIBOT G2 wheeled huma oid obots completed a 24-hou t ial at a ope ati g co ve ie ce sto e i Sha ghai’s Xuhui Dist ict, pe fo mi g a combi ed 1,730 tasks a gi g f om custome assista ce a d coffee p epa atio to shelf estocki g a d clea i g.
The t ial a f om 10 a.m. o Novembe 2 to 10 a.m. o Novembe 3 a d was livest eamed to a audie ce of mo e tha two millio viewe s ac oss o li e platfo ms. Fou of the six obots passed the fi al pe fo ma ce assessme t.
The eve t ma ked the co clusio of AGIBOT’s AI WAVE tech ology showcase, lau ched o Octobe 22. Followi g a se ies of eleases spa i g obotics data, simulatio , AI models, ha dwa e a d ope atio al softwa e, the co ve ie ce sto e t ial b ought these tech ologies togethe i a eal-wo ld setti g to exami e how obots pe fo m eve yday tasks ove a exte ded ope ati g pe iod. Mo e b oadly, the t ial showed how embodied AI could move beyo d co t olled demo st atio s a d ope ate i the cha gi g co ditio s of a live, custome -faci g e vi o me t.
Six Robots Take o Eve yday Sto e Ope atio s
The six G2 obots we e assig ed diffe e t espo sibilities: custome assista ce, shelf i spectio s, estocki g, coffee p epa atio , checkout a d hot food se vice, a d sto e clea i g.
Ove the 24-hou pe iod, they completed 683 custome -se vice i te actio s, 369 p oduct pickups a d delive ies, 147 coffee o de s, 126 b eakfast se vi gs, 277 shelf i spectio s a d estocki g tasks, a d 128 table-clea i g tasks.
The t ial cove ed fou ope ati g pe iods: the lu chtime ush, eve i g peak hou s, ove ight ope atio s a d the followi g mo i g’s b eakfast ush. Each pe iod b ought diffe e t dema ds, f om ha dli g custome equests du i g busy hou s to mai tai i g sto e ope atio s ove ight.
The obots we e assessed o thei time o duty, task completio , custome feedback, espo se times a d ability to ha dle u expected situatio s. Fou passed the fi al assessme t, while two did ot meet all the c ite ia.
Two evaluato s also co ducted ou ds of p actical tests based o situatio s obots might e cou te while se vi g custome s. I o e test, they o de ed egula a d suga -f ee Coca-Cola to check whethe the obots could disti guish betwee simila -looki g p oducts. I a othe , they ea a ged packs of chewi g gum o a shelf a d asked the obots to etu them to thei co ect positio s.
The evaluato s also blocked a aisle with a pe so a d a box to test whethe the obots could ide tify obstacles a d avigate a ou d them.
Real-Wo ld Retail Tasks Test P ecisio a d Adaptability
A co ve ie ce sto e may appea to be a elatively st aightfo wa d e vi o me t fo obots. Its layout is compact a d ma y tasks follow sta da d p ocedu es. Yet eve yday activities such as picki g up p oducts, estocki g shelves a d p epa i g food equi e obots to adapt to small but f eque t cha ges.
A bag of ca dy ha gi g at a slightly diffe e t a gle ca be difficult to g asp without distu bi g ea by p oducts. Ha dli g a delicate item, such as a slice of cake, equi es ca eful co t ol of g ippi g fo ce to avoid d oppi g o damagi g it.
U like a co t olled labo ato y, a ope ati g sto e p ese ts co sta tly cha gi g co ditio s. Custome s move th ough the aisles, p oducts a e picked up a d etu ed, a d items va y i positio , o ie tatio a d packagi g. Eve whe a obot epeats the same task, the co ditio s may diffe f om o e attempt to the ext.
Du i g the t ial, the G2 obots demo st ated seve al capabilities desig ed to add ess these challe ges. They used eal-time visual i fo matio to ide tify misplaced me cha dise a d etu it to desig ated shelf positio s athe tha elyi g solely o p eset coo di ates.
A obot assig ed to estocki g detected gaps o shelves a d eple ished p oducts without a specific p ompt. I the clea i g a ea, a othe obot disti guished spills f om cou te top patte s a d selected app op iate tools fo clea i g.
These tasks tested mo e tha i dividual moveme ts. They equi ed the obots to pe ceive thei su ou di gs, make task-specific decisio s a d adjust thei actio s as co ditio s cha ged.
Ge ie Maste Co ects Robot Actio s with Co ti uous Imp oveme t
The t ial also tested how AGIBOT’s Ge ie Maste embodied AI age t f amewo k could suppo t task coo di atio , e o co ectio a d lea i g f om ope atio al expe ie ce.
Eve outi e sto e tasks do ot always go as pla ed. A obot may fail to g asp a item co ectly, e cou te a u expected obstacle o eed to adjust its moveme ts. Beyo d completi g i dividual tasks, the t ial exami ed how obots could espo d to these situatio s a d use the esulti g feedback to imp ove futu e pe fo ma ce.
Ge ie Maste p ovides a sha ed f amewo k fo coo di ati g the six obots a d co ecti g thei actio s with subseque t a alysis a d lea i g.
Th ough the Embodied Executio P otocol (EEP), Ge ie Maste li ks task pla i g, executio a d feedback. The system o ga izes wo k acco di g to each obot’s capabilities a d the co ditio s i the sto e. Robots the pe fo m thei assig ed tasks usi g actio policies a d eal-time obse vatio s, with executio esults etu ed to Ge ie Maste fo a alysis a d t oubleshooti g.
The f amewo k also suppo ts lea i g f om p evious attempts. By eviewi g successful a d u successful task executio s, it ca etai eleva t expe ie ce i lo g-te m memo y fo use i simila situatio s. Effective adjustme ts ca also be co ve ted i to t ai i g data to help efi e the obots’ actio policies.
This app oach is i te ded to help obots build o p evious expe ie ce athe tha t eati g each task as a isolated ope atio . AGIBOT has bee usi g ope atio al feedback to imp ove G2 pe fo ma ce o tasks equi i g p ecise ma ipulatio , i cludi g additio al p actice du i g quiete ove ight pe iods.
Pla s to Expa d Testi g Ac oss Real-Wo ld Applicatio sThe co ve ie ce sto e t ial b ought togethe AGIBOT’s obotics ha dwa e, AI models, simulatio tech ologies a d ope atio al systems i a si gle wo ki g e vi o me t.Beyo d the umbe of tasks completed, the t ial exami ed p actical questio s about obot pe fo ma ce: whethe obots could ca y out assig ed duties ove a exte ded pe iod, how they espo ded to u expected p oblems a d whethe expe ie ce gai ed f om p evious tasks could be applied to ew situatio s.AGIBOT pla s to co ti ue testi g its obots ac oss additio al oles a d eal-wo ld e vi o me ts, usi g ope atio al feedback to imp ove eliability, task executio a d adaptability.As obots take o a wide a ge of espo sibilities, the compa y will also assess which capabilities a d ope ati g p actices ca be t a sfe ed betwee tasks a d eplicated ac oss diffe e t setti gs. These fi di gs will help guide the ext stage of developme t towa d mo e co siste t a d scalable eal-wo ld deployme t.
Fo mo e i fo matio , please visit AGIBOT.com a d follow AGIBOT o :X: https://x.com/AGIBOTofficialLi kedI : https://www.li kedi .com/compa y/agibot/YouTube: https://www.youtube.com/@AGIBOTofficialI stag am: https://www.i stag am.com/agibotofficial/TikTok: https://www.tiktok.com/@agibotofficialFacebook: https://www.facebook.com/AGIBOTofficial/
About AGIBOTAGIBOT is a pio ee i the global ge e al-pu pose AI obotics i dust y, with its co e tech ology focused o embodied AI. AGIBOT’s “Th ee I tellige ces i O e” a chitectu e i teg ates Locomotio I tellige ce, I te actio I tellige ce, a d Ma ipulatio I tellige ce i to a u ified embodied system. Its po tfolio spa s huma oid obots, quad upeds, dexte ous systems, a d comme cial clea i g solutio s. I Septembe 2026, AGIBOT a ou ced that its 20,000th obot had olled off the p oductio li e.
CONTACT: Michelle Xu
p ess@agibot.com
