SHANGHAI, July 22, 2026 (GLOBE NEWSWIRE) — ACE Robotics a ou ced that its self-developed Kai os Wo ld Model has a ked fi st ac oss fou leadi g global embodied AI be chma ks: RoboTwi 2.0, LIBERO-Plus, Wo ldModelBe ch Robot, a d D eamGe . The model also achieved multiple ew state-of-the-a t (SOTA) esults. Togethe , these be chma ks evaluate bima ual ma ipulatio , lo g-ho izo task pla i g, text-d ive embodied ge e atio , a d physical i te actio easo i g. The esults validate Kai os Wo ld Model’s leade ship i sce e u de sta di g, actio p edictio , physical co siste cy, a d c oss-task ge e alizatio fo eal-wo ld obotics.
O RoboTwi 2.0, a i te atio al be chma k fo bima ual obotic ma ipulatio , Kai os Wo ld Model delive ed supe io simulatio quality a d st o ge sim-to- eal t a sfe . Its high-quality t ai i g data sig ifica tly imp oved ze o-shot pe fo ma ce o eal dual-a m obot tasks.
O LIBERO-Plus, which measu es lo g-ho izo obotic pla i g, the model mai tai ed co siste t sce e u de sta di g th oughout complex multi-step tasks. It also avoided commo issues such as object defo matio a d b oke task logic.
O Wo ldModelBe ch Robot, Kai os Wo ld Model a ked fi st i text-to-visual ge e atio a d dy amic sce e p edictio . It accu ately t a slated atu al la guage i st uctio s i to ealistic obot behavio s while p ese vi g physical co siste cy.
O D eamGe , developed by NVIDIA GEAR Lab, Kai os Wo ld Model led both Physical Alig me t (PA) a d I te actio Fidelity (IF). Its ge e ated motio s followed ealistic physical dy amics, p oduci g simulatio s that ca mo e effectively suppo t eal-wo ld obotic deployme t.
U like the widely used i dust y app oach of combi i g a ge e al-pu pose video ge e atio model with co ti ued t ai i g a d fi e-tu i g o embodied sce a ios, the Kai os Wo ld Model follows a fu dame tally diffe e t tech ical path. I Decembe 2025, Ace Robotics i t oduced what it desc ibed as the wo ld’s fi st atively i teg ated a chitectu e fo multimodal u de sta di g, ge e atio , a d p edictio . The a chitectu e b i gs wo ld pe ceptio , sce e ge e atio , a d actio decisio -maki g i to a si gle u ified backbo e etwo k, c eati g a complete i tellige ce loop f om u de sta di g the wo ld to p edicti g it a d i te acti g with it. I Ma ch 2026, the a chitectu e was successfully validated fo edge deployme t. Nvidia’s subseque tly eleased Cosmos 3.0 adopted a a chitectu e based o the same u de lyi g desig , i dicati g that the tech ical system pio ee ed by Kai os has become a impo ta t di ectio fo the developme t of wo ld model i dust y sta da ds.
I te ms of pe fo ma ce, Kai os Wo ld Model 3.0 suppo ts lo g-du atio video ge e atio of up to seve mi utes, while p oduci g high-fidelity static a d dy amic sce es at i fe e ce speeds that exceed those of mai st eam wo ld models. Its lightweight ve sio , Kai os 3.0-4B, u s 72 times faste tha Cosmos 2.5 o a A800 GPU, setti g a ew global pe fo ma ce eco d fo embodied wo ld models. Mo e impo ta tly, Kai os 3.0-4B is the wo ld’s fi st wo ld model capable of d ivi g embodied obot co t ol at the edge. It is also the fi st embodied wo ld model i the i dust y to achieve a 1:1.5 atio of video ge e atio time to video du atio o the THOR edge platfo m, e abli g eal-time ge e atio . Robots powe ed by the model ca ca y out complex tasks equi i g lo g-du atio a d lo g-ho izo easo i g, i cludi g p epa i g b eakfast, pe fo mi g household cho es, a d wate i g pla ts. The model also add esses widely ecog ized tech ical challe ges such as fluid i te actio a d defo mable-object ma ipulatio . Its “o e b ai , multiple embodime ts” capability allows it to adapt seamlessly to obotic ha ds, huma oid obots, obotic a ms, a d othe ha dwa e platfo ms f om diffe e t ma ufactu e s.
At the applicatio level, Ace Robotics, CUHK MMLab, a d She zhe Loop A ea I stitute joi tly eleased Kai os-HomeWo ld, the wo ld’s fi st whole-home i te active 3D wo ld model. Fo the fi st time, the model e ables e d-to-e d ge e atio of a complete eside tial 3D e vi o me t f om a si gle text p ompt, with globally co siste t st uctu es, physically plausible layouts, a d i te active objects. The esea ch team also ope -sou ced the wo ld’s la gest whole-home 3D dataset a d the fi st desig ed specifically fo Chi ese households. It i cludes 300,000 floo pla s f om eal Chi ese homes a d 5,000 complete simulated e vi o me ts with i te active fu itu e a d objects. Kai os-HomeWo ld is ow used i Ace Robotics’ outi e embodied i tellige ce t ai i g, suppo ti g full-p ocess simulatio t ai i g fo complex, lo g-ho izo household tasks such as c oss- oom avigatio a d multi- oom object o ga izatio .
F om ge e ati g video to co t olli g physical obots, a d f om si gle-sce e ge e atio to fully i te active whole-home e vi o me ts, the Kai os Wo ld Model is t a sfo mi g wo ld models f om labo ato y demo st atio s i to i dust ial i f ast uctu e that suppo ts eal-wo ld obot ope atio .
Ace Robotics focuses o fou datio al esea ch a d developme t fo embodied i tellige ce wo ld models. B i gi g togethe i te atio ally ecog ized schola s a d docto al esea ch pa t e s, the compa y pio ee ed a complete tech ical loop that i teg ates ACE ambie t data collectio , wo ld models, a d ge e alizatio modules. It delive s obot b ai solutio s ac oss i dust ial applicatio s, u ba gove a ce, etail, cultu e, a d tou ism.
