DEEP Robotics Unveils Five Research Advances in Embodied AI

New esea ch spa s wo ld models, visio -la guage-actio models, ei fo ceme t lea i g, p edictive memo y a d eal-wo ld obot be chma ks

BOSTON, Oct. 24, 2026 (GLOBE NEWSWIRE) — DEEP Robotics has u veiled five esea ch adva ces suppo ti g the developme t of DEEPRobotics-B ai , its embodied i tellige ce a chitectu e fo eal-wo ld obot ope atio . Spa i g physical p edictio , avigatio a d ma ipulatio , policy optimizatio , p edictive memo y, a d data-d ive evaluatio , the esea ch add esses a sha ed challe ge: helpi g obots wo k eliably i e vi o me ts that cha ge.

The five esea ch a eas comp ise the DR-WM-1.0 wo ld model, the DR-VLA-2.0 embodied fou datio model, the DR-APO policy optimizatio f amewo k, the DR-Age t embodied age t, a d the DR-Be chma k data a d evaluatio platfo m. Togethe , they explo e how obots ca a ticipate actio outcomes, t a slate i st uctio s i to motio , imp ove lea ed policies, evise outdated spatial beliefs, a d measu e p og ess usi g eal-wo ld data.

The five adva ces add ess i te co ected challe ges ac oss this a chitectu e. The esea ch i volves collabo atio s with esea che s f om the Massachusetts I stitute of Tech ology, the U ive sity of Califo ia Sa Diego, the U ive sity of Illi ois U ba a-Champaig , Tsi ghua U ive sity, Zhejia g U ive sity, the Ho g Ko g U ive sity of Scie ce a d Tech ology, a d the Beiji g I stitute of Tech ology. Togethe , the studies co ect physical p edictio , avigatio a d ma ipulatio , policy lea i g, dy amic memo y, a d eal-wo ld data.

The esea ch suppo ts DEEP Robotics’ b oade goal of developi g obots that ca espo d to huma i st uctio s a d p actical ope ati g co ditio s, with capabilities evaluated th ough simulatio , eal- obot expe ime ts, a d sha ed be chma ks.

DR-WM-1.0 B i gs Physical G ou di g to Wo ld ModelsWo ld models offe a way to evaluate actio s befo e a obot executes them. Co ve tio al obot t ai i g elies o epeated eal-wo ld t ial a d e o , which is costly a d time-co sumi g. Video-ge e ative wo ld models ca simulate visually ich e vi o me ts, but visual plausibility alo e does ot establish that a p edicted actio is physically feasible. A co vi ci g video of a obot c ossi g s ow, fo example, may fail to captu e slippi g caused by low f ictio .

Developed with esea che s f om Zhejia g U ive sity, the Beiji g I stitute of Tech ology, Tsi ghua U ive sity, MIT, a d the U ive sity of Illi ois U ba a-Champaig , DR-WM-1.0 add esses this gap by i co po ati g obot–te ai co tact dy amics i to actio co seque ce p edictio . The associated DR-WM study combi es eco st ucted sce e geomet y with de se physical p io s fo f ictio , ough ess, a d stiff ess, calib ated usi g eal-wo ld actio – espo se data.

The esea ch d aws o 2.58 millio image–mask pai s a d mo e tha 100,000 eal- obot clips. Its dual-Co t olNet a chitectu e co ditio s video ge e atio o sce e geomet y a d came a motio de ived f om simulated obot moveme t. Du i g auto eg essive ollouts, obot a d co t olle states a e ca ied fo wa d, mai tai i g co ti uity with the u de lyi g co t ol loop.

Figu e 1. DR-WM wo kflow a d motio outcomes u de diffe e t co tact co ditio s.

This gives the obot a fo m of physical ehea sal: a way to estimate what might happe alo g a oute a d compa e alte ative actio s. By co ecti g pe ceptio , physical p edictio , a d decisio -maki g, DR-WM p ovides a testable app oach to a owi g the gap betwee visual ealism a d physically g ou ded p edictio .

A illust ative applicatio is substatio i spectio afte ai o s ow. Befo e choosi g betwee a slippe y sho tcut a d a lo ge oute, a obot could simulate both optio s, assess the likelihood of slippi g, u stable motio , o limited p og ess, a d weigh those isks agai st t avel efficie cy.

DR-VLA-2.0 Co ects Huma Expe ie ce with Robot Navigatio a d Ma ipulatio Fo obots to ope ate i st eets, campuses, a d delive y e vi o me ts, avigatio must exte d beyo d path pla i g. They eed to follow i st uctio s, avoid obstacles, a ticipate cha ges, a d move i ways that a e compatible with huma activity.

Collecti g obot demo st atio s th ough teleope atio is expe sive a d difficult to scale ac oss dive se e vi o me ts a d i te actio s. Huma egoce t ic web videos p ovide exte sive examples of obstacle avoida ce, yieldi g, a d oute selectio . Howeve , diffe e ces i came a viewpoi t, imagi g cha acte istics, a d motio p eve t di ect t a sfe to obots, while the videos lack executable obot actio labels.

Wo ki g with esea che s f om MIT, the U ive sity of Illi ois U ba a-Champaig , Tsi ghua U ive sity, a d Zhejia g U ive sity, DEEP Robotics explo es a lea i g pipeli e that combi es actio -f ee p et ai i g o huma video with adaptatio usi g obot demo st atio s. The associated DR-VLA study i t oduces EgoLy x, comp isi g 10,000 hou s of cu ated egoce t ic web video a d 2,000 teleope ated obot avigatio clips.

Figu e 2. DR-VLA ove view. Huma egoce t ic videos supe vise u de sta di g a d p edictio ; obot demo st atio s additio ally supe vise executable actio s.

The ce t al mecha ism is a domai -i va ia t sce e ep ese tatio (DISR). It combi es sema tic, spatial, a d elatio al i fo matio , allowi g the model to focus o what is p ese t, whe e it is, a d how sce e eleme ts elate to o e a othe . This st uctu al ep ese tatio p ovides a i te face fo t a sfe i g avigatio k owledge ac oss huma a d obot obse vatio s.

A Mixtu e-of-T a sfo me s f amewo k b i gs togethe visio -la guage u de sta di g, late t wo ld modeli g, a d a flow-matchi g actio expe t. Huma video supplies supe visio fo sce e u de sta di g a d p edictio ; obot demo st atio s supply the actio supe visio eeded to ge e ate executable t ajecto ies. The app oach is i te ded to suppo t socially complia t behavio s such as waiti g, bypassi g obstacles, a d yieldi g to pedest ia s.

O the held-out EgoLy x-Robot test split, DR-VLA imp oves the Quad uped Navigatio Sco e (QNS) by app oximately 5% a d educes ope -loop fi al headi g e o by 14.6%, each elative to the best competi g esult fo that met ic. QNS combi es safety, i st uctio complia ce, comfo t, a d p og ess. These esults poi t towa d mo e atu al obot avigatio i sha ed spaces, i cludi g last-mile delive y i eside tial commu ities a d campuses.

Ma ipulatio p ese ts a elated challe ge: t a slati g la guage i to eliably co t ollable motio . I existi g visio -la guage-actio (VLA) models, st o g visual p io s ca ove shadow la guage i st uctio s. Cha gi g a comma d may fail to edi ect a a m, while co tact- ich tasks such as i se tio a d wipi g equi e motio to adapt to fo ce feedback.

DEEP Robotics a d its esea ch collabo ato s add ess this p oblem with Atomic Motio Coo di ate (AMC), a f amewo k fo la guage-stee able a d fo ce- espo sive ma ipulatio . AMC defi es 13 i te p etable motio atoms fo each a m: positive a d egative t a slatio s alo g th ee axes, positive a d egative otatio s about th ee axes, a d a hold actio . A i itial t ai i g stage excludes visual i puts to alig la guage with geomet ic motio . Subseque t stages use visual co text to ealize those motio te de cies a d i co po ate fo ce feedback to adjust ge e ated actio s withi the same coo di ate f amewo k.

Figu e 3. With images a d obot states fixed, opposite motio -atom p ompts p oduce opposi g e d-effecto t a slatio a d otatio espo ses.

AMC achieves opposite-atom sepa atio ates of 92.5% fo si gle-atom p ompts a d 83.1% fo two-atom p ompts. I eal bima ual tasks with u see layouts, o malized task p og ess eaches 87.8%. Addi g fo ce feedback aises task p og ess to 78.5% fo the d awe -a d-plug task a d 75.2% fo vase wipi g. The evaluatio s demo st ate imp oveme ts i la guage stee ability, ge e alizatio to held-out co figu atio s, a d co tact adaptatio elative to the tested baseli es, i cludi g fi e-tu ed π0.5 a d a ep oduced LA4VLA-style app oach.

Pote tial applicatio s i clude elect ical cabi et ope atio , co ecto i se tio , a d compo e t alig me t. The esults suppo t mo e i te p etable co t ol of ma ipulatio .

DR-APO Imp oves Rei fo ceme t Lea i g Ac oss Policy Rep ese tatio sAs avigatio a d ma ipulatio models g ow mo e capable, t ai i g methods must suppo t i c easi gly dive se actio ep ese tatio s. Existi g ei fo ceme t lea i g app oaches a e ofte tailo ed to a pa ticula policy pa amete izatio , limiti g thei euse ac oss models a d tasks.

Developed by DEEP Robotics a d Zhejia g U ive sity, DR-APO (Decoupled-Rep ese tatio Adva tage-Guided Policy Optimizatio ) sepa ates policy imp oveme t f om the u de lyi g policy ep ese tatio . The f amewo k p ovides a sha ed optimizatio p i ciple fo Gaussia , diffusio , a d flow-matchi g policies, suppo ti g both co ti uous-co t ol lea i g f om sc atch a d post-t ai i g of p et ai ed embodied avigatio policies.

Figu e 4. DR-APO combi es adva tage-guided ta get imp oveme t with ep ese tatio -compatible actio fitti g fo co ti uous co t ol a d avigatio post-t ai i g.

Its adva tage-guided ta get imp oveme t mecha ism co st ucts imp oved actio ta gets usi g local adva tage g adie ts at executed actio s a d adva tage-based eweighti g. Policy imp oveme t is fo mulated as ta get optimizatio i depe de t of the policy pa amete izatio . This avoids explicit policy-likelihood evaluatio fo acto updates a d does ot equi e c itic-based selectio amo g multiple ca didate actio s at i fe e ce time.

Rep ese tatio -compatible di ect actio fitti g the t a sfe s these ta gets to Gaussia , diffusio , a d flow policies. This is desig ed to p eve t local imp oveme t sig als f om bei g diluted du i g multistep ge e atio .

Ac oss five MuJoCo-v5 co ti uous-co t ol tasks, DR-APO achieves the best esult i i e of te diffusio -policy compa iso s a d all te flow-policy compa iso s epo ted i the study. It delive s the best agg egate policy-o ly pe fo ma ce fo both policy families, without i fe e ce-time multi-ca didate c itic selectio , while emai i g competitive with Soft Acto -C itic (SAC) fo Gaussia policies.

I embodied visual avigatio post-t ai i g, DR-APO aises p et ai ed NavDP’s simulated success ate f om 70.33% to 89.55%, a i c ease of 19.22 pe ce tage poi ts. I eal-wo ld avigatio tests o the DEEP Robotics LYNX M20 wheel-legged obot, success ises f om 25% to 65%, a gai of 40 pe ce tage poi ts. These fi di gs demo st ate the f amewo k ac oss co ti uous-co t ol lea i g a d avigatio post-t ai i g, p ovidi g a basis fo fu the esea ch o othe embodied policies.

DR-Age t Updates Spatial Beliefs as the Wo ld Cha gesReliable lo g-te m ope atio equi es obots to easo about cha ges that occu outside thei field of view. Objects a e moved i homes, hotel spaces a e ea a ged, a d i ve to y shifts i wa ehouses. A p eviously accu ate memo y ca become obsolete befo e the obot eaches its desti atio .

Developed with esea che s f om the U ive sity of Califo ia Sa Diego a d Zhejia g U ive sity, DR-Age t d aws o Evolvi gNav, a avigatio method built a ou d p edictive 4D belief. The method add esses pe siste t avigatio whe ta gets may move while u obse ved, i cludi g du i g the obot’s jou ey.

Evolvi gNav co st ucts a time-i dexed belief f om timestamped 3D object histo ies. A eve t-d ive p edict–obse ve– epla filte estimates whe e a ta get is likely to be at the expected i spectio time, i co po ates ew visual evide ce, a d evises the obot’s pla . This e ables memo y a d locatio beliefs to be updated as the e vi o me t cha ges.

Figu e 5. Ove view of the p edictive belief a d avigatio f amewo k u de lyi g DR-Age t.

The accompa yi g EvoWo ld-Be ch be chma k is g ou ded i huma activity t aces a d comp ises 54 simulated sce es, 803,680 task i sta ces, a d app oximately 239,000 dy amic state cha ges. Expe ime ts show imp oveme ts i avigatio success a d sea ch efficie cy ove the evaluated baseli es, with pa ticula ly clea gai s whe tempo al patte s ca be lea ed. Evaluatio s o Fi di gDo y a d GOAT-Be ch fu the demo st ate applicability to embodied avigatio tasks.

Pote tial uses a ge f om sustai ed auto omous se vice i homes a d sha ed spaces to object sea ch i wa ehouses a d i spectio – oute adjustme t i i dust ial facilities. P edictive, evisable memo y helps add ess a ce t al deployme t equi eme t: mai tai i g useful behavio ove epeated tasks i cha gi g e vi o me ts.

DR-Be chma k U ifies Evaluatio o Real-Wo ld Robot DataAll of these capabilities depe d o data that co ects pe ceptio , i te t, a d co ti uous obot motio . Existi g obot a d la guage- avigatio datasets ofte p ovide o ly pa t of this i fo matio . Pe ceptio , avigatio , futu e p edictio , a d eco st uctio a e also f eque tly evaluated usi g diffe e t datasets a d p otocols, maki g co siste t compa iso difficult.

DEEP Robotics developed DR-Be chma k with esea che s f om the U ive sity of Illi ois U ba a-Champaig , the Ho g Ko g U ive sity of Scie ce a d Tech ology, Zhejia g U ive sity, Tsi ghua U ive sity, a d MIT. At its co e is Ly xSce es, a la ge-scale eal-wo ld dataset a d be chma k fo outdoo wheel-legged obot ope atio .

Figu e 6. Ly x Sce es ove view, i cludi g eal- obot data collectio , multimodal obse vatio s, spatial a d behavio al a otatio s, a d fou dow st eam tasks.

Collected usi g the DEEP Robotics LYNX M20, Ly xSce es spa s 833.6 hou s a d 3,294.7 kilomete s of eal-wo ld ope atio . It co tai s app oximately 100,000 avigatio clips a d 3 millio a otated f ames, with supe visio cove i g 3D objects, oad st uctu es, avigatio i st uctio s, a d expe t demo st atio t ajecto ies. The a otatio s e compass 27 object catego ies, fou oad-st uctu e catego ies, a d 20 avigatio i st uctio s o ga ized i to th ee levels.

The be chma k b i gs fou compleme ta y tasks o to a sha ed eal-wo ld dataset: i st uctio -co ditio ed t ajecto y p edictio , bi d’s-eye-view (BEV) pe ceptio of 3D objects a d oad st uctu es, futu e sce e p edictio , a d multi-view sce e eco st uctio . The esea ch evaluates 25 ep ese tative models, e abli g mo e co siste t a alysis ac oss avigatio , pe ceptio , wo ld modeli g, a d eco st uctio .

Applicatio s i clude studyi g whe delive y obots should p oceed, wait, o detou i sha ed spaces; a alyzi g avigatio th ough a ow passages a d tempo a y obstacles du i g i dust ial i spectio ; a d eplayi g eco ded sce a ios to diag ose model behavio . DR-Be chma k p ovides a basis fo model developme t a d compa iso , while li ki g eal-wo ld ope ati g data with subseque t esea ch a d imp oveme t.

Togethe , these esea ch adva ces suppo t the developme t of DEEPRobotics-B ai a d its u de lyi g a chitectu e. They add ess compleme ta y equi eme ts fo eal-wo ld obotics: physically g ou ded p edictio , i st uctio -d ive avigatio a d ma ipulatio , policy imp oveme t, evisable spatial memo y, a d co siste t evaluatio o eco ded obot expe ie ce.

DEEP Robotics is pu sui g this coo di ated esea ch p og am to suppo t developme t ac oss obot embodime ts. By co ecti g these esea ch di ectio s with ve ifiable methods a d sha ed be chma ks, the compa y aims to adva ce embodied AI towa d eliable i dust ial applicatio s a d sustai ed huma – obot collabo atio .

The esea ch p ep i ts a e available i the esea ch sectio of the DEEP Robotics Develope Ce te : https://develope .deep obotics.c / esea ch/ .

Media Co tact

Compa y: DEEP RoboticsCo tact: Vivia Che Email: che li gjia@deep obotics.c Website: https://www.deep obotics.c /e Disclaime :  This co te t is p ovided by DEEP Robotics. The stateme ts, views, a d opi io s exp essed i this co te t a e solely those of the co te t p ovide a d do ot ecessa ily eflect the views of this media platfo m o its publishe . We do ot e do se, ve ify, o gua a tee the accu acy, complete ess, o eliability of a y i fo matio p ese ted. This co te t is fo i fo matio al pu poses o ly a d should ot be co side ed fi a cial, i vestme t, o busi ess advice. All i vestme ts ca y i he e t isks, i cludi g the pote tial loss of capital. Reade s a e st o gly e cou aged to co duct thei ow due dilige ce a d co sult with a qualified fi a cial adviso befo e maki g a y i vestme t decisio s. Neithe the media platfo m o the publishe shall be held espo sible fo a y i accu acies, mis ep ese tatio s, o fi a cial losses esulti g f om the use o elia ce o the i fo matio i this p ess elease. Speculate o ly with fu ds you ca affo d to lose. I the eve t of a y legal claims o co ce s ega di g this a ticle, we accept o liability o espo sibility. Globe ewswi e does ot e do se a y co te t o this page.

Legal Disclaime : This media platfo m p ovides the co te t of this a ticle o a “as-is” basis, without wa a ties o ep ese tatio s of a y ki d, exp ess o implied. We assume o espo sibility o liability fo the accu acy, co te t, images, videos, lice ses, complete ess, legality, o eliability of the i fo matio co tai ed he ei . A y complai ts, copy ight issues, o i qui ies ega di g this a ticle should be di ected to the co te t p ovide listed above.

Photos accompa yi g this a ou ceme t a e available at

https://www.globe ewswi e.com/NewsRoom/Attachme tNg/d9904360-3a8b-4d37-9a4c-617f11d853e2

https://www.globe ewswi e.com/NewsRoom/Attachme tNg/1c879e9e-e245-4fa4-8edc-bd16f32aa8fe

https://www.globe ewswi e.com/NewsRoom/Attachme tNg/58acf734-f667-40ed-b8dc-79ae827b4276

https://www.globe ewswi e.com/NewsRoom/Attachme tNg/a633fd25-bbf8-427e-b0a4-b88bd249222a

https://www.globe ewswi e.com/NewsRoom/Attachme tNg/a2325259-2967-45dc-a9c5-91eccb5ddd21

https://www.globe ewswi e.com/NewsRoom/Attachme tNg/c4e24e92-21af-4333-83a6-45e6d4a0a231