Robotsko pregledovanje in manipulacija tekstila in tkanin (RTFM)
Oznaka in naziv projekta
J2-4457 Robotsko pregledovanje in manipulacija tekstila in tkanin (RTFM)
J2-4457 Robot Textile and Fabric Inspection and Manipulation (RTFM)
Logotipi ARRS in drugih sofinancerjev
Projektna skupina
Vodja projekta: izr. prof. dr. Andrej Gams (25638)
Sodelujoče raziskovalne organizacije:
Institut "Jožef Stefan", Odsek za avtomatiko, biokibernetiko in robotiko (0106-014), (IJS)
Univerza v Ljubljani, Fakulteta za računalništvo in informatiko, Laboratorij za umetne vizualne spoznavne sisteme, (1539-011), (FRI)
Sestava projektne skupine:
Peter Nimac (55794) - IJS
Matija Mavsar (51232) - IJS
Simon Reberšek (39258) - IJS
Vitjan Zavrtanik (53924) - FRI
Vsebinski opis projekta
Vsebinski opis projekta
Manipulacija tekstila in tkanin oz. blaga je pomembno področje robotskih raziskav z aplikacijami tako v industriji kot v domovih. Vendar pa je napredek pri robotski manipulaciji takih deformabilnih predmetov zaostal za napredkom manipulacije togih predmetov zaradi veliko bolj zapletene dinamike in konfiguracijskega prostora. V tem projektu bomo uporabili nove, napredne metode globokega učenja in prenosa simulacije v realnost za reševanje praktičnega in obstoječega problema robotskega pregledovanja in manipulacije tekstila in tkanin. Nadgradili bomo najsodobnejše metode zaznavanja/pregledovanja ter robotske manipulacije tekstila in blaga ter premostili tehnološko vrzel, ki onemogoča avtomatizacijo tovrstnega ravnanja z materiali. Tehnološki napredek bomo demonstrirali na stopnji tehnološke zrelosti TRL 4 – tehnologija, prikazana v laboratoriju. Rezultati tega projekta bodo osnova za potencialne prihodnje, aplikativne izvedbe, ki bodo povečale konkurenčnost slovenskih in evropskih podjetij, ki se ukvarjajo s proizvodnjo in logistiko izdelkov iz blaga in tekstila. V okviru projekta bomo reševali naštete probleme 1) Razvili bomo vizualni sistem, ki omogoča segmentacijo, karakterizacijo in pregled manipuliranega tekstila/tkanine. Temeljil bo na robustni večmodalni segmentaciji na osnovi globokega učenja in zaznavanju ključnih točk, pomembnih za prijemanje ter na nenadzorovanem učenju za odkrivanje napak. 2) Razvili in prikazali bomo učinkovito ciljno usmerjeno manipulacijo tekstilov/tkanin. To bomo doseženo z učenjem ustreznih načinov gibanja, kjer bomo za učenje uporabili napredno globoko ter spodbujevano učenje v simulaciji in v resničnem svetu, metode simulacija-resničnost in trenirali nove, celostne nevronske mreže od vida do gibanja 3) Razvili bomo metode za načrtovanje (planiranje) zaporedja dejanj, ki temeljijo na novi predstavitvi stanja kosa tekstila oz. blaga. Uporabljal se bo za generiranje grafov stanj, kjer bo vsaka povezava predstavljala prehod med stanjema vključno s potrebno robotsko akcijo za izvedbo takega prehoda. Za prikaz tehnološkega napredka bomo implementirali dvoročno robotsko celico za logistiko blaga in tekstila na nivoju tehnološke zrelosti TRL4. Celica bo zaznavala, gladila, pregledovala in zlagala kose tekstila – oblačil oz. blaga v želeno, ciljno zloženo stanje. Demonstracija bo zajemala vse glavne vidike manipulacije ter zaznavanja oz. pregledovanja tekstila.
Textile and fabric manipulation is an important area of robotics research that has applications both in the industry and in homes. Yet, advances in robotic manipulation of such deformable objects have lagged behind work on rigid objects due to the far more complex dynamics and configuration space. In this project we will apply novel, advanced deep-learning and sim-to-real transfer learning methods on a real-world problem of textile and fabric manipulation and inspection. We will advance the state of the art of perception/inspection and robotic manipulation of textile and fabric, in order to bridge the technological gap and enable automation of such material handling. We will demonstrate technological advances at technology readiness level TRL 4 – technology demonstrated in lab. The outcomes of this project will serve as foundation for future, applied implementations, which will increase the competitiveness of Slovenian and European companies that deal with production and logistics of textile and fabric items. Through the course of the project we will solve these problems. We will 1) develop a vision-based system that allows segmentation, characterization and inspection of the manipulated textiles/fabrics. It will be based on robust deep-learning-based multimodal segmentation and detection of key points relevant for grasping, as well as on unsupervised learning for defect detection. 2) We will develop and demonstrate effective goal directed handling and manipulation of textile/fabric objects. This will be achieved through learning of appropriate motion policies, where advanced deep learning and reinforcement learning methods in simulation and in the real world, sim-to-real methods and training of new, end-to-end vision-to-motion deep neural networks will be applied. 3) Finally, we will develop means to plan a sequence of actions based on a novel state representation of textile. It will be used to form graphs of states where each edge is a transition with information on the needed robot action. To demonstrate the technological advances, we will implement a bimanual robot cell for textile and fabric logistics at TRL4. It will detect, flatten, inspect and fold textiles and fabrics into desired goal states. The presented demonstration will cover all the major aspects of the project in perception/inspection and handling/manipulation of such deformable objects.
Osnovni podatki sofinanciranja so dostopni na spletni strani SICRIS.
Faze projekta in opis njihove realizacije
1. Faza - Razvoj vizualnega sistema, ki omogoča segmentacijo, karakterizacijo in pregled manipuliranega tekstila/tkanine.
2. Faza - Razvoj in prikaz učinkovite ciljno usmerjene manipulacije z uporabo metod globokega učenja.
3. Faza - Razvoj metod načrtovanja zaporedja in prehodov med stanji tekstila.
1. Phase - Development of vision-based system that allows segmentation, characterization and inspection of manipulated textiles/fabrics.
2. Phase - Development and demonstration of effective goal directed handling and manipulation with the use of deep learning methods.
3. Phase - Development of methods to plan and execute a sequence of actions to transition between different states of textile.
Bibliografske reference
D. Tabernik, J. N. Muhovič, in D. Skočaj, „Dense center-direction regression for object counting and localization with point supervision“, Pattern recogn., let. 153, št. ] 110540, str. 1–13, 2024, [Na spletu]. Dostopno na: https://www.sciencedirect.com/science/article/pii/S0031320324002917
[COBISS.SI-ID - 195818755 ]
D. Tabernik, J. N. Muhovič, in D. Skočaj, „Lokalizacija in ocenjevanje lege predmeta v treh prostostnih stopnjah s središčnimi smernimi vektorji“, 2023, str. 359–362. [Na spletu]. Dostopno na: https://erk.fe.uni-lj.si/2023/papers/tabernik(lokalizacija_in).pdf
[COBISS.SI-ID - 167408387 ]
D. Tabernik, M. Šuc, in D. Skočaj, „Automated detection and segmentation of cracks in concrete surfaces using joined segmentation and classification deep neural network“, Constr. build. mater., let. 408, št. ] 133582, str. 1–12, 2023, [Na spletu]. Dostopno na: https://www.sciencedirect.com/science/article/abs/pii/S0950061823033007?via%3Dihub
[COBISS.SI-ID - 168164611 ]
D. Tabernik, P. Mlakar, J. Božič, L. Čehovin Zajc, V. Rijavec, in D. Skočaj, „Demonstracijska celica za prikaz globokega učenja v praktičnih aplikacijah“, 2024, str. 25–36. [Na spletu]. Dostopno na: https://press.um.si/index.php/ump/catalog/view/859/1246/3897
[COBISS.SI-ID - 191481347 ]
P. Nimac, M. Mavsar, in A. Gams, „Cloth smoothing simulation with vision-to-motion skill model“, 2022, str. 220–223. [Na spletu]. Dostopno na: https://erk.fe.uni-lj.si/2022/erk22.pdf
[COBISS.SI-ID - 122679043 ]
P. Nimac in A. Gams, „Cloth flattening with vision-to-motion skill model“, 2023, str. 367–374. [Na spletu]. Dostopno na: https://link.springer.com/chapter/10.1007/978-3-031-32606-6_43
[COBISS.SI-ID - 154215939 ]
P. Nimac in A. Gams, „Vpliv parametrov barvnega modela pri robotskem ravnanju tekstila“, 2023, str. 189–192. [Na spletu]. Dostopno na: https://erk.fe.uni-lj.si/2023/papers/nimac(vpliv_parametrov).pdf
[COBISS.SI-ID - 174868483 ]
P. Nimac in A. Gams, „Determining Sample Quantity for Robot Vision-to-Motion Cloth Flattening“, 2024, str. 3–11. [Na spletu]. Dostopno na: https://link.springer.com/chapter/10.1007/978-3-031-59257-7_1
[COBISS.SI-ID - 197714179 ]
Dosežki
Zmagovalci izziva PERCEPTION na 2nd Edition Cloth Manipulation and Perception, 7th Robotic Grasping and Manipulation Competition (RGMC) 2023
Domen Tabernik, Matej Urbas, Jon Muhovič in Danijel Skočaj
ICRA 2023, ViCoS-FRI team from the University of Ljubljana in Slovenia
2. mesto na ICRA 2024 - Cloth Competition Cloth Manipulation track, 9th Robotic Grasping and Manipulation Competition (RGMC) 2024
Domen Tabernik, Andrej Gams, Peter Nimac, Matej Urbas, Jon Muhovič, Danijel Skočaj in Matija Mavsar
ICRA 2024, Team Ljubljana
The BEST Application PAPER FINALIST AWARD za prispevek predstavljen na 33. International Conference on Robotics in Alpe-Adria-Danube Region z naslovom Determining Sample Quantity for Robot Vision-to-Motion Cloth Flattening RAAD 2024 Best Paper Awards
Peter Nimac in Andrej Gams
RAAD 2024
Povezava na SICRIS