Towards Efficient Robotic Manipulation of Deformable Objects by Learning Dynamics Models and Adaptive Policies

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
School of Electrical Engineering | Doctoral thesis (article-based) | Defence date: 2024-04-19
Degree programme
71 + app. 57
Aalto University publication series DOCTORAL THESES, 64/2024
Recent years have witnessed significant progress in developing intelligent robotic systems that are able to perform manipulation tasks. One reason for this success has been the advent of learning-based approaches, which driven by improvements in deep learning techniques, have endowed robots with greater generalisation capabilities to manipulate objects varying in shape, size, and texture. However, the majority of these accomplishments have been restricted to the domain of rigid objects, while our world is replete with diverse objects that deform when manipulated. This introduces a new set of challenges, such as the need for representing their deformation and adapting the robotic manipulation actions accordingly. Nevertheless, attempts have been made to improve the efficiency of current approaches by either reducing the number of interactions required to succeed in these tasks or reducing the amount of data collected in the real world using simulation engines. Although methods have been proposed for learning to manipulate deformable objects such as garments, their adaptation capabilities still remain limited. Therefore, this dissertation proposes methods to bridge the gap in the adaptive capabilities of robotic systems for manipulating a variety of materials and objects. More specifically, it investigates methods that can learn to efficiently manipulate deformable objects in simulation, transfer the learnt skills to the real world, and examine the challenges that arise when transferring these skills. To accomplish this, the thesis first investigates the representation and modelling of deformable object dynamics using data-driven approaches, resulting in two methods for modelling the dynamics using graph-based representations. Subsequently, the thesis continues by investigating methods for enabling the learning of policies that can adapt and generalise to different objects and material properties. Thus, the dissertation proposes two approaches: adapting manipulation primitives when performing high-level planning and implementing closed-loop feedback for adapting the actions according to the object's deformation. Finally, this thesis studies a major challenge limiting approaches that learn to manipulate deformable objects in simulation: the reality gap. Here, a benchmark data set is proposed to evaluate the gap when performing a dynamic manipulation task. The results of the work comprising this dissertation show that policies learnt in simulation can adapt to a wide variety of deformable objects and can efficiently manipulate them, where closed-loop feedback can mitigate the reality gap in these approaches. Consequently, approaches based on learning in simulation can enhance the adaptability of manipulation systems, where closed-loop feedback plays a vital role in successfully transferring the learnt skills to the real world.
Supervising professor
Kyrki, Ville, Prof., Aalto University, Department of Electrical Engineering and Automation, Finland
Thesis advisor
Alcan, Gokhan, Asst. Prof., Tampere University, Finland
robotics, machine learning, deformable object manipulation
Other note
  • [Publication 1]: David Blanco-Mulero, Markus Heinonen and Ville Kyrki. Evolving-Graph Gaussian Processes. In International Conference on Machine Learning, Time Series Workshop, Virtual Workshop., July 24 2021.
    DOI: 10.48550/arXiv.2106.15127 View at publisher
  • [Publication 2]: Neea Tuomainen, David Blanco-Mulero and Ville Kyrki. Manipulation of Granular Materials by Learning Particle Interactions. IEEE Robotics and Automation Letters (RA-L), vol. 7, issue 2, pp. 5663-5670, April 2022.
    DOI: 10.1109/LRA.2022.3158382 View at publisher
  • [Publication 3]: David Blanco-Mulero, Gokhan Alcan, Fares J. Abu-Dakka and Ville Kyrki. QDP: Learning to Sequentially Optimise Quasi-Static and Dynamic Manipulation Primitives for Robotic Cloth Manipulation. Accepted for publication in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Detroit, US., October 2023.
    DOI: 10.1109/IROS55552.2023.10342002 View at publisher
  • [Publication 4]: Julius Hietala, David Blanco-Mulero, Gokhan Alcan and Ville Kyrki. Learning Visual Feedback Control for Dynamic Cloth Folding. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Kyoto, Japan. pp. 1455-1462, October 2022.
    DOI: 10.1109/IROS47612.2022.9981376 View at publisher
  • [Publication 5]: David Blanco-Mulero, Oriol Barbany, Gokhan Alcan, Adrià Colomé, Carme Torras, Ville Kyrki. Benchmarking the Sim-to-Real Gap in Cloth Manipulation. Accepted for publication in IEEE Robotics and Automation Letters (RA-L), vol. 9, issue 3, pp. 2981-2988, March 2024.
    DOI: 10.1109/LRA.2024.3360814 View at publisher