Abstract
Over the years, significant advancements have occurred in robotic grasping and manipulation techniques, transitioning from early analytical methods reliant on explicit mathematical model of grasps to modern learning-based approaches capable of generating high-quality grasps across a broad spectrum of objects from a single image. However, many of these methods achieve such remarkable performance by making assumptions pertaining to the object's physical properties, including uniform rigidity and friction across the surface. These assumptions limit the effectiveness of such methods in handling more complex artefacts, such as multi-material or deformable objects, which are commonly encountered in domains such as healthcare and household tasks. This dissertation seeks to explore the potential of explicitly estimating and harnessing two crucial physical properties of the target objects -- surface friction and deformability -- to develop resilient grasp and manipulation planners beyond these assumptions. First, a probabilistic method is proposed to estimate the surface friction properties of a target object. This estimation method employs exploratory actions to obtain visuo-tactile feedback, which is then used to determine the friction coefficient values. The resulting comprehensive object representation, incorporating both shape and surface friction information, is then integrated into the grasp planning process to notably enhance the grasp success rate. Subsequently, the dissertation investigates how to harness object deformability in grasp planning, developing a generative, deformation-aware deep grasp synthesis approach that enables planning high-quality grasps while considering object stiffness. Simultaneously, the aim is to overcome the existing limitations of obstructive and time-consuming grasp evaluation techniques for deformable objects. To achieve this goal, a computationally inexpensive analytical approach and a novel grasp quality metric are proposed to facilitate the evaluation of grasps on deformable objects in a matter of seconds, significantly accelerating the data generation process. In addition to grasping, the dissertation also explores how to harness object deformability in deformable object manipulation task planning by first introducing a learning-based model to predict the interactions between a volumetric deformable tool and rigid objects, and then using the learned model in task planning. Finally, instead of object deformability, the focus shifts to soft robotic hands, where deformability is built into robotic grippers. Specifically, the dissertation investigates means of integrating position and contact force sensing capabilities into a soft robotic hand to grasp deformable objects safely without causing damage. Together, the results indicate that acquiring and harnessing knowledge of an object's physical properties beyond its shape increases the robustness and performance of grasp planning and manipulation planning methods. Therefore, the hope is that this dissertation will motivate roboticists to move beyond current assumptions and consider deeper object understanding when developing new grasping and manipulation approaches.
Translated title of the contribution | Harnessing the physical properties of objects for robotic grasping and manipulation |
---|---|
Original language | English |
Qualification | Doctor's degree |
Awarding Institution |
|
Supervisors/Advisors |
|
Publisher | |
Print ISBNs | 978-952-64-1583-3 |
Electronic ISBNs | 978-952-64-1584-0 |
Publication status | Published - 2023 |
MoE publication type | G5 Doctoral dissertation (article) |
Keywords
- robotic grasping
- robotic manipulation
- deep learning