Robotic Manipulation of Deformable Objects across Object Variations and Human Feedback

QC 20261030

Abstract

Robots operating in human environments must handle objects whose appearance, geometry, physical behavior, and desired mode of manipulation vary across tasks and users. This is especially challenging for deformable objects, such as garments and textiles, whose behavior depends on contact, material response, deformation, and task intent. This thesis investigates methods for improving the generalization, robustness, and personalization of robotic manipulation systems for deformable objects.The thesis addresses this problem through three complementary directions. First, it studies object-centric representations, including skeleton-graphs and keypoints, as compact descriptions of task-relevant structure. These representations support the transfer of human-specified folding plans across garment instances and improve the generalization of imitation learning policies. Second, it considers mechanical-property variation, studying interaction-derived representations that capture object-specific mechanical context and support prediction across different deformation behaviors. Third, it investigates human feedback as a mechanism for refining policies after offline training, improving robustness through corrective interventions and personalization through feedback on preferred manipulation behavior.Overall, the results show that robust deformable object manipulation requires different mechanisms for different sources of variation. Structured representations help address visible object changes, interaction reveals mechanical properties that affect action outcomes, and human feedback provides information about failures and preferences that task success alone does not capture. The thesis contributes toward manipulation systems that can operate more reliably across changing objects, dynamics, and user-specific requirements.

Link to DiVA