A Unified Deep Learning System for Category-Agnostic Deformable Object Manipulation in Unstructured Environments

Deformable object manipulation is difficult because soft materials change shape in complex and unpredictable ways, especially in unstructured environments. This work presents a unified deep learning system that learns category agnostic representations of deformable object states directly from sensory observations, without relying on object specific models or templates, and uses these representations to support robust manipulation across varied materials and task settings. Using a contrastive representation learning approach within an end to end framework that links perception and control, the system is evaluated on multiple deformable manipulation tasks and demonstrates improved generalisation to unseen objects and configurations compared with category specific baselines, suggesting a scalable pathway towards more flexible real world deformable manipulation.