Deformable Object Manipulation

Deformable object manipulation (DOM) in robotics focuses on enabling robots to skillfully interact with and reshape objects like cloth, dough, or tissue, a task significantly more complex than handling rigid objects due to their infinite degrees of freedom and unpredictable dynamics. Current research heavily emphasizes learning-based approaches, often employing neural networks (e.g., Graph Neural Networks, Transformers, and recurrent state-space models) combined with differentiable physics simulators and techniques like model predictive control or trajectory optimization to predict and control object deformation. This field is crucial for advancing robotics in areas like surgery, manufacturing, and everyday tasks, with recent progress focusing on improving sim-to-real transfer, handling constraints, and achieving greater generalization across diverse deformable objects.