Manipulation of deformable objects is crucial for countless applications in household, caregiving, industrial, and surgical robotics. However, deformable objects violate many fundamental assumptions of rigid object manipulation, requiring different approaches to the problems of perception, planning, grasp selection, and control. Shape estimation can already be a major challenge, and planning further requires an understanding of how an object's shape changes under manipulation. High computational cost limits what we can do with simulation under practical time constraints, and deformable object simulation is particularly susceptible to the "reality gap".
The complexity of their dynamics and the scarcity of simplifying assumptions make deformable objects a final frontier in object manipulation, as well as an excellent testbed for cutting-edge AI research. Neural Networks and Kernel-based Probabilistic Models have made it possible to model object dynamics and extract computationally manageable features from high-dimensional state spaces, making it possible to learn sophisticated manipulation skills from human demonstrations, or iteratively learn control policies under the Reinforcement Learning framework. For practical applications, the union between AI-based and model-based techniques shows potential for improving generalization capabilities and reducing training costs. These and other innovations have led to a flurry of activity and interest in the field.
From a more theoretical angle, the topic is fascinating because humans manipulate deformables with ease, on basis of hard-to-formalize intuitions. Exploring these skills may help us understand how humans cognitively handle sub-symbolic, high-dimensional problem spaces.
This topic aims to weave together some of the many threads of research in this field in a collection of work on recent advances. We naturally welcome work representing progress on the field’s open problems, as well as novel approaches with future potential. With an eye toward practical application and integration of deformable object manipulation into broader robotic intelligence, we welcome integrative work on deformable object handling in the context of scene understanding and task understanding. Finally, in consideration of the long-term goal of obtaining more human-like manipulation skills, we welcome work elucidating the cognitive processes humans apply in deformable object manipulation with potential application to robotics.
Potential article topics include, but are not limited to:
• Application of reinforcement learning and model learning in deformable object manipulation
• Learning deformable object manipulation through demonstration
• Manipulation planning for deformable objects
• Practical applications of deformable object manipulation
• Shape estimation & shape understanding
• Prediction of object shape change under manipulation
• Estimation of material properties through visual, tactile, and active perception
• Model-predictive control for deformable objects
• Grasp selection strategies for deformable objects
• Dimensionality reduction techniques for deformable object manipulation tasks
• Deformable object simulation for robotics
• Strategies for bridging the reality gap in simulation-trained systems
• Integration of deformables in scene understanding and task understanding
• Cognitive processes involved in deformable object manipulation
Manipulation of deformable objects is crucial for countless applications in household, caregiving, industrial, and surgical robotics. However, deformable objects violate many fundamental assumptions of rigid object manipulation, requiring different approaches to the problems of perception, planning, grasp selection, and control. Shape estimation can already be a major challenge, and planning further requires an understanding of how an object's shape changes under manipulation. High computational cost limits what we can do with simulation under practical time constraints, and deformable object simulation is particularly susceptible to the "reality gap".
The complexity of their dynamics and the scarcity of simplifying assumptions make deformable objects a final frontier in object manipulation, as well as an excellent testbed for cutting-edge AI research. Neural Networks and Kernel-based Probabilistic Models have made it possible to model object dynamics and extract computationally manageable features from high-dimensional state spaces, making it possible to learn sophisticated manipulation skills from human demonstrations, or iteratively learn control policies under the Reinforcement Learning framework. For practical applications, the union between AI-based and model-based techniques shows potential for improving generalization capabilities and reducing training costs. These and other innovations have led to a flurry of activity and interest in the field.
From a more theoretical angle, the topic is fascinating because humans manipulate deformables with ease, on basis of hard-to-formalize intuitions. Exploring these skills may help us understand how humans cognitively handle sub-symbolic, high-dimensional problem spaces.
This topic aims to weave together some of the many threads of research in this field in a collection of work on recent advances. We naturally welcome work representing progress on the field’s open problems, as well as novel approaches with future potential. With an eye toward practical application and integration of deformable object manipulation into broader robotic intelligence, we welcome integrative work on deformable object handling in the context of scene understanding and task understanding. Finally, in consideration of the long-term goal of obtaining more human-like manipulation skills, we welcome work elucidating the cognitive processes humans apply in deformable object manipulation with potential application to robotics.
Potential article topics include, but are not limited to:
• Application of reinforcement learning and model learning in deformable object manipulation
• Learning deformable object manipulation through demonstration
• Manipulation planning for deformable objects
• Practical applications of deformable object manipulation
• Shape estimation & shape understanding
• Prediction of object shape change under manipulation
• Estimation of material properties through visual, tactile, and active perception
• Model-predictive control for deformable objects
• Grasp selection strategies for deformable objects
• Dimensionality reduction techniques for deformable object manipulation tasks
• Deformable object simulation for robotics
• Strategies for bridging the reality gap in simulation-trained systems
• Integration of deformables in scene understanding and task understanding
• Cognitive processes involved in deformable object manipulation
