**AdaptiGraph: Material-Adaptive Graph-Based Neural Dynamics for Robotic Manipulation** **Authors:** Kaifeng Zhang, Baoyu Li, Kris Hauser, Yunzhu Li **Institution:** University of Illinois Urbana-Champaign, Columbia University **Abstract:** This paper introduces AdaptiGraph, a learning-based dynamics modeling approach that enables robots to predict, adapt to, and control a wide array of challenging deformable materials with unknown physical properties. AdaptiGraph leverages a graph-based neural dynamics (GBND) framework, which represents material bits as particles and employs a graph neural network (GNN) to predict particle motion. The key innovation is a unified physical property-conditioned GBND model capable of predicting the motions of diverse materials with varying physical properties without retraining. Upon encountering new materials during online deployment, AdaptiGraph utilizes a physical property optimization process for a few-shot adaptation of the model, enhancing its fit to the observed interaction data. The adapted models can precisely simulate the dynamics and predict the motion of various deformable materials, such as ropes, granular media, rigid boxes, and cloth, while adapting to different physical properties, including stiffness, granular size, and center of pressure. Experiments on real-world objects demonstrate superior prediction accuracy and task proficiency over non-material-conditioned and non-adaptive models. **Introduction:** The paper addresses the challenge of constructing accurate predictive models for deformable objects with unknown physical properties. AdaptiGraph integrates a physical property-conditioned dynamics model with online physical property estimation, enabling robots to adaptively manipulate diverse objects. The framework captures relational biases and complex motions of deformable objects using graph representations and GNNs. Unlike previous works that focus on single material types, AdaptiGraph considers a wide range of materials and variations in physical properties, allowing it to adapt to unseen objects through interaction. **Methods:** The problem formulation involves learning a dynamics model conditioned on the material type and continuous physical property variables. The model predicts how the environment will change if the robot applies a given action. The physical property estimation is achieved through an inverse optimization process, where the robot interacts with the object to estimate its physical properties. The adapted model is then used for closed-loop control in downstream manipulation tasks. **Experiments:** The framework is evaluated across four types of objects: rigid objects, granular objects, rope-like objects, and cloth-like objects. Experiments show that AdaptiGraph can accurately predict the dynamics of objects with varied physical properties and effectively estimate real-world physical properties through few-shot interactions. The adapted model performs better in model-based planning tasks, achieving lower dynamics prediction errors and higher success rates. **Conclusion:** AdaptiGraph is a flexible framework for modeling multiple materials with unknown physical properties. Future work includes extending the method to include more object materials and a more comprehensive set of physical properties.**AdaptiGraph: Material-Adaptive Graph-Based Neural Dynamics for Robotic Manipulation** **Authors:** Kaifeng Zhang, Baoyu Li, Kris Hauser, Yunzhu Li **Institution:** University of Illinois Urbana-Champaign, Columbia University **Abstract:** This paper introduces AdaptiGraph, a learning-based dynamics modeling approach that enables robots to predict, adapt to, and control a wide array of challenging deformable materials with unknown physical properties. AdaptiGraph leverages a graph-based neural dynamics (GBND) framework, which represents material bits as particles and employs a graph neural network (GNN) to predict particle motion. The key innovation is a unified physical property-conditioned GBND model capable of predicting the motions of diverse materials with varying physical properties without retraining. Upon encountering new materials during online deployment, AdaptiGraph utilizes a physical property optimization process for a few-shot adaptation of the model, enhancing its fit to the observed interaction data. The adapted models can precisely simulate the dynamics and predict the motion of various deformable materials, such as ropes, granular media, rigid boxes, and cloth, while adapting to different physical properties, including stiffness, granular size, and center of pressure. Experiments on real-world objects demonstrate superior prediction accuracy and task proficiency over non-material-conditioned and non-adaptive models. **Introduction:** The paper addresses the challenge of constructing accurate predictive models for deformable objects with unknown physical properties. AdaptiGraph integrates a physical property-conditioned dynamics model with online physical property estimation, enabling robots to adaptively manipulate diverse objects. The framework captures relational biases and complex motions of deformable objects using graph representations and GNNs. Unlike previous works that focus on single material types, AdaptiGraph considers a wide range of materials and variations in physical properties, allowing it to adapt to unseen objects through interaction. **Methods:** The problem formulation involves learning a dynamics model conditioned on the material type and continuous physical property variables. The model predicts how the environment will change if the robot applies a given action. The physical property estimation is achieved through an inverse optimization process, where the robot interacts with the object to estimate its physical properties. The adapted model is then used for closed-loop control in downstream manipulation tasks. **Experiments:** The framework is evaluated across four types of objects: rigid objects, granular objects, rope-like objects, and cloth-like objects. Experiments show that AdaptiGraph can accurately predict the dynamics of objects with varied physical properties and effectively estimate real-world physical properties through few-shot interactions. The adapted model performs better in model-based planning tasks, achieving lower dynamics prediction errors and higher success rates. **Conclusion:** AdaptiGraph is a flexible framework for modeling multiple materials with unknown physical properties. Future work includes extending the method to include more object materials and a more comprehensive set of physical properties.