A Robot-Object Unified Modeling Method for Deformable Object Manipulation in Constrained Environments

Deformable object manipulation (DOM) holds significant importance in a variety of robotic applications. However, due to the absence of computationally efficient and accurate models, manipulating such objects remains a challenge. This complexity arises from the intricate laws of deformation and the high dimensionality of shape states. While prevailing solutions address DOM primarily using explicit servo-control methods in a model-free manner for task-specific local shape attainment, these methods falter when confronting more complicated tasks that demand global model-based planning. In response, we present a unified modeling method for DOM planning within constrained environments. Our approach integrates manipulating motions, object shapes, and environmental constraints into a singular physics-based deformation model, ensuring accurate computation of a unified robot-object state at each computational phase. By harnessing the alternating direction method of multipliers-based parallel numerical recipe with a learning-based sim2real parameter estimation strategy, we achieve superior computational efficiency and modeling accuracy. The detailed numerical evaluations and sim-to-real experiments show that our model outperforms the existing methods on DOM tasks with an updating rate
$>\!25$
FPS and a relative deformation error
$<\!10\%$
. Furthermore, we demonstrate the practical utility of our model in planning a global manipulation task.