Parameter Estimation for Deformable Objects in Robotic Manipulation Tasks.

We consider the problem of identifying material parameters of a deformable object, such as elastic moduli, by non-destructive robotic manipulation. We assume known geometry and mass, a reliable fixed grasp, and the ability to track the positions of a few points on the object surface. We collect a dataset of grasp pose sequences and corresponding point position sequences. We represent the object by a tetrahedral Finite Element Method (FEM) mesh and optimize the material parameters to minimize the difference between the real and predicted observations. We use a collocation-type formulation where the sequence of FEM mesh states are decision variables, and the dynamics are encoded as constraints. Sparsity patterns in the constraints make this problem tractable despite the large number of variables. Experiments show that our approach is computationally feasible and able to adequately re-identificy simulated material parameters.