SculptBot: Pre-Trained Models for 3D Deformable Object Manipulation

Deformable object manipulation presents a unique set of challenges in robotic manipulation by exhibiting high degrees of freedom and severe self-occlusion. State representation for materials that exhibit plastic behavior, like modeling clay or bread dough, is also difficult because they permanently deform under stress and are constantly changing shape. In this work, we investigate each of these challenges using the task of robotic sculpting with a parallel gripper. We propose a system that uses point clouds as the state representation and leverages pre-trained point cloud reconstruction Transformer to learn a latent dynamics model to predict material deformations given a grasp action. We design a novel action sampling algorithm that reasons about geometrical differences between point clouds to further improve the efficiency of model-based planners. All data and experiments are conducted entirely in the real world. Our experiments show the proposed system is able to successfully capture the dynamics of clay, and is able to create a variety of simple shapes.

The entire dynamics prediction pipeline. We first use farthest point sampling and k-nearest neighbors to cluster the original point cloud into 64 clusters. These 64 clusters become a much smaller and sparser point cloud. We then use the pre-trained dVAE from Point-BERT to tokenize each cluster. The centroid point cloud is passed through a simple physics-based dynamics approximator to predict the next state centroid point cloud given the grasp action.  This predicted next state centroid point cloud is passed to the point token predictor dynamics model along with the state centroid tokenization and the grasp action. The point token predictor predicts the tokens for each next state centroid, which represent the geometrical structure of the points within that region of the cloud. These predicted tokens along with the predicted centroid point cloud are then passed through the dVAE decoder to reconstruct the full dense predicted next state point cloud.

The full preprocessing pipeline for the point clouds. a) The original point cloud of the scene. b) the scene after position-based cropping to eliminate the elevated stage. c) The point cloud after color-based thresholding. d) The point cloud after removing statistical outliers and adding in a base plane. e) The point cloud downsampled to 2048 points.