We initiate the first method for solving the impractical goal specification problem in robotic deformable object shape servoing, by architecting a neural network that predicts the object’s goal shapes using human demonstrations.

We pioneer a novel machine learning pipeline for robotic manipulation of deformable objects in surgery and warehouses, by developing a novel neural network that predicts robot actions based on point cloud observation. My work outperforms the previous learning-based method by 240% in terms of Chamfer distance.

We engineer a state-of-the-art deep neural network capable of predicting the complete shape of tendon-driven surgical robots, outperforming the previous physics-based model by 26.5 times in terms of proximity to the ground truth shape.