Active Learning of the Cutting of Cooking Ingredients using Simulation with Object Splitting

Food preparation is hard to automate on general-purpose robot hardware. This work focuses on acquisition of cutting manipulations for food preparation. We construct a physical simulation environment for virtually performing cutting operations on food ingredients, including routines for splitting food object models at arbitrary positions, and design a compact manipulation description format. Using the simulation environment, we generate data for training a neural network architecture (EMD-net) to predict the state transformation induced by a given cutting manipulation and evaluate the network's performance. We also perform automatic generation of cutting manipulations for obtaining a given goal state from a given initial state, and evaluate the obtained manipulations by performing them in the physical simulation environment.