CLUTCH: Developing an Interaction Model to Pick Food More Accurately

Chef Robotics developed CLUTCH, a neural network that ingests an ingredient point cloud and target weight to predict robotic utensil pick depth directly from data, replacing the geometric VPD (volumetric pick depth) model. In tests on randomly varying target weights such as 50g servings of mashed potatoes, CLUTCH achieved roughly a 2x reduction in the standard deviation of z diff and improved picks within 2mm tolerances, though it requires training data for each ingredient-utensil combination.