Researchers have created a robotic manipulation system that uses a two-stage learning process to enable a robot to perform complex dough-manipulation tasks. The method, detailed in a new paper, allows robots to do things like cutting and spreading dough or gathering pieces of dough from around a cutting board. “It may sound funny, but pizza-making is an extraordinary test for robots,” AI researcher Adrian Zidaritz, who was not involved in the study, told Lifewire in an email interview. “A robot views objects through a camera, so it must work with 2-dimensional images of that object while trying to cobble these images together into a 3-dimensional object. Now add to that the fact that the dough of the pizza is continuously being deformed, and the test becomes even more extraordinary.”
Spreading the Dough
For a robot, working with a deformable object like dough is difficult because the shape of dough can change in many ways, which are difficult to represent with an equation. And creating a new form out of that dough requires multiple steps and the use of different tools. It is challenging for a robot to learn a manipulation task with a long sequence of steps—where there are many possible choices—since learning often occurs through trial and error. Now, scientists at MIT, Carnegie Mellon University, and the University of California at San Diego say they have developed an improved method of teaching robots to make pizza. They created a framework for a robotic manipulation system that uses a two-stage learning process, which could enable a robot to perform complex dough-manipulation tasks over a long period. The new method involves a “teacher” algorithm that solves each step the robot must take to complete the task. Then, it trains a “student” machine-learning model that learns abstract ideas about when and how to execute each skill it needs during the lesson, like using a rolling pin. With this knowledge, the system reasons how to manage the skills to complete the entire task. “This method is closer to how we as humans plan our actions,” Yunzhu Li, a graduate student at MIT and one of the paper’s authors about the method, said in the news release about the project. “When a human does a long-horizon task, we are not writing down all the details. We have a higher-level planner that roughly tells us what the stages are and some of the intermediate goals we need to achieve along the way, and then we execute them.”
The Pi of Pie
A surprising amount of mathematics goes into creating pizza dough, Zidaritz said. The dough can be described using algebraic or parametric surfaces. “Then there is the question of choosing the formalism with which to represent the deformations, usually a set of differential equations,” he added. “Things can get difficult here because these differential equations have high computational complexity. The dough of the pizza cannot be frozen in the air while the robot works out what it may be deformed into at the next step.” Yariv Reches, the co-founder of Hyper Food Robotics, which builds robotic fast-food stores, said in an email interview that manipulating pizza dough is a tough challenge. Working with a deformable object like dough is more complex than handling a rigid one. “Static objects are being examined at the end of a series of actions, while in deformable objects, the subject matter is always changing shape and consistency—the learning then, is the annotation process mechanism needs to adapt on the fly,” he added. But recent advances in robotics could lead to great things for pizza lovers, Reches said. Food handling, assembly, cooking, preparation, and packaging often change shape while being handled by robots. “Integrating AI into food preparation means all food ingredients that experience a change of state, and need to flow through robotic dispensers, can be managed with technology,” Reches added. “For instance, pizza toppings that need an application, spreading and even corrections on the fly can be handled—or even hamburger patty and bun application and assembly.”
