The disproportionate number of food processing robots we’ve seen has focused on pizza. I have long said that these are the two main reasons. First: People love pizza. We eat it a lot. Americans eat three billion pizzas a year alone. Second: It is relatively easy to build. The dough provides a very simple platform to which ingredients are added.
I say “fair” here because there are still problems. When it comes to building robots that can perform a wide range of options, nothing is really easy. Here it is the dough that presents the problem. Turning a soft and flexible ball of dough into a pizza roll is one of those things that people have figured out how to do effectively, but it’s still hard for robot workers.
A team of researchers at MIT, CMU and UC San Diego tried to create what they considered a “complex dough manipulation”. The system is divided into a two-step process in which the robot must first determine the task and then perform it using a tool such as a rolling pin. The DiffSkill system involves robots learning complex tasks in simulations.
The “teacher” algorithm solves each step of the robot to complete the task. It then learns a “student” machine learning model that provides abstract ideas about when and how to perform each skill it needs during a task, such as using a pin. With this knowledge, the system is based on how to perform the skills to complete the entire task.
Researchers say the system outperformed those trained in the traditional reinforcement learning model. “Our framework provides a new way for robots to acquire new skills. Then these skills can be chained up to solve more complex tasks beyond the capabilities of previous robot systems,” said Yunju Li, a CSAIL graduate student.
It’s robotic, so the whole pizza thing is actually a primary challenge for a system designed to solve more tasks. The company points to the emergence of robotics for the elderly as a potential future use.