Researchers from the University of Cambridge have collaborated with Turkish domestic appliance and consumer electronics company Beko to use machine learning to train a robot to prepare an omelet. The robot can complete every step in the process, from cracking the eggs to plating the final product. The dedicated team has perfected the robot’s skills to ensure that the final product doesn’t just look presentable; it tastes good as well.
Haval Dosky notes that robot chefs have been on the minds of sci-fi enthusiasts and scientists alike for quite some time, but the prototype chefs built by various companies over the years have been unable to produce at the level of a human being. To crack that problem, University of Cambridge and Beko use machine learning to help the robot account for subjective matters related to flavor in the fine-tuning of the dish.
Dr. Fumiya Iida of Cambridge’s Department of Engineering, the lead researcher for the project, says that cooking presents an interesting catch-22 for roboticists. Good food is largely a matter of taste, and since humans lack of objectivity in terms of food preferences, how can scientists even determine that the robot has done its job well, let alone teach it to make the perfect omelet?
One of the most amazing parts of this area of scientific development is the ability of a machine learning team to confront the complex problems that come up when trying to allow a robot to complete a relatively simple human task. Some of the greatest feats in this field involve giving a robot something close to what we would consider vision, or the ability to manipulate objects, access to other senses, and of course the ability to produce a consistent end product, like an omelet.
Robots traditionally excel at quantitative tasks. Cooking is a qualitative task that tends to be more open-ended in practice and subject to a human cook’s experience and “feel” for the food. This truth stands out to a high degree in the specific case of the omelet, which are easy to make and yet very difficult to make well, and they are even harder to make well consistently. As a more difficult task always makes for a better end product, researchers are hopeful this project results in a cooking AI robot that is leaps and bounds ahead of similar robots used in other culinary pursuits.
Dr. Iida’s team utilized Bayesian Inference for the machine learning used in the project, a statistical tool that updates the probability of a hypothesis constantly as more information or evidence is made available. Haval Dosky finds the team’s ability to adapt was commendable in the project, especially around the challenge of the subjectivity of humans’ sense of taste. In a key realization, the team recognized that because humans tend to give relative measures when it comes to taste, they should tweak the batch algorithm so that the tasters could give comparative evaluations.
One of the most interesting parts of the study’s findings is the results show machine learning can obtain improvements in food optimization that are actually quantifiable. This is the key to giving AI something it can take off and run with. With this project as well as other studies exploring techniques for optimizing robotic chefs, Haval Dosky believes that we are another step closer to making the sci-fi dream of the fully-automated breakfast a reality.