An Egg-Cellent Machine

Researchers from the University of Cambridge in collaboration with Beko, a domestic appliance and electronic brand, used machine learning to train a robot to account for highly subjective matters of taste.

As artificial intelligence has advanced, many commercial companies are building prototype robot chefs, although none of these robots are currently available, and they lack the precision and swiftness that their human counterparts have in terms of skill.

Cooking is a really interesting problem for roboticists, as humans can never be totally objective when it comes to food, so how do we as scientists assess whether the robot has done a good job?
— Dr Fumiya Iida, Cambridge's Department of Engineering

Training a robot to prepare and cook food is one of the most stimulating tasks, since it must deal with the complexities in robot manipulation, computer vision, sensing and human-robot interaction, and assembling a consistent finished product.

Moreover, every person has a different taste palette - cooking is a qualitative task whereas robots are generally more proficient at quantitative tasks. Special tools are needed to be developed for robots in order for food to be prepared as taste is not universal.

Many other research groups have trained robots to make pancakes, spaghetti and even pizza but these robot chefs have not been optimised for the many subjective variables involved in cooking.

 
 
An omelette is one of those dishes that is easy to make, but difficult to make well. We thought it would be an ideal test to improve the abilities of a robot chef, and optimise for taste, texture, smell and appearance.
— Iida

In partnership with Beko, Iida and his associates trained their robot chef to prepare an omelette. The work was conducted in Cambridge's Department of engineering, using a test kitchen provided by Beko plc and Symphony Group.

Here is a demonstration of the robot chef lifting up the omelette from its tray. Image Credit: Google

Here is a demonstration of the robot chef lifting up the omelette from its tray. Image Credit: Google

A statistical tool, called Bayesian Interference, is used to extract out as much information as possible from a limited amount of data, which is mandatory in avoiding over-stuffing the human tasters with omelettes.

Another challenge we faced was the subjectivity of human sense of taste – humans aren’t very good at giving absolute measure, and usually give relative ones when it comes to taste. So we needed to tweak the machine learning algorithm – the so-called batch algorithm – so but human tasters could give information based on comparative evaluations rather than sequential ones.
— Iida

Results show that machine learning can be used to acquire quantifiable improvements in food optimisation. Additionally, this approach can easily be extended to multiple robot chefs. Although, further studies are needed in order to investigate other optimisation techniques and their ability to work successfully.