Scientists from the University of Cambridge have successfully trained a robotic ‘chef’ to learn and replicate recipes by watching cooking videos. The team utilised a set of eight simple salad recipes to train the robot, which was able to identify and recreate the recipes after observing a human demonstrator prepare them in the videos.
The study, which was published in the journal ‘IEEE Access’, shows the potential of exploiting video information as a valuable data source for automated food production. The deployment of robot chefs becomes more accessible and cost-effective by allowing robots to learn incrementally from films.
While the concept of robotic chefs has long been a source of fascination in science fiction, the reality of cooking has offered substantial obstacles for robots. Although several businesses have produced prototype robot chefs, they are not yet on the market and fall short of matching human chefs’ culinary skills.
The Cambridge researchers wanted to train their robot similarly to how humans learn. They came up with eight salad recipes and recorded themselves making them. The robot was then taught using a publicly available neural network that could already recognise numerous objects, including the fruits and vegetables used in the salad recipes.
Using computer vision techniques, the robot analysed each frame of the films to recognise various things such as ingredients, a knife, and the activities of the human demonstration. The robot could determine the dish being produced by recognising the ingredients and actions of the human chef, even allowing for changes or faults.
Impressively, the robot achieved a 93 per cent success rate in correctly identifying the recipes from the 16 videos it watched, despite only detecting 83 per cent of the human chef’s actions.
Furthermore, the robot successfully incorporated a new, ninth salad recipe into its repertoire. The researchers emphasised the robot’s ability to perceive subtle nuances, such as understanding that different quantities of chopped apples and carrots can still constitute the same recipe. They also noted that the training videos for the robot were distinct from fast-paced, visually captivating food videos commonly found on social media platforms. As technology advances, future iterations of robot chefs could potentially learn a wide range of recipes from platforms like YouTube.
This study, which was funded in part by Beko plc – a major consumer electronics brand – and the Engineering and Physical Sciences Research Council, demonstrates the potential of using video information to train robots in the field of food preparation. It puts us one step closer to a future in which robot chefs can help in kitchens and streamline the cooking of numerous cuisines.