Robotic chef masters cooking skills through video learning, innovating new recipes

A team of researchers from the University of Cambridge has achieved a significant breakthrough in the field of culinary robotics. They have successfully trained a robotic “chef” to learn and recreate dishes by watching cooking videos. This innovative approach could revolutionize automated food production and pave the way for the widespread use of robot chefs.

The scientists began by equipping the robot with a cookbook containing eight simple salad recipes. They then introduced the robot to a variety of cooking videos where human chefs demonstrated the preparation of these recipes. Through careful observation and analysis, the robot was able to identify the recipe being followed in each video and replicate it with precision.

Furthermore, the videos served as a valuable source of learning for the robotic chef. Over time, the robot incrementally expanded its culinary repertoire by incorporating elements from the videos into its cookbook. Astonishingly, by the end of the experiment, the robot even devised a unique ninth recipe entirely on its own.

This breakthrough research, published in the journal IEEE Access, highlights the immense potential of utilizing video content as a rich data source for automated food production. It offers a promising solution to the challenges associated with teaching robots to cook and could potentially facilitate the easier and more cost-effective deployment of robot chefs.

While the concept of robotic chefs has long been a staple of science fiction, the reality of teaching a robot to cook has proven to be a complex problem. Although a few companies have developed prototype robot chefs, these machines are not yet commercially available and still fall short of their human counterparts in terms of culinary skill.

By leveraging the power of video-based learning, the researchers at the University of Cambridge have taken a significant step forward in bridging this gap. Their work opens up exciting possibilities for the future of automated cooking and paves the way for advancements in the capabilities of robot chefs.

Human demonstrating one of eight pre-programmed recipes for a robot ‘chef’ using a neural network. Credit: University of Cambridge

Grzegorz Sochacki, the first author of the paper and a Ph.D. candidate in Professor Fumiya Iida’s Bio-Inspired Robotics Laboratory at the University of Cambridge, explained the team’s approach to training the robot chef. They created eight straightforward salad recipes and recorded themselves preparing them. To train the robot, they utilized a publicly available neural network that had already been programmed to recognize various objects, including the fruits and vegetables featured in the salad recipes (such as broccoli, carrot, apple, banana, and orange).

Employing computer vision techniques, the robot analyzed each video frame to identify different objects and attributes, such as the ingredients, a knife, and even the human demonstrator’s arms, hands, and face. Both the recipes and videos were converted into vectors, and the robot conducted mathematical operations on these vectors to determine the similarity between a demonstration and a particular vector.

By accurately identifying the ingredients and the actions performed by the human chef, the robot could deduce which recipe was being prepared. For instance, if the human demonstrator held a knife in one hand and a carrot in the other, the robot could infer that the carrot needed to be chopped up.

This method enabled the robot to learn in an incremental manner, similar to how humans learn, by understanding the ingredients and their interactions within a dish. By analyzing the visual cues and actions in the videos, the robot could identify and replicate the recipes accurately.

Example of robot recreating a ‘salad’ after watching a human make it. Credit: University of Cambridge

The robot chef showcased impressive capabilities during the experiment. Out of the 16 videos it watched, the robot correctly identified the corresponding recipe with a remarkable accuracy rate of 93%, even though it only detected 83% of the human chef’s actions. Additionally, the robot displayed an understanding of slight variations in recipes, such as doubling the portion size or accounting for human errors, recognizing them as variations rather than entirely new recipes. It successfully integrated a new, ninth salad recipe into its cookbook, replicating it flawlessly.

Grzegorz Sochacki expressed his amazement at the robot’s ability to discern subtle nuances. While the salad recipes themselves were not complex, primarily consisting of chopped fruits and vegetables, the robot proved adept at recognizing equivalences. For instance, it correctly identified that two chopped apples and two chopped carrots were the same recipe as three chopped apples and three chopped carrots.

It is worth noting that the videos used to train the robot differed from typical food videos found on social media platforms, which often feature fast cuts, visual effects, and rapid transitions between the chef and the dish being prepared. The robot faced challenges in identifying objects like a carrot if the human demonstrator’s hand obscured it. To ensure accurate identification, the demonstrator had to hold up the carrot, enabling the robot to view the entire vegetable.

Sochacki clarified that the robot’s focus was not on the popular viral food videos found on social media platforms, as they are often challenging for the robot to follow. However, as robotic chefs continue to enhance their ability to identify ingredients in food videos, platforms like YouTube could become valuable resources for the robots to learn a diverse range of recipes.

Source: University of Cambridge

Leave a Reply

Your email address will not be published. Required fields are marked *