Machine learning allows a robot to properly peel a banana

A machine-learning-trained robot that mimics gestures produced by a demonstrator human is able to peel a banana without crushing it, or even mashing it.

Handling soft fruit is challenging for robots, who lack the dexterity and subtle touch needed to handle items without destroying them. Furthermore, the irregular shapes – which can vary significantly even between fruits of the same type – can also confuse the computer vision algorithms that often act as the brains of these robots.


Heecheol Kim of the University of Tokyo led a team of researchers that developed a machine learning system by imitation in a robot with two arms and hands with pincer-shaped “fingers”.

First, a person operating the robot peeled hundreds of bananas, creating 811 minutes of demo data to train the machine to do on its own. The task was divided into nine steps, from holding the banana to picking it up from the table with one hand, picking up the end with the other, peeling it and handling the fruit to remove the rest of the skin.

For large movements, which do not pose a risk of damaging the banana, the machine learning model maps a trajectory, mimicking what a human does in a natural way. However, when arms are needed to manipulate the banana precisely, the system switches to a reactive approach, in which it responds to unexpected changes in its environment.

In tests, the robot was able to successfully peel a banana 57% of the time, in an action that takes less than 3 minutes. “What’s really interesting in this case is that the process that a human uses was transferred to training the robotic system through deep imitation learning,” says the co-author of the study, which was published on the preprint server arxiv. org, Jonathan Aitken, University of Sheffield, UK.

Kim says his approach uses 13 hours of training data instead of hundreds or thousands of hours than other similar experiments, which makes it far more efficient. “Still requires a lot of GPUs [unidades de processamento gráfico]but using our framework, we can reduce the large amount of computation [necessária]” he says.

Aitken would like to see how the robot handles more deformed fruit. “But with more precise motor control, it can work even better,” he says. According to the Newscientist website, the technology won’t just be used for bananas, however: the goal is a more general feeding system that can handle tasks that don’t require final motor skills.

Have you watched our videos on YouTube? Subscribe to our channel!

Related Posts