A robot capable of peeling bananas without reducing them to mush

0
17
a-robot-capable-of-peeling-bananas-without-reducing-them-to-mush

You may also like this partner content (after the ad)


It doesn’t look like that, but peeling a banana isn’t that simple: especially for a robot. By developing a specific “machine learning” system, researchers at the University of Tokyo have managed to perform this delicate task with two robotic arms.

We humans can often be considered virtuosos when it comes to dexterity compared to robots. Indeed, we do not really need to think in order to intuitively adapt to all kinds of situations. For example, handling delicate fruit requires a lot more skill than it looks. You have to take into account the particular shape of each fruit, its texture, its size, adjust your movements according to the deformation applied to the fruit…

It is therefore because of the challenge of peeling a banana that scientists from the University of Tokyo have made it a personal matter. They have developed an automatic learning system, or “machine learning”, specially designed to allow a robot to tackle this task successfully. The robot used for this intensive training has two arms, each with two “fingers”.

Taking on human skill, the researchers sought to create a system based on the imitation of our gestures: “ Learning by deep imitation, which aims to train the behavior of the robot using of human-generated demonstration data with deep neural networks, is promising because it can transfer implicit human knowledge of dexterous manipulation into a robotic system without a predefined object knowledge manipulation rule”, explain the researchers in their study. The results are available on the server arXiv, awaiting the peer validation.

Deep neural network learning systems, intensely explored for all kinds of uses at the moment, are inspired by the brain functioning. The idea is to “feed” them with a lot of data to allow an “artificial intelligence” to learn. Here is how the researchers proceeded in this specific case.

First, a human manipulated the machine himself to peel bananas. 811 minutes of data were thus recorded to serve as a learning base for the robot. It is on this data that the latter then had to base himself to carry out the task “by himself”.

Eleven steps to peel a banana

Banana peeling has been broken down into eleven stages of manipulation: grabbing the banana, lifting it, grabbing the stem, pulling… The particularity of the learning system developed by the scientists lies in its duality. For large, “banana safe” movements, the software performs a simple trajectory calculation. If, on the contrary, it is a delicate and more random step, the robot adopts a “reactive” approach, to better adapt to the situation.

The famous robot is not yet a model banana peeler, but the scientists have nevertheless reached 57% of success. He is able to complete his task in three minutes. “ What is really interesting in this case is that the process that a human uses has been carried over into the training of the robotic system thanks to the ‘deep imitation learning‘, explains to the New Scientist Jonathan Aitken from the University of Sheffield, UK.

doubt, the ambition of the research team is not limited to bananas. The idea is ultimately to be able to create a system capable of performing delicate tasks in terms of dexterity in a much more general way, adapting to all kinds of situations.

Video of the robot in action:

Source: arXiv