A robot trained using machine learning managed to peel a banana without crushing it, which is a milestone due to the inconveniences that these prototypes have when handling fruit.
Researchers at the University of Tokyo have developed a machine learning system that powers a robot with two arms and hands that is capable of grasping objects between two “fingers.”
For this learning, a human peeled hundreds of bananas, resulting in 811 minutes of demo data used to train the robot. The prototype divided the training process into several stages, to pick it up, grab the tip, peel it and remove the shell.
The tests showed that the robot was able to peel the fruit without damaging it in 57 percent of the tests carried out, in a time that did not exceed three minutes.
“What is really interesting in this case is that the process that a human uses has been transferred to the training of the robotic system through deep imitation learning,” explains Jonathan Aitken, a researcher at the University of Sheffield.
For their part, the scientists in charge of the experiment assured that their method is efficient in terms of data because it uses 13 hours of training information instead of hundreds or thousands of hours. “Still needs a lot of GPUs [unidades de procesamiento gráfico] expensive, but by using our structure, we can reduce the large number of calculations required.”
Scientists now want to see how the robot behaves with the most misshapen fruit. The idea is also that in the future this prototype can perform tasks that require fine motor skills.