Motion sequences: How robots learn to open doors – WELT

Status: 06.04.2022 | Reading time: 3 minutes

Open the door, go through: A real task for robots

When robots move in the everyday human environment, simple things can become real problems. Researchers have now succeeded in an experiment with doors. And there has also been progress in dealing with clothing.
The little robot rolls smoothly towards the door. He stops, stretches out his right arm and wraps his hand around the blade. He then pushes the door open, rolls through the opening, and uses his left arm to shove the door closed behind him.
What is an everyday process for a human – opening a door – is a challenging and complex task for a robot. To make it easier for them, Japanese researchers have tested a new method. They developed several action modules, between which the robot switches based on a probability calculation. The study by the group led by Hiroshi Ito from Waseda University in Tokyo has been published in the journal Science Robotics.

Also read

The use of robots in the immediate living environment of humans changes the nature of the necessary instructions for the machines. "Since cities and houses are more complicated than a well-defined factory, a robot must react autonomously to changes in the environment," the researchers write. Even opening a door becomes a challenge that you as a person do not initially expect.
This requires different movement sequences – depending on whether it is a sliding door or a door that opens outwards or inwards, for example. The shape of the handle, its position and color are also challenges when choosing the option for action.

Training sessions by remote control

Certain actions can be taught to robots with artificial neural networks through reinforcement learning. However, this requires a great deal of training. To reduce the effort, Ito and his team developed modules for so-called deep predictive learning.

Robot dog patrolling the ruins of Pompeii

A four-legged robot in the ruins of Pompeii: "Spot" is intended to provide more security and protect the facility. Boston Dynamics' robo-dog isn't alone: it works alongside a flying laser scanner.
The method works so well that the robot is 96 percent able to open the door when approaching from different positions. It becomes difficult when the lighting changes and the image comparison does not lead to good results. For example, if the handle is white instead of dark brown, as in training, the robot will only be successful if it is close to the handle.

Also read

Nevertheless, the researchers draw a positive conclusion: "We believe that our approach of realizing complex tasks in the real world with little effort for design and movement exercises could be applied to tasks other than opening doors and to a wide range of areas. "
In the same issue of Science Robotics, Fan Zhang and Yiannis Demiris from Imperial College London (UK) present a robot capable of dressing a human – at least a hospital gown. This could help patients who can no longer move their arms. The robot can not only appropriately grab the hospital gown hanging on a hanger, but also slip the sleeves over the immobilized arms. The robot managed to get dressed on a human-sized doll in more than 90 percent of the tests.

Related Posts

Leave a Reply

%d bloggers like this: