Robot with a box: push and learn (Ki photo: freepik)
Cambridge/Santa Clara – With a training program, researchers make the Massachusetts Institute of Technology (With) together with experts from the chip manufacturer Nvidia Household robots on valuable helpers. The robots serve food, clean toilets and sweep away dirt. If you make a mistake, you have to practice again. Targeted thrusts are also instructive for the machines, say the engineers.
Lay technology
In contrast to other methods for correcting robot behavior, users do not have to collect new data in the new procedure and the mechanical learning model that drives the robot’s brain.
“We cannot expect laypersons to fine -tune a neural network model. The user expects the robot to be ready for use immediately, and if this is not the case, he wants an intuitive procedure to adapt it to a task, ”said co-chief developer Felix Yanwei Wang.
What a robot has learned is not one hundred percent of what he has to do in the real world. Perhaps he was trained to take boxes out of a shelf without reversing them, but he may not be able to grab a box that is located in a different shelf.
In order to fix this defect, additional data is usually necessary that illustrate the new task. A new training follows that is expensive and time -consuming. Above all, it requires specialist knowledge in the field of machine learning, it is said.
Correct intuitive
The co-and Nvidia researchers go a different way and want to enable users to control the behavior of the robot during use if it makes a mistake. They offer him 3 options.
One of them is pushing – the user physically guides the robot arm in the desired direction. It can also point to the object that is to be moved, or demonstrate the necessary movement of the robot arm with gestures. All of this remembers the helper for the cases in which he has to perform similar tasks.
Source: www.pressetext.com
(PTE015/07.03.2025/12:30)