Many creatures learn through observation, above all humans: Babies observe their parents and siblings. They look, imitate, and repeat what they see to learn skills and behaviors. Although the process is particularly pronounced in the first years of life, it does not stop with age; adults are also constantly learning new hand movements and movements by copying them from others.
Advertisement
Researchers at Carnegie Mellon University, in cooperation with Meta AI, are using the way babies learn and explore their surroundings to train robots. So far, most robots have been programmed for very specific tasks. In order to be an everyday helper, however, they must also be able to cope with unfamiliar tasks and new situations. The RoboAgent that is now being presented is designed to do this – and at the same time have the learning ability of a three-year-old child.
The experimental robot is equipped with 12 basic motor skills, including picking up objects, opening and closing drawers and sliding doors, opening and closing containers, and wiping and wiping. His test lab is a small kitchen equipped with various appliances and kitchen utensils. The researchers aim to get the robot to apply its basic skills to new situations and tasks.
Active and passive learning
The RoboAgent learns the new processes in two ways. Firstly, he can passively copy tasks, although this does not take place with a camera in reality, but by feeding the system with videos from the Internet. “RoboAgent uses the information contained in these videos to learn how humans interact with objects and use different skills to successfully complete tasks,” says robotics expert Mohit Sharma, who helped develop it.
Secondly, the RoboAgent can also be introduced to new tasks in a targeted manner through human input. The system can also be remotely controlled and thus learns how to deal with new situations – as if you were literally taking a small child by the hand and showing him how to stack building blocks.
Advertisement
“This is an important milestone on the way to robots that learn efficiently, are effective in new situations, and can augment their behavior over time,” said Vikash Kumar, an associate professor at Carnegie Mellon University’s Robotics Institute. This could make robots more useful in “unstructured environments” such as homes, hospitals and other public spaces.
The RoboAgent software framework should also work with robots other than the comparatively crude test system. Thus, other robotic agents could also be trained in this way. The team wants to make its trained models, the source code, the data sets collected so far and the hardware drivers available to other researchers under an open source license.
(jl)
To home page