Robot Learning by Imitation

Humans prefer to demonstrate movements, rather than describe linguistically. They offer a visual model by demonstration, which can be used for learning by imitation. Thus, from the learning motor perspective of skills humanoid robots have an unmatched advantage on other robots. They have a body shape that allows them to imitate humans.

The most straightforward way to force a robot to imitate of human movements is to take control completely over its actions, moving it by tele-manipulation. For instance, NASA JSC (Johnson Space Center) has a full immersion tele-presence testbed, which allows operators to be virtually immersed in the environment where a two-arm dexterous anthropomorphic robot operates. The operator headset allows the human to see trough eyes of robot, the cameras mounted on the robot head, and special gloves allow the operator to move the robot arms, while also getting force feedback.

It is possible for extension forcing overall body imitation if the body is covered with appropriately placed sensors. Capturing and imitation of elements of human movement is of great interest not only to robotics engineers but also to computer assisted movie and game makers. For such users, Sarcos has developed the SenSuit, which enables real time tele-operator control of computer generated icons and robotic figures.

Early references of the use of imitation for anthropomorphic or humanoid robots include. The topic was not much in the attention of researchers, partly because the whole field of humanoids was largely non existent outside Japan. The only notable was the COG project, which in its early phase focused more on using ideas related to the subsumption architecture and behavioral robotic, that has changes a few years later to emphasize the interaction with human users. Learning by watching was a precursor of the learning by imitation. The focus was on task learning and not how to move. More currently, imitation learning has received a much larger attention, the role of learning by imitation for humanoid robots being well argued in the work of Schaal and Vijayakumar and also Mataric.