Applying Unsupervised Learning and Action Selection to Robot Teleoperation

Unsupervised learning and supervised remote teleoperator control for robots may seem an unlikely combination. This paper argues that the combination holds advantages for both parties. The operator would like to “instruct” the robot without any special effort, and then be able to hand over some or all of the tasks to be performed without loss of overall supervisory control. In return, the learning algorithm receives a continuous stream of exemplar data relevant to the tasks it might later be asked to perform. We consider an unsupervised learning method, the Dynamic Expectancy Model, and a teleoperator “architecture” offering just such a serendipitous combination.