This paper presents, to the best of our knowledge, the first instance of real-time human-robot interaction using motion capture (mocap) data obtained from fully wireless, on-body sensor networks. During the learning phase, data for motion such as waving of the hands, standing on a leg, performing sit-ups and squats is captured from a human strapped with the orient motion capture specks. Key features are extracted from the captured motion data using unsupervised learning algorithms. During subsequent interactions with the robot, the motion of the operator, speckled with orients, is classified and the robot selects to play the closest motion. This approach is particularly useful in situations where the robot operates a well defined vocabulary of motion, and the advantages are the real-time interaction and the rapidity (in a matter of minutes) in programming new behaviour compared to a heuristics-based approach. This paper compares the performances of three unsupervised learning algorithms: c-means, k-means and expectation maximisation (EM) for the four motion scenarios. Nine best candidates for the three learning algorithms for each of the four motion scenarios were selected in the Webots robot simulator and then transferred to the real robot. Metrics were defined for each motion scenario and their performances compared for the three learning algorithms. In all the cases the motions were able to be imitated; c-means was the best, followed closely by the k-means algorithms, and the reasons have been analysed.
|Title of host publication||Robot and Human Interactive Communication, 2009. RO-MAN 2009. The 18th IEEE International Symposium on|
|Publisher||Institute of Electrical and Electronics Engineers (IEEE)|
|Publication status||Published - 1 Mar 2009|