Unsupervised Learning of Action Primitives

Action representation is a key issue in imitation learning for
humanoids. With the recent finding of mirror neurons there has been a
growing interest in expressing actions as a combination meaningful
subparts called primitives. Primitives could be thought of as an
alphabet for the human actions. In this paper we observe that human
actions and objects can be seen as being intertwined: we can interpret
actions from the way the body parts are moving, but as well from how
their effect on the involved object. While human movements can look
vastly different even under minor changes in location, orientation and
scale, the use of the object can provide a strong invariant for the
detection of motion primitives. In this paper we propose an
unsupervised learning approach for action primitives that makes use
of the human movements as well as the object state changes. We group
actions according to the changes they make to the object state
space. Movements that produce the same state change in the object
state space are classified to be instances of the same action
primitive. This allows us to define action primitives as sets of
movements where the movements of each primitive are connected through
the object state change they induce.