Human infants display a remarkable capacity to learn
collaborative behavior from a single demonstration,
and to use this knowledge to take either agent’s role in
the collaborative behavior. They are able to extract
individual’s actions in terms of their object
manipulation goals and attribute these to the
appropriate agent, forming a “bird’s eye view” of the
collaborative action.
The current research exploits these concepts to allow
the iCub humanoid to learn collaborative tasks via
single observations of human demonstration. The tasks
involve two agents performing coordinated,
collaborative sequences of simple object manipulations.
Action perception is organized around physical
properties of objects – their appearance and
disappearance. The robot has a pre-learned action
repertoire that mirrors this perceptual capability for
actions. During human demonstration our real-time
action parser extracts the sequence of actions, including
agent attribution. The human and robot then agree on
“who goes first” and the shared plan is used by the
robot to collaborate, taking the appropriate role in the
learned action plan. We present results from 2
experiments in which distinct collaborative behaviors
are learned in real-time. We argue that this approach
provides a powerful compliment to existing
programming by demonstration methods.
| 
