RoboErgoSum Project

Self-awareness and Social awareness

Is self-awareness a necessary capability for interacting with others? The answer to this question depends on the level of interaction. In neuroscience the role of mirror neurons in interaction is well documented [1], although many questions are open on their exact role in imitation learning or perspective taking. In the field of human-robot interaction, little work has been done on applying perspective-taking mechanisms for ambiguity resolution. Trafton et al. in [2][3] proposed a system by which the robot is able to figure out which of several cones a human is referring to in different situations (visible/not visible for one of the interacting agents). Berlin et al. [4] focused on the use of visual perspective taking skills for learning from a human teacher. Visual perspective taking has been also used to aid action recognition between two robots [5]. Marin-Urias in [6] used visual perspective taking to ease clarification of referential utterances in scenarios with multiple objects.
Several theories dealing with collaboration [7][8][9] emphasize that collaborative tasks have specific requirements. Since the robot and the person share a common goal, they have to agree on the manner to realize it, they must show their commitment to the goal during execution, etc. Several robotic systems have already been built based on these theories [10][11][12][13] and they all have shown benefits of this approach. They have also shown how difficult it is to manage turn-taking between communication partners and to interleave task realization and communication in a generic way. Finally, today only few systems [14], [13][15] take humans into account at all levels.

Approach. We hypothesize that social awareness can be envisaged through the interplay between self awareness as considered in previous sections, and the capability of the system to be aware of the behaviors of other agents in its environment. We will study the robot ability to build and exploit a theory of mind associated to other agents and how it can be influenced or, conversely, can influence them. This will be based on models of others and perspective taking as a key approach to social awareness. The objective is to endow the robot with the capability to understand interaction, and consequently to be able to exploit it and to enrich its existing interaction skills or explore and learn new ones. In fact, when performing tasks in interaction with others, e.g., with humans, the robot supervision system is not only responsible for the refinement and the correct execution of the robot plan, but also for the appropriate set of communications and monitoring activities within and around task realization. It is also in charge of monitoring human commitment and activities in order to provide appropriate responses based on current context.

Another issue will consist in extending the previously developed reinforcement learning system to social feedback and verify that the model can be well applied to the human-robot interaction. The system will be tested in a social norms learning task in a context involving the interaction with several humans. Whereas the system was previously tested with abstract scalar rewards, in this scenario the robot will have to maximize a reward function determined by the amount of attention humans give to the robot. Put in the middle of a group of humans, the robot will have to choose to orient towards those that pay the highest attention to the robot. The robot shall engage interaction with this chosen humans and shall maintain interaction until humans stop paying attention to the robot.

Back to Work Packages