Man-Robot Dialogue – MaRDi

In this project, we intend to study the human-computer interaction in situated manner. We believe that the interaction must have a physical realization, anchored in the real world to be natural and effective. In order to embody interactive systems, we propose to use humanoid robots. Robots, endowed with perceptions, but also means to act in the environment, allow the integration of a physical context in the interaction for the machine as well as for humans.
So it is a situated approach of man-machine dialogue that will prevail to the innovations brought by this project. To do so, three processes will interact. First, it will be necessary for the machine to build a context of interaction allowing making decisions about the future of the interaction. This environment will incorporate the spoken input provided by humans but also the perceptions of the environment and the robot’s proprioception. Based on this background, potentially uncertain because of errors introduced by the automatic analysis of speech, the machine will take decisions. To help humans to achieve effective interaction, the robot will produce an expressive feedback of its understanding of the context.
The reconstruction of the context will be based of course on the recognition and understanding of the spoken inputs uttered by the user but not only. The originality of this project is to anchor the interaction in the physical world, in close link with the aim of this interaction. For example, as part of a collaborative task manipulation of objects, the context will incorporate information about the configuration of the objects from the perspective of the robot but also of the humans. Thus this project will continue the research conducted in the field of spatial reasoning, and especially perspective taking. Perspective taking is a process by which a machine adopts the viewpoint of another agent (human or artificial) to reason about what can be seen by one or the other. Thus, ambiguities in the context may be raised due to the physical impossibility of certain assumptions.
Language processing methods, among others, are stochastic processes that provide hypotheses about the context of the interaction, associated with confidence levels indicating the degree of certainty on these assumptions. Decisions taken by the machine must consider the potentially generated ambiguities and keep track of them throughout the interaction. This issue is discussed in the community of man-machine dialogue, which includes Supélec and the LIA, through statistical optimization models of decision making processes but remains an important lock.
The physical attitude that robots should adopt to make the interaction more natural and effective will be among the areas of research. Also, the voice of the robot, including the lexical content, but also the intonations will be investigated. Specifically, the Acapela group company will work on speech synthesis methods based on voice transforms.