Robot Control System of Interaction Manager

The robot control system can be implemented using different publicly available software packages, for instance the popular Player application or the Carmen robot navigation toolkit. It has both throughout our preliminary experiments. The aim of this component is to handle tasks such as mapping, path planning and localization. We do not discuss this component further as it is somewhat orthogonal to the main focus.

The Interaction Manager acts as the core decision making unit in the robot architecture. This module is responsible ultimately for selecting the behavior of the robot throughout the interaction with the user.

The Interaction Manger can be seen as an Input/Output device, where information about the world is received via the grammar system and the low level robot navigation system. The unit then outputs actions in the form speech and display responses, or issuing the control commands to the navigation unit. These actions are processed through the behavior manager, to extract a preset sequence of low level operations, before being sent to the respective modules (visuo tactile unit, speech synthesis, robot control system).

The aim of the Interaction Manager is to provide a robust decision making mechanism capable of handling the complexity of the environment. This is challenging target due to the high degree of noise in the environment. While the semantic grammar can help handle some of the noise, even properly transcribed speech can consist ambiguities. Accounting for this set of possible outcomes can be crucial in providing robust action selection.

The POMDP (Partially Observable Markov Decision Process) paradigm has been shown to be a powerful tool for modeling a wide range of robot related applications featuring uncertainty, including dialogue management, robot navigation, and behavior tracking, one of the advantages of the POMDP model is its ability to capture the idea of partial observability, namely that the state of the world can not be observed directly, but instead must be inferred through noisy observations.