Towards Learning Human-Robot Dialogue Policies Combining Speech and Visual Beliefs

We describe an approach for multi-modal dialogue strategy learning combining two sources of uncertainty: speech and gestures. Our approach represents the state-action space of a reinforcement learning dialogue agent with relational representations for fast learning, and extends it with belief state variables for dialogue control under uncertainty. Our approach is evaluated, using simulation, on a robotic spoken dialogue system for an imitation game of arm movements. Preliminary experimental results show that the joint optimization of speech and visual beliefs results in better overall system performance than treating them in isolation.