Using AI Planning for Managing Affective States in Social Robotics

Social robotics has recently focused on developing AI agents that recognise and respond to human emotions. The use of plan-based approaches is promising, especially in domains where collecting data in advance is challenging (e.g., medical domains). However, we observe that the appropriate use of the user' affective state will vary with the particular interaction, the expected impact of the robot's behaviours on the user, and the opportunity and accuracy of affective sensing. We observe that there are different ways of modelling the user's affective state, and the appropriate choice will take into consideration the relationship between the user's affective state and the robot's behaviour. We propose alternative methods of modelling the user's affective state, and use lessons learnt from a recent project in order to discuss the relevant factors in each approach. We use simulated data in order to demonstrate the flexibility of model-based generation of interaction strategies.