On the Application of Classical Planning to Real Social Robotic Tasks

Automated Planning is now a mature area offering several techniques and search heuristics extremely useful to solve problems in realistic domains. However, its application to real and dynamic environments as Social Robotics requires much work focused, not only in the efficiency of the planners, but also in tractable task modeling and efficient execution and monitoring of the plan into the robotic control architecture. This paper identifies the main issues that must be taken into account while using classical Automated Planning for the control of a social robot and contributes some practical solutions to overcome such inherent difficulties. Some of them are the discrimination between predicates for internal control and external sensing, the concept of predicted nominal behavior with corrective actions or plans, the continuous monitoring of the plan execution and the handling of action interruptions. This manuscript highlights the dependencies between all the design and deployment activities involved: task modeling, plan generation, and action execution and monitoring. A task of Comprehensive Geriatric Assessment (CGA) is used as an illustrative example that can be easily generalized to any other interactive task.