Experience-based Learning and Adaptive Behavior for Improved Situation-Awareness of a Social Robot

Abstract When robots work in close contact with their human counterparts, they have to perceive their surroundings adequately. Hence have adequate situation-awareness. In such cases, an encounter will include the robot itself, the surrounding and the humans. Existing frameworks to interpret such situations, considered a limited number of observable human cues and inputs from the surrounding so far. Therefore only a few aspects of a situation were effectively perceived by these mechanisms. In our work, we trained a Deep Q-Network (DQN) which takes a number of attributes in a situation as input before a robot makes decisions regarding an interaction in a domestic environment. Large number of situations were simulated and trained, so that adaptive responses were generated through the model for different individuals a robot will encounter. Hence an experience-based, adaptive cognitive framework for robots has been developed for initiating an interaction with their humans. This creates situation-awareness in a robot and enables user-centered perception, demolishing the requirement of the human to adopt a restricted behavior to accommodate the robot’s limited perceptive skills. Human acceptance towards the system has been evaluated through a set of experiments conducted over a period of time. Critical observations during the experiments have been highlighted.