Deep Reinforcement Learning for the Autonomous Adaptive Behavior of Social Robots

Social assistive robots are conceived to cooperate with humans in many areas like healthcare, education, or assistance. In situations where the workforce is scarce and when these machines work with special populations like older adults or children, the behavior must be appropriate and seem natural. In this contribution, we present a Deep Reinforcement Learning model for the autonomous adaptive behavior of social robots. The model emulates some aspects of human biology by generating artificial biologically inspired functions, like sleep or entertainment, to endow robots with long-term autonomous behavior. The Deep Reinforcement Learning system overcomes classical Reinforcement Learning problems such as high dimensional state-action spaces learning which actions better suit each situation the robot is experiencing. Besides, the system aims at maintaining the robot's internal state in the best possible condition sustaining human-robot interaction. The results show that our robot Mini correctly learns how to regulate the deficits in its biological processes by selecting from six actions in a high diversity of situations that merge the state of the biological process and the external stimuli the robot perceives from the environment.