Learning to interact with humans using goal-directed and habitual behaviors

International audienceIn order to improve adaptation capabilities of robots for human-robot interaction, we take inspiration from psychology and neuroscience to propose a hybrid control architecture. This architecture is based on the multiple Experts approach that is mainly used for mammal behavior modelling. We propose to couple a human-aware task planner (HATP) with a model-free reinforcement learning to allow the robot to learn behaviors relevant to solve tasks in interaction, taking advantage from the a-priori knowledge provided to the planner and the cheap decision capability of the reinforcement learning agent. We evaluate this architecture in a HRI task of cleaning a table and show that the combination of Experts (planner and reinforcement learning agent) increases the learning speed of the learning agent