Flexible POMDP Framework for Human-Robot Cooperation in Escort Tasks

International audienceService robotics in public spaces with HRI constraints provide several challenges to the robot. In particular, humans may have an unpredictable behavior, may be distracted by the environment and may lack commitment to shared tasks. Such issues require novel decision-making abilities capable of ensuring joint intention and a robust execution of policies. We describe a novel method for ensuring cooperation between human and robot. First, we present a flexible and hierarchical framework based on POMDPs. Second, we introduce a set of cooperative states within the state-space of the POMDP. Third, for ensuring an efficient scalability, the framework partitions the overall task into independent planning modules. Lastly, for a robust execution of the POMDP policies we use Petri Net Plans, which have already been used to execute MDP policies. To this end, we describe how to convert a POMDP policy into an executable Petri Net Plan. We implement our approach and develop experiments on simulation and on a real robot in an escorting task where the robot guides a customer to the desired place in a public space