Modeling and Solving Human-Robot Collaborative Tasks Using POMDPs

Robots have helped in the automation of repetitive tasks. However, they have limited roles as co-workers or assistants to humans because of their limited sensing and perception abilities, which in turn leads to limited inference capabilities. To help with a task a robot can learn to infer its human user’s state in a process, which is hidden information. This inference is multimodal and over a large set of observations. Previous dialogue modeling work [22] has framed collaborative tasks, but they do not consider physical state-action spaces. In this work we first model a joint task between a human and a robot as a POMDP [7] with turn taking and common goals. Next we present a POMDP solver capable of handling large state spaces and an infinite number of observations to perform multimodal inference. We compare this POMDP solver with other state of the art POMDP solvers. Further we used this solver for the collaborative task of changing a child’s diaper on a robotic platform.