A model-based description of environment interaction for mobile robots

We consider a mobile robot that attempts to accomplish a task by reaching a given goal, and interacts with its environment through a finite set of actions and observations. The interaction between robot and environment is modelled by Partially Observable Markov Decision Processes (POMDP). The robot takes its decisions in presence of uncertainty about the current state, by maximizing its reward gained during interactions with the environment. It is able to self-locate into the environment by collecting actions and perception histories during the navigation. To make the state estimation more reliable, we introduce an additional information in the model without adding new states and without discretizing the considered measures. Thus, we associate to the state transition probabilities also a continuous metric given through the mean and the variance of some significant sensor measurements suitable to be kept under continuous form, such as odometric measurements, showing that also such unreliable data can supply a great deal of information to the robot. The overall control system of the robot is structured as a two-levels layered architecture, where the low level implements several collision avoidance algorithms, while the upper level takes care of the navigation problem. In this paper, we concentrate on how to use POMDP models at the upper level