Dynamic Control of Robot Perception Using Stochastic Spatial Models

Abstract

Robot perception has traditionally been addressed as a passive and incidental activity, rather than an active and task-directed activity. Consequently, although sensor systems are essential to provide the information required by the decision-making and actuation components of a robot system, no explicit planning and control of the sensory activities of the robot is performed. This has lead to the development of sensor modules that are either excessively specialized, or inefficient and unfocused in their informational output. In this paper, we develop strategies for the dynamic control of robot perception, using stochastic sensor and spatial models to explicitly plan and control the sensory activities of an autonomous mobile robot, and to dynamically servo the robot and its sensors to acquire the information necessary for successful execution of robot tasks. We discuss the explicit characterization of robot task-specific information requirements, the use of information-theoretic measures to model the extent, accuracy and complexity of the robot's world model, and the representation of inferences about the robot's environment using the Inference Grid, a multi-property tesselated random field model. We describe the use of stochastic sensor models to determine the utility of sensory actions, and to compute the loci of observation of relevant information. These models allow the development of various perception control strategies, including attention control and focussing, perceptual responsiveness to varying spatial complexity, and control of multi-goal perceptual activities. We illustrate these methodologies using an autonomous multi-sensor mobile robot, and show the application of dynamic perception strategies to active exploration and multi-objective motion planning.