POMDP-based Planning for Visual Processing Management on a Robot

Recent progress in sensor technology [10, 24], and the use of state of the art algorithms to process the input from a variety of sensors, has resulted in the deployment of mobile robots in several specific applications [2, 17, 22]. A key requirement for the widespread deployment of mobile robots is the ability to autonomously tailor the sensory processing to the task at hand. Our work represents a significant effort towards such general-purpose processing of visual input. We pose visual processing management as an instance of probabilistic sequential decision making, and specifically as a Partially Observable Markov Decision Process (POMDP). Our prior work introduced a hierarchical POMDP decomposition that enables a robot to plan a sequence of visual operators that reliably and efficiently analyze the state of the world represented by salient regions-of-interest (ROIs) in input images [20]. Here, we significantly enhance the capabilities of the existing system by: (a) extending our POMDP framework to autonomously adapt to a change in state space dimensions, thereby enabling the robot to effectively process partially overlapping objects in the image; and (b) enabling the robot to autonomously trade-off planning speed and plan quality, by theoretically and empirically evaluating the estimation errors involved in policy caching. All algorithms are implemented and tested on a physical robot platform. We show that the hierarchical planner performs significantly better than a modern planner that has been applied successfully to human-robot interaction domains [1].