Active perception and scene modeling by planning with probabilistic 6D object poses

Abstract

This paper presents an approach to probabilistic active perception planning for scene modeling in cluttered and realistic environments. When dealing with complex, multi-object scenes with arbitrary object positions, the estimation of 6D poses including their expected uncertainties is essential. The scene model keeps track of the probabilistic object hypotheses over several sequencing sensing actions to represent the real object constellation. To improve detection results and to tackle occlusion problems a method for active planning is proposed which reasons about model and state transition uncertainties in continuous and high-dimensional domains. Information theoretic quality criteria are used for sequential decision making to evaluate probability distributions. The probabilistic planner is realized as a partially observable Markov decision process (POMDP). The active perception system for autonomous service robots is evaluated in experiments in a kitchen environment. In 80 test runs the efficiency and satisfactory behavior of the proposed methodology is shown in comparison to a random and a step-aside action selection strategy. The objects are selected from a large database consisting of 100 different household items.