Observation planning with on-line algorithms and GPU heuristic computation

Abstract

When making an useful description of its environment, a robot has to identify both the free space and the objects location. SLAM algorithms are used for computing the free space map and an image processing algorithm is used in order to identify the objects. Those algorithms are time consuming (the time to go to the observation location and the image processing time) and are not perfect (their outcomes are stochastic). Furthermore, the agent may have multiple target to identify at the same time, and so has to build a policy for identication. We propose a Markov Decision Process (MDP)-based approach to compute those policies. Since in our application, the policy has to be computed on-line, in a limited time, optimal algorithms are too slow for those on-line purposes with a large state space. We show how on-line approaches oer solutions to the observation planning problem, and how to eciently use those algorithms by computing an accurate admissible heuristic on a GPU.