Active exploration and identification of kinematic devices

Abstract

As an important part of solving the lockbox problem, this thesis deals with the problem of identifying kinematic devices based on data generated using an Active Learning strategy. We model the belief over different device types and parameters using a discrete multinomial distribution. We discretize directions as a Geodesic sphere. This allows an isotropic distribution without being biased towards certain directions. The belief update is based on experience using a Bayes Filter. This allows to localize the correct states, even if an action fails to generate movement. Our action selection strategy aims to minimize the number of actions necessary to identify devices by considering the expected future belief. We evaluate the effectiveness of different information measures and compare them with a random strategy within a simulation. Our experiments show that the use of the MaxCE strategy creates the best results. We were able to correctly identify prismatic, revolute, and fixed devices in 3D space.