Cooperative Rao-Blackwellized Particle Filter based SLAM framework using Geometric Information and Inter-Robot Measurements

In unknown environments, multiple robots must have capabilities to sense and interpret their surroundings, and localize themselves before performing some missions such as exploring the mineral resources and rescuing people. It is usually called multi-robot simultaneous localization and mapping. To perform multi-robot SLAM more accurately, robots are required to build maps of their surroundings accurately. In addition, inter-robot measurements should be properly utilized in the SLAM process. In this dissertation, a novel Rao-Blackwellized particle filter based SLAM framework is presented using geometric information and inter-robot measurements for accurate multi-robot SLAM. For SLAM, a RaoBlackwellized particle filter (RBPF) is basically one of representative methods. It takes advantage of linear time-complexity which is linearly proportional to the number of features by factoring the full SLAM posterior into the product of a robot path posterior and landmark posteriors. Additionally, it deals with multihypothesis data association using particles with their own data association. They makes it more robust than extended Kalman filter based SLAM. The proposed SLAM framework is divided into two major parts. First, the RBPF is improved using cooperation among particles in case of single robot SLAM, which is called Relational RBPF-SLAM. Here, the framework basically follows the process of the factored solution to SLAM using the unscented Kalman filter (UFastSLAM), which is an accurate instance of RBPFSLAM. A concept of particle to particle cooperation is considered in the importance weight step and the resampling step to increase the SLAM accuracy and solve some inherent problems such as the particle depletion problem and the data association problem. The particle depletion problem is almost eliminated using the formation maintenance of particles which is controlled without any rejection or replication of particles during the resampling step. In addition, to overcome the data association problem, the posterior distribution is estimated more accurately by compensating improperly assigned weights of particles. Secondly, to reduce the accumulated robot pose errors and feature errors, inter-robot measurements are utilized in the proposed RBPF-SLAM framework. They can be measured when a rendezvous between robots occurs or robots share common features. To deal with the inter-robot measurements, a Kalman consensus filter scheme is involved in the proposed RBPF-SLAM framework, which is robust than the covariance intersection method. Several simulations and experiments show significant improvements of the proposed RBPF-SLAM framework in both the accuracy of robot poses and map quality by comparing the state of the art techniques, i.e. FastSLAM 2.0, particle swarm optimization (PSO) based FastSLAM, UFastSLAM, particle fission based UFastSLAM and PSO based UFastSLAM.