A new approach to simultaneous localization and map building with learning: NeoSLAM (Neuro-Evolutionary Optimizing)

This paper addresses a novel approach to the solution of the simultaneous localization and mapping (SLAM) problem bared on a neuro evolutionary optimization (NeoSLAM) method. The proposed algorithm first casts SLAM as a global optimization problem using the cost function which represents the quality of robot pose trajectory and the feature positions in world coordinate frame. In our algorithm, the neural network trained to estimate the pose difference of the two consecutive positions accurately from the corresponding sensor data and the previous pose difference. The cost function is formulated as the importance of the full SLAM assumptions of EKF. Evolutionary programming (EP) is used to evolve the neural model that is most consistent with the actual data measurement. Prediction and correction is simultaneously performed in our neural model that combines both the motion model and sensor model. By way of learning and evolution, our algorithm does not need prior assumption on the motion and sensor models, and therefore shows a robust performance regardless of the actual noise type. Further, our method can generate an accurate map even without the data association step, paving the way to deal with practical applications. Both the simulation and real experimental results conducted made various environments and noise/sensor types demonstrate that NeoSLAM ensures a consistently robust and accurate performance.