Neural network based FastSLAM for autonomous robots in unknown environments

Map learning and self-localization based on perception of the environment's structure are fundamental capacities required for intelligent robots to realize true autonomy. Simultaneous Localization and Mapping (SLAM) is an effective technique for such robots, as it addresses the problem of incrementally building an environment map from noisy sensory data and tracking the robot's path with the built map. As a popular SLAM solution, FastSLAM suffers from limitation on error accumulation introduced by incorrect odometry model and inaccurate linearization of the SLAM nonlinear functions. To overcome the problem, a new Jacobian free neural network (NN) based FastSLAM algorithm is derived and discussed in this paper. The main contribution of the algorithm is twofold: on the one hand, the odometry error is online compensated by using a multilayer NN, and the NN is online trained during the SLAM process; on the other hand, the third-degree Cubature rule for Gaussian weighted integral, which calculates nonlinear transition density of Gaussian prior up to the 3rd order nonlinearity, is utilized to estimate the SLAM state (i.e., the robot path and environment map) and to online train the NN compensator. The performance of proposed SLAM is investigated and compared with that of popular FastSLAM2.0 in simulations and experiments. Results show that the proposed method improves the SLAM performance.