Sim-to-Real RNN-Based Framework for the Precise Positioning of Autonomous Mobile Robots

This work proposes a recurrent neural network-based sim-to-real method to learn mobile robot localization using lidar data in dynamic environments. The main aim of the algorithm is to estimate a Cartesian position error relative to a saved position by means of stored lidar readings in a two-dimensional environment, using lidar data as input. To achieve this, we propose a method that first trains a model on synthetic and augmented LiDAR data to embed rigid transformations into the deep learning model and then fine-tunes the model on real positions using real-world data and external camera measures to produce training labels. This pre-training and fine-tuning approach considerably reduces the time, the computation power, and the amount of real-world data needed to have an accurate model, allowing running the fine-positioning model on the edge of autonomous mobile robots(AMRs). After optimizing the model architecture and hyperparameters, the devised model is tested in different scenarios, comparing the precise positioning capability of AMRs with that of a classical iterative closest point and advanced Monte Carlo localization.