Underwater Rat-SLAM with Memristive Spiking Neural Networks

• Autores: Pirozzo, Bernardo M.; De Paula, Mariano; Villar, Sebastián; Acosta, Gerardo G.
• Año de publicación: 2024
• Idioma: inglés
• Tipo de recurso: documento de conferencia
• Materia: Ciencias de la Computación e Información; RAT-SLAM; Memristors; Neuromorphic Computing; Neuroscience; Autonomous Underwater Vehicles
• Institución: Comisión de Investigaciones Científicas de la Provincia de Buenos Aires

Autonomous Underwater Vehicles (AUVs) are suitable platforms for a wide variety of applications in oceanic environments. However, to successfully carry out their tasks, they require a navigation system that can estimate their location with bounded error. Localization in underwater environments is a non-trivial problem with various proposed solutions, such as the use of buoys, computer vision, acoustic beacons, among others. Nevertheless, all these solutions require expensive sensors and hardware to keep positioning errors below a threshold. For these reasons, the use of bioinspired systems with cognitive capabilities is attractive and valuable for addressing this type of problem, as they allow for learning to fuse sensors and reset growing positioning errors through loop closure. To take advantage of these features in underwater environments, we present in this work a method for solving AUV localization using Rat-SLAM. Our proposal has two central modules, the Pose Cells (PC) network, implemented with Spiking Neural Networks (SNN), which allow us to have a form of information processing closer to the biological one and with the possibility of hardware implementation to have on-board Neuromorphic Computing (NC). Each spiking neuron in the SNN is modelled using the fourth element of electronics, the memristor. With this building block, a network of Pose cells estimates the 3D positioning coordinates (x, y, z) from linear velocities, while other network of Pose Cells estimates the heading (yaw) for each position in the plane, from the x and y components of the linear velocities and the z components of the angular velocity. The linear velocities are provided by a DVL. A monocular grayscale camera provides input to Local View Cells (LVC) network to perform loop closure when a known scene is seen again. The information from the LVC, together with that from the DVL and the activity in the PC-MLIF, is used to build the experience map, which will contain all the localization estimates. To evaluate the performance of this proposal, we carried out different trajectories using a Rexrov AUV in the Gazebo simulation environmet, which is part of the Robot Operating System (ROS). From these results, we obtained some conclusions and determined the future work of our research.