Automatic calibration of a biologically inspired neural network for robot SLAM

Neural networks have long been a promising model for creating high performance robotic systems, from robot navigation and SLAM to modern deep learning techniques for tasks like manipulation. Traditional neural network systems typically relied heavily on a large number of hand-tuned parameters, while many modern implementations perform end-to-end learning, often with extreme data and computational requirements. Past work has focused on achieving high performance in real world environments, but with extensive hand tuning. In this paper, we instead present a new framework for automatically calibrating and optimising the performance of a biologically inspired neural network SLAM system. This framework combines a preset network structure with learning procedures. We use simulations with realistic noise to demonstrate the system's ability to learn the basic components of SLAM: odometry integration, landmark learning and landmark-driven relocalisation. We also show the framework is able to calibrate a large range of network sizes, allowing rapid development and deployment of a bio-inspired SLAM system. Our work serves as a bridging contribution between traditional hand-crafted neural networks and modern end-to-end learning approaches.