Fast Tunnel Traversal for Ground Vehicles by Bearing Estimation with Neural Networks

Underground environments are challenging scenarios for autonomous robotics, as they have a set of properties that make most common self-localization and navigation methods struggle. We propose a system to allow wheeled autonomous vehicles to traverse tunnels at high speeds. Our proposal is built around a Convolutional Neural Network (CNN) trained to predict the bearing that centers the robot into the tunnel as it traverses it. This exploits the generalization capabilities of CNNs, so that the method works in environments outside of the training data, achieving speeds of up to 10 m/s.