Physical Interactive Localization Learning

Localization is fundamental for mobile robots, especially in unconstrained outdoor environments. Earlier work showed unsupervised localization learning on landmarks to be suitable for large-scale scenes. However, this relied on hand-labeled data to train a CNN for recognizing landmarks. We propose a new approach that allows a robot to learn landmarks for localization with a human cooperatively. This approach uses pre-trained detectors of common objects for learning new landmarks in a scene, requiring only minimal human supervision. Hence, the method bootstraps the landmark learning process and removes the need to manually label large amounts of data. We present localization results using the learned landmarks in simulated and real-world outdoor environments and compare the results to models based on complete images and PoseNet. The landmark-based localization shows improved performance than the baseline methods in challenging scenarios. Our results further show that localization accuracy increases with the number of learned landmarks. The human teacher has complete control over selecting new landmarks, which allows learning unique, robustly detectable, and semantic landmarks.