Machine Learning for Wireless Distance Estimation Model Parameter Estimation for Breadcrumb Localization Applications

Autonomous robots commonly rely on global positioning systems (GPS) and easily recognizable, and re-identifiable, landmarks in order to achieve localization. These elements of infrastructure are not always available in the environments where autonomous robots operate. GPS is easily denied and obstructed due to its low power design. Finding and re-observing unique landmarks in an environment can also be problematic in landscapes with limited unique features. This work builds upon a system of deploy-able electronic way-points, known as BreadCrumbs [1]. Designed to provide path planning and localization information in outdoor environments through the use of GPS, these BreadCrumbs are expanded to be capable of self-localization and the ability to provide adaptive path-loss-exponent determination based on environmental placement via a Deep Deterministic Policy Gradient algorithm for use in RSSI based range determination. Thus, enabling the BreadCrumbs to act as easily re-identifiable landmarks in a variety of environments where GPS may not be accessible and natural landmarks are sparse. We present an analysis of this system in an indoor lab and outdoor forest trail setting.