Feature Representations for Enhanced Multi-Agent Navigation Strategies

ABSTRACT: Systems that perform autonomous mapping build their representations by extracting and matching feature descriptors from the environment. However, these systems are limited in their ability to adapt to significant changes in operating conditions, such as novel scenes or unexpected objects. While researchers have created many different feature descriptors, there is only limited guidance as to which features work best in different environments, how to balance computation complexity with the level of detail stored, and how to share information between multiple agents that are working jointly. In this Phase 1 effort, we will develop a testbed for systematic testing of descriptors in a range of operating conditions. Based on the results of tests on a wide set of features, we will develop a novel feature using a statistical hierarchical framework that combines multiple features to allow it to adapt to the environment, while also having a compact encoding for efficient storage and transmittal. We will evaluate the descriptor with multi-agent mapping applications, and, in Phase 2, we will build a number of prototype agents to demonstrate the enhanced multi-agent mapping capabilities in realistic scenarios. BENEFIT: This effort will create both a systematic evaluation of feature descriptors in a range of environments, and also a new feature descriptor that is designed both for adaptability to different environments, and compactness for sharing between agents. The new descriptor will have application to mapping, object detection, target recognition, and other autonomous recognition algorithms, from a variety of sensors. It will also have commercial application to emergency search and rescue, remote exploration, and other robotic applications.