Cooperative Lunar Surface Exploration using Transfer Learning with Multi-Agent Visual Teach and Repeat

From in-situ resource utilization to lunar habitation and rocket launch sites, autonomous robots will play a large role in supporting the infrastructure and development of permanent habitation on the Moon's surface. Current perception, localization, and path planning algorithms provide mobility to autonomous robots at the cost of heavy computation and data volume, expensive sensors, and communication limited by latency and volume. This paper aims to introduce an extension to traditional teach-and-repeat algorithms which use a teleoperation learning phase to inform an autonomous repetition phase. Rather than a single robot repeating a previously explored path, we explore the performance of a heterogenous robotic system sharing pathing data amongst different agents. A transfer-learning based teach-and-repeat system provides flexible and scalable mobility to a swarm of robots. Monocular image feed and basic odometry data is shared between robots for traversal. This research paves the way towards synchronous multi-agent algorithms that are lightweight and scalable for repetitive navigation tasks. The results show the performance of transfer teach-and-repeat algorithms across terrains of varied topological conditions, path characteristics (length, angle change, elevation gain, hazard level), and the characteristics of the repeat robots in a simulated environment. The goal of this work is to create a distributed system that will enable the mobility of robots designed for tasks beyond exploration such as mining and construction.