Real World Data Collection for Navigation via Learning Over Subgoals Planning 

Abstract

We seek to enable robots to navigate through unfamiliar environments with Learning over Subgoals Planning (LSP), an algorithm that uses supervised learning to make predictions about space that the robot cannot see to inform navigation. Current LSP planning models are trained on artificially generated data, which translates into poor performance in the real world. We propose a data collection pipeline from which a high performance real-world LSP model can be trained. Our pipeline produces labeled data after a robot navigates around a building. The pipeline identifies frontiers (boundaries between familiar and unfamiliar space) and labels them and their corresponding images with whether they lead to the target. This data is utilized to train the LSP model. We aim to find that real agents navigating with LSP on real-world floor plans make reasonable predictions about what lies in unseen space.