Long-Range Markerless Pose Estimation for Planetary Multi-Robot SLAM

Abstract

Simultaneous Localization and Mapping (SLAM) is a key software component of rover systems to enable long-term autonomy in previously unknown settings. Multi-robot SLAM approaches, in contrast to single-agent solutions, rely on perceptual information from multiple networked robots, to produce a consistent representation of the environment in a shared and collaborative manner, with beneficial implications regarding accuracy, time efficiency, and redundancy. A key element of multi-robot SLAM is the capability of all robots in a team to observe each other with their respective perceptual inputs, and com- pute their relative poses, to join and better constrain their measurements in a global repre- sentation. Traditional means for this task rely on visual markers (e.g., AprilTags) mounted on the robots, which have limitations regarding detection range and environmental condi- tions, e.g. visibility, occlusion and reflections. We propose to complement the traditional, marker-based approach with markerless pose estimation (MPE), built on a deep-learning based object detection and pose estimation pipeline. The markerless pipeline is trained on synthetic data, and tested on both synthetic and real-world data as part of a multi-robot and multi-session visual SLAM system for a team of planetary robots. Under real-world conditions, the inclusion of our markerless pipeline reduces overall localization errors by 21% in an otherwise identical system.