Ground-Challenge: A Multi-sensor SLAM Dataset Focusing on Corner Cases for Ground Robots

TL;DR

Ground-Challenge provides a comprehensive multi-sensor dataset focusing on challenging corner cases for ground robot SLAM, revealing current system limitations and aiding future research.

Contribution

This paper introduces a new multi-sensor dataset with diverse corner cases for ground robot SLAM, enabling better evaluation and development of robust SLAM algorithms.

Findings

SLAM systems exhibit drift and failure on certain sequences

The dataset includes challenging scenarios like aggressive motion and severe occlusion

Multiple sensors are synchronized for comprehensive data collection

Abstract

High-quality datasets can speed up breakthroughs and reveal potential developing directions in SLAM research. To support the research on corner cases of visual SLAM systems, this paper presents Ground-Challenge: a challenging dataset comprising 36 trajectories with diverse corner cases such as aggressive motion, severe occlusion, changing illumination, few textures, pure rotation, motion blur, wheel suspension, etc. The dataset was collected by a ground robot with multiple sensors including an RGB-D camera, an inertial measurement unit (IMU), a wheel odometer and a 3D LiDAR. All of these sensors were well-calibrated and synchronized, and their data were recorded simultaneously. To evaluate the performance of cutting-edge SLAM systems, we tested them on our dataset and demonstrated that these systems are prone to drift and fail on specific sequences. We will release the full dataset and relevant materials upon paper publication to benefit the research community. For more information, visit our project website at https://github.com/sjtuyinjie/Ground-Challenge.