DiTer++: Diverse Terrain and Multi-modal Dataset for Multi-Robot SLAM in Multi-session Environments

TL;DR

This work tackles robust SLAM benchmarking in GPS-denied, diverse campus-like environments by introducing DiTer++, a multi-robot, multi-session, multi-modal dataset collected with legged robots across day and night. It provides survey-grade ground-truth via a Leica RTC360-based prior map and a scan-to-map localization pipeline, with dynamic-object removal to maximize accuracy. The dataset includes sequences LAWN, PARK, and FOREST and offers odometry and place recognition benchmarks using LiDAR-based methods, highlighting challenges from dynamic motion and monocular scale ambiguity. The work enables evaluation of cross-robot collaboration and long-term place recognition in complex terrains, with promising potential for campus-scale SLAM research. Future directions include expanding to broader multi-robot scenarios and DiTer# campus-scale data acquisitions with full survey-grade maps.

Abstract

We encounter large-scale environments where both structured and unstructured spaces coexist, such as on campuses. In this environment, lighting conditions and dynamic objects change constantly. To tackle the challenges of large-scale mapping under such conditions, we introduce DiTer++, a diverse terrain and multi-modal dataset designed for multi-robot SLAM in multi-session environments. According to our datasets' scenarios, Agent-A and Agent-B scan the area designated for efficient large-scale mapping day and night, respectively. Also, we utilize legged robots for terrain-agnostic traversing. To generate the ground-truth of each robot, we first build the survey-grade prior map. Then, we remove the dynamic objects and outliers from the prior map and extract the trajectory through scan-to-map matching. Our dataset and supplement materials are available at https://sites.google.com/view/diter-plusplus/.

Figures (10)

• Figure 1: Obtained survey-grade map for each sequence. PARK, LAWN, FOREST maps are provided with accurate photo-metric representation, global intensity, and geometric scale obtained from survey device. For versatility in evaluation step, each survey-grade prior map are given its manual origin.
• Figure 2: Illustration of robots with perceptual sensors and different ranging sensor setups. Agent-A utilizes RGB-D for terrain information. Agent-B utilized built-in LiDAR sensor for geometric terrain information.
• Figure 3: LAWN and PARK sequences are recorded by our multi-robot systems. Agent-A (blue) and Agent-B (yellow) traverse different regions for efficient mapping. Each sequence contains partial overlap in trajectory (red box).
• Figure 4: Tightly coupled factor-graph structure for our map-localization and ground-truth generation. Given factors are tightly coupled by scan-to-map registration factors in the optimization window.
• Figure 5: Original survey-grade map contains dynamic objects and outliers caused by pedestrians, vehicles, sun glare, and reflected points captured during the survey. To obtain an accurate trajectory, dynamic object and outlier removal were conducted.