M2UD: A Multi-model, Multi-scenario, Uneven-terrain Dataset for Ground Robot with Localization and Mapping Evaluation

TL;DR

M2UD addresses the lack of challenging, uneven-terrain SLAM data for ground robots by offering a multi-modal, multi-scenario dataset collected from two specialized robots. It couples 58 sequences across 12 categories (>50 km) with RTK-smoothed localization ground truth, high-precision ground-truth maps, and a development-kit ecosystem, plus a novel efficiency-aware metric $EA$-$Drift$ for localization evaluation. The work benchmarks turnkey SLAM methods across LiDAR, visual, and fused approaches, revealing strengths of LiDAR-inertial systems under degraded conditions and highlighting persistent Z-axis drift, while also enabling robust mapping evaluation via $AC$, $CD$, $AWD$, and $SCS$. Overall, M2UD enables realistic benchmarking and acceleration of SLAM research for ground robots in challenging environments, with future plans for 3D semantics and terrain analysis.

Abstract

Ground robots play a crucial role in inspection, exploration, rescue, and other applications. In recent years, advancements in LiDAR technology have made sensors more accurate, lightweight, and cost-effective. Therefore, researchers increasingly integrate sensors, for SLAM studies, providing robust technical support for ground robots and expanding their application domains. Public datasets are essential for advancing SLAM technology. However, existing datasets for ground robots are typically restricted to flat-terrain motion with 3 DOF and cover only a limited range of scenarios. Although handheld devices and UAV exhibit richer and more aggressive movements, their datasets are predominantly confined to small-scale environments due to endurance limitations. To fill these gap, we introduce M2UD, a multi-modal, multi-scenario, uneven-terrain SLAM dataset for ground robots. This dataset contains a diverse range of highly challenging environments, including cities, open fields, long corridors, and mixed scenarios. Additionally, it presents extreme weather conditions. The aggressive motion and degradation characteristics of this dataset not only pose challenges for testing and evaluating existing SLAM methods but also advance the development of more advanced SLAM algorithms. To benchmark SLAM algorithms, M2UD provides smoothed ground truth localization data obtained via RTK and introduces a novel localization evaluation metric that considers both accuracy and efficiency. Additionally, we utilize a high-precision laser scanner to acquire ground truth maps of two representative scenes, facilitating the development and evaluation of mapping algorithms. We select 12 localization sequences and 2 mapping sequences to evaluate several classical SLAM algorithms, verifying usability of the dataset. To enhance usability, the dataset is accompanied by a suite of development kits.

Figures (12)

• Figure 1: Acquisition Zone trajectories. We visualize the trajectories across 12 categories using distinct colors. (A)–(C) show satellite maps of our outdoor data recording zones; (D)–(E) present the floor plans of two indoor environments; (F)–(Q) provide snapshots of the robot and its surroundings during data collection, offering a visual representation of the recording conditions and robot status.
• Figure 2: Acquisition Platforms. (A) Six-wheeled special robot primarily used for outdoor data collection. (B) Four-wheeled mobile robot primarily used for indoor data collection.
• Figure 3: The Coordinate System for the Acquisition platform.
• Figure 4: Snapshot of the LiDAR-Camera Calibration Procedure. The first row presents a snapshot of the calibration procedure for the six-wheel robot platform, while the second row displays a snapshot of the calibration procedure for the four-wheel robot platform.
• Figure 5: The First-person Snapshots during The Collection. (A)–(H) illustrate the impact on vision; (I)–(K) depict extreme weather conditions; (L)–(O) highlight instances of uneven-terrain and aggressive motion; (P)–(T) show highly-dynamic scenes; (U)–(Y) demonstrate LiDAR degradation elements within the dataset.