SMapper-light: SLAM Dataset

• SMapper-light is a comprehensive dataset providing tightly synchronized LiDAR, camera, and IMU recordings with sub-centimeter ground truth for accurate SLAM evaluations.
• It includes detailed sensor modalities such as high-density LiDAR scans at 10/20 Hz, multi-camera arrays capturing 2K images, and high-frequency inertial measurements to support robust mapping.
• Its open-hardware design and reproducible calibration pipeline enable rigorous benchmark testing and development of cross-modal SLAM algorithms in diverse environments.

The SMapper-light dataset is a publicly available resource for Simultaneous Localization and Mapping (SLAM) research, designed as part of the SMapper open-hardware platform. It offers tightly synchronized multimodal sensor recordings—LiDAR, multi-camera, and inertial readings—with sub-centimeter ground truth trajectories and dense 3D reconstructions. The dataset is explicitly intended to advance benchmarking, reproducibility, and cross-modal algorithm development in both indoor and outdoor SLAM scenarios.

1. Dataset Composition and Sensor Modalities

SMapper-light contains raw, temporally synchronized sensor data from three primary modalities:

• LiDAR: The Ouster OS0 64-beam LiDAR provides panoramic sensing at 10 or 20 Hz. Its coverage is 360° horizontally and 90° vertically, with a range up to 100 meters and 1024 × 64 points per rotation. This high-density geometry ensures that maps and trajectories generated from SMapper-light are detailed and robust.
• Visual Sensors: Two distinct camera subsystems are included.
  • The e-CAM200 CUOAGX four-camera array produces 2K resolution rolling shutter RGB images at 30 Hz, collectively spanning roughly 270° field-of-view with ~30° overlap among adjacent cameras.
  • An Intel RealSense D435i records synchronized global shutter RGB-D data at 30 Hz. Field-of-view is approximately 69° × 42° for RGB and 87° × 58° for depth.
• Inertial Measurement Units (IMUs): IMUs integrated into both the LiDAR and the RealSense camera collect three-axis acceleration and angular velocity data at up to 400 Hz.

All data streams are tightly aligned using a custom synchronization pipeline. Three selectable timestamping modes are supported:

• TIME_FROM_ROS (ROS system clock)
• TIME_FROM_TSC (hardware counter from NVIDIA Jetson AGX)
• TIME_FROM_PTP (Precision Time Protocol master-slave alignment for sub-millisecond synchronization)

Each data sequence is delivered as .mcap-format ROS bag files containing all synchronized modalities.

2. Ground-Truth Trajectory Acquisition and Accuracy

Ground-truth poses are generated by offline LiDAR-SLAM on the captured point clouds. This post-processing pipeline accumulates high-density geometric observations to create 6-DoF trajectories with errors under 3 cm across the dataset. The ground truth is therefore suitable for rigorous evaluation of both LiDAR- and visual-SLAM frameworks, as well as hybrid sensor fusion approaches.

In addition to trajectory data, dense 3D reconstructions are provided for each scene. These reconstructions form a reference for validating scene structure, supporting both traditional SLAM and advanced semantic mapping or cross-modal learning methods.

3. Benchmarking SLAM Algorithms

State-of-the-art SLAM algorithms were evaluated using SMapper-light, as detailed in (Soares et al., 11 Sep 2025). Results span both LiDAR and visual SLAM paradigms:

• LiDAR SLAM: Frameworks such as GLIM and S-Graphs were applied to the dataset. S-Graphs leverages semantics in addition to geometry, whereas GLIM focuses on efficient geometric mapping. Experiments demonstrated that LiDAR-based approaches consistently yield dense and complete reconstructions due to the geometric richness of the data.
• Visual SLAM: Algorithms like ORB-SLAM3 and vS-Graphs utilize the image and depth streams, generating sparser maps but capturing appearance-based scene details. These methods were robust on both indoor and outdoor sequences, benefiting from the camera system’s broad spatial coverage.

No explicit RMSE or ATE metrics are quoted; however, qualitative benchmarking emphasizes that the sub-centimeter ground truth and multimodal coverage allow rigorous and reproducible cross-system comparisons.

SLAM System	Data Modality	Notable Characteristics
GLIM	LiDAR	Dense geometry-based maps
S-Graphs	LiDAR+Semantics	Semantic+geometric reconstruction
ORB-SLAM3	Visual	Sparse, appearance-driven maps
vS-Graphs	Visual+Semantics	Semantic scene understanding

4. Open-Hardware Platform and Replicability

SMapper was conceived as an open-hardware solution and is fully documented for reproducibility and extension. Key aspects include:

• Compact construction: The full sensor+computer assembly weighs 2.5 kg (or 1.7 kg bare) and measures approximately 15 cm × 15 cm × 38.4 cm (or half-length without the handle).
• A custom 3D-printed base and aluminum mounting plate integrate all sensors and protect the LiDAR.
• Computation and storage are handled by an onboard NVIDIA Jetson AGX Orin Developer Kit, featuring a 12-core ARM Cortex-A78AE CPU and 2048 CUDA cores.
• Complete technical documentation, mechanical CAD files, and open-source software for calibration, synchronization, and data acquisition (e.g., smapper_toolbox) are available online. Calibration utilizes established tools such as Kalibr and generates transformation trees (TF) for sensor registration.

This open design enables reproducible experimental setups and encourages hardware extensions for new modalities or custom research platforms.

5. Data Format, Synchronization, and Storage

Each SMapper-light sequence is distributed as a ROS bag in .mcap format, encapsulating aligned streams from all modalities. Temporal alignment is achieved using selectable timestamping, with TIME_FROM_PTP supplying sub-millisecond accuracy by synchronizing all sensors to the Jetson Orin’s grandmaster clock.

The abstract synchronization relationship may be expressed as

$$t_\text{corrected} = f(\text{TSC}, \delta_\text{ptp})$$

where $\delta_\text{ptp}$ is the inter-sensor offset computed via PTP negotiation, ensuring global alignment for SLAM and sensor fusion.

This infrastructure supports precise multimodal time-correlation, critical for evaluating spatio-temporal modeling, calibration quality, and sensor fusion algorithm performance.

6. Practical Applications and Research Impact

SMapper-light encompasses a diverse range of environments and trajectory profiles:

• Indoor scenes: Spanning single-room setups, large multi-room areas, and scenarios with multiple loop closures.
• Outdoor scenes: Covering urban campus paths with both linear and circular routes.

The dataset has been used to benchmark SLAM frameworks, analyze the effects of temporal synchronization, and support investigations into multimodal sensor fusion. It is also positioned for research into:

• Scene graph generation and semantic mapping leveraging dense 3D reconstructions,
• Cross-modal alignment between visual and LiDAR data,
• Evaluations of inertial, depth, and appearance data fusion for enhanced SLAM robustness in challenging conditions.

A plausible implication is that the comprehensive multimodal coverage, precise ground truth, and open-hardware reproducibility of SMapper-light will facilitate the development of more robust, multimodal SLAM systems and reproducible benchmarking standards within the research community (Soares et al., 11 Sep 2025).