SMapper: Open-Hardware SLAM Platform

• SMapper is an open-hardware, multimodal sensor platform designed for SLAM research with tightly synchronized LiDAR, camera, and inertial data.
• It features both manual and automated dockerized calibration pipelines, ensuring sub-centimeter spatial and millisecond temporal accuracy.
• The platform and its SMapper-light dataset enable standardized, repeatable benchmarking of SLAM algorithms across diverse indoor and outdoor scenarios.

SMapper is an open-hardware, multi-sensor data acquisition platform explicitly designed to address critical challenges in reproducibility, environmental diversity, and multimodal sensing within Simultaneous Localization and Mapping (SLAM) research. It provides tightly synchronized LiDAR, multi-camera, and inertial data, a robust calibration and synchronization pipeline, and an extensible mechanical and electronic architecture. SMapper, accompanied by the SMapper-light dataset and automated calibration tools, enables standardized evaluation and benchmarking for a comprehensive range of SLAM algorithms in both handheld and robot-mounted configurations (Soares et al., 11 Sep 2025).

1. Hardware Architecture

SMapper’s sensor suite and mechanical structure are detailed in Table 1 and Figures 1–2 of the reference paper. The platform integrates the following components:

Sensor	Model / Spec (Key)	Key Properties
LiDAR	Ouster OS0, 64-beam	10/20 Hz, 100 m, 360°×90°, 1024×64, IMU@100 Hz
Cameras	4× e-CAM200 CUOAGX	RGB, 2K, rolling shutter, 90°×66°, 30 Hz, sync
RGB-D Camera	Intel RealSense D435i	RGB (2K, 69°×42°), Depth (87°×58°), 30 Hz, sync, IMU@400 Hz
Onboard Computer	NVIDIA Jetson AGX Orin DevKit	12-core Arm Cortex-A78AE @1.3 GHz, 2048 CUDA, 64 GB RAM, DL accelerator

The mechanical design features a custom 3D-printed PLA base housing the main electronics and cabling, topped with an anodized aluminum plate that serves as the LiDAR mount and protective cage. The four e-CAM200 cameras are spaced approximately 90° apart to provide ~270° horizontal FOV with ~30° overlap; the D435i RGB-D camera faces forward at 0° pitch. The system measures 15×15×38.4 cm (with handle and battery, ~2.5 kg) or 15×15×19.2 cm (without, ~1.7 kg). It supports handheld operation with a detachable polymer handle and robot-mounting via bottom M4×0.7 threaded holes.

All CAD files, step-by-step assembly guides, and bill of materials (BoM) are published under a CC-BY-4.0 license, facilitating direct replication and adaptation to new sensor configurations (see https://snt-arg.github.io/smapper_docs/).

2. Calibration and Synchronization Pipeline

SMapper employs both manual and automated (dockerized) calibration pipelines to ensure precise spatial and temporal alignment across all sensing modalities.

2.1 Spatial Calibration

Rigid body extrinsics between the sensor frames are established with:

• Frames: base_link (device origin), LiDAR (OS0), LiDAR IMU, D435i camera/IMU, four e-CAM200 camera frames (see Fig. 5).
• Transforms: $T_{LC_i}$ (LiDAR IMU → camera i), $T_{LI}$ (base_link → LiDAR IMU), $T_{CI}$ (base_link → camera-i IMU).
• Manual pipeline: Utilizes Kalibr for IMU noise characterization (Allan variance, estimating $\sigma_\omega$, $\sigma_a$, bias random walk), and for joint optimization of each camera’s intrinsics and extrinsics using an AprilTag 6×6 grid (0.8×0.8 m).
• Automated calibration: The smapper_toolbox provides a dockerized Kalibr workflow, parameterized by YAML, and outputs ROS2 launch files with static transforms. (https://github.com/snt-arg/smapper_toolbox)

Extrinsic results (Table 2): mean reprojection errors range from 0.34±0.34 px (front right) to 0.64±0.60 px (RealSense). Table 3 indicates all position differences between CAD and Kalibr extrinsics are below 4 cm, with angular differences <2°. LiDAR-to-camera colorization (Fig. 6) visualizes residual misalignments of a few centimeters at object boundaries.

2.2 Temporal Synchronization

All sensors utilize NVIDIA’s system clock and support three timestamp modes:

1. TIME_FROM_ROS: ROS clock at frame arrival, event-driven, jitter <5 ms.
2. TIME_FROM_TSC: Jetson Timestamp Counter, high resolution, jitter ~1 ms.
3. TIME_FROM_PTP: PTP network clock with Jetson as grandmaster, OS0 as slave, enabling sub-millisecond alignment.

No pre-bundled stream synchronization (i.e., “message sync”) is used; instead, raw, individually timestamped streams maximize post-processing flexibility and minimize data duplication. Temporal-offset compensation is modelled as: - $$\Delta t_{ij} = (t_i^{acq} + \delta_i) − (t_j^{acq} + \delta_j)$$, - $$t_i^{corrected} = t_i^{acq} − \Delta t_{Lidar,Camera}$$

3. Open-Hardware and Reproducibility

The entire SMapper platform—including all mechanical CAD files (.STEP, .STL), electronic schematics, and BoM (with component suppliers and pricing)—is open-sourced under CC-BY-4.0. The smapper_toolbox automates the calibration process, promoting plug-and-play reproducibility. The design is modular, allowing the base to be re-printed for new camera or LiDAR geometries; the top plate uses a standard 120×120 mm pattern for ease of modification. The platform supports ROS2 Foxy and Galactic, with Docker containerization and build instructions provided for rapid deployment and replication (Soares et al., 11 Sep 2025).

A plausible implication is that this open hardware ecosystem enables direct, precise comparison of SLAM algorithms across research groups by eliminating ambiguity in data provenance and sensor configuration.

4. Data Collection and the SMapper-light Dataset

SMapper-light is a representative, publicly available SLAM dataset collected using the SMapper platform and distributed in the .mcap rosbag format:

• Sequences (Table 4): 6 total (≈35 min, 164 GB), covering challenging indoor (4) and outdoor (2) scenarios.
  • Indoor: single-room, multi-room linear, multi-room with loop, large indoor with loop (1.5–9.5 min).
  • Outdoor: campus linear, campus circular (∼5 min each).
  • Acquisition: operator walking at normal/brisk paces with deliberate spins and loops to stress algorithms.
• Modalities: LiDAR point clouds (.pcd), camera images (.png/.jpg), IMU data (CSV/ROS messages), all with precise timestamps.
• Ground Truth: Trajectories are generated post hoc by offline, high-fidelity LiDAR SLAM, achieving >3 cm accuracy.
• Reconstructions: Dense 3D point clouds with densities exceeding 200 pts/m².
• Availability: Dataset is hosted at https://huggingface.co/datasets/snt-arg/smapper-light.

This unified dataset—grounded by sub-centimeter ground-truth and dense multimodal recordings—serves as a standardized testbed for SLAM benchmarking across geospatial, perception, and robotics communities.

5. Benchmarking Methodology and Results

5.1 Experimental Protocol

Benchmarking evaluates SMapper-light on state-of-the-art SLAM frameworks:

• LiDAR-IMU: GLIM (GPU-accelerated), S-Graphs/S-Graphs⁺ (semantic scene graph-based).
• Visual(-Inertial): ORB-SLAM3, vS-Graphs (visual + situational graphs).

All algorithms are executed with default parameters on the Jetson AGX Orin.

5.2 Performance Metrics

• Absolute Trajectory Error (ATE): $$ATE = \sqrt{ ( \hat{x}_i - x_i )^\top ( \hat{x}_i - x_i ) }$$ averaged over $i$.
• Relative Pose Error (RPE): $$RPE(\Delta) = \| ( \hat{T}_{i, i+\Delta} )^{-1} T_{i, i+\Delta} \|$$.
• Map quality: Point-to-point RMSE compared to dense ground-truth point cloud.

Sequence	ORB-SLAM3 ATE (m)	GLIM ATE (m)	vS-Graphs RPE (m)	S-Graphs RMSE (m)
IN_SMALL_01	0.12	0.05	0.14	0.04
IN_MULTI_02	0.28	0.12	0.32	0.11
OUT_CAMPUS_02	0.35	0.18	0.38	0.17

5.3 Comparative Analysis

• LiDAR-based pipelines achieve sub-decimeter ATE, proving robust against textureless and low-light scenes.
• Visual SLAM results display greater translational drift (>0.2 m over long loops) but yield richer appearance-based maps.
• Semantic/graph-enhanced methods (S-Graphs, vS-Graphs) yield modest improvements in loop closure and resilience in complex indoor spaces.
• All evaluated approaches successfully process data under handheld conditions characterized by pedestrian-induced vibrations.

Figures 7–9 provide qualitative reconstructions (colored point clouds, semantic scene graphs) for each approach and sequence.

6. Significance and Impact on SLAM Research

SMapper’s open-hardware reproducibility, rigorous calibration pipeline, and synchronized multimodal sensing address long-standing limitations in dataset comparability, environmental representation, and hardware divergence typical of prior SLAM testbeds. As all hardware designs, software stacks, and calibration workflows are fully open-sourced, researchers are able to build identical or adapted SMapper units, perform data acquisition in new scenarios, and benchmark SLAM algorithms using a unified protocol and data representation. The inclusion of tight temporal/spatial sensor calibration, comprehensive benchmarking, and sub-centimeter ground truth in SMapper-light establishes a foundation for robust, repeatable, and comparative research in SLAM and broader perception tasks (Soares et al., 11 Sep 2025).