Hilti-Research/hilti-trimble-slam-challenge-2026

Hilti x Trimble SLAM Challenge 2026

The Hilti x Trimble SLAM Challenge 2026 dataset is a real-world robotics benchmark for evaluating visual-inertial SLAM and localization systems on active construction sites.

The dataset combines synchronized dual-fisheye imagery and inertial measurements with building floor plan priors and LiDAR-derived reference trajectories. It was created through a collaboration between Hilti, Trimble, and the Dynamic Robot Systems Group at the University of Oxford.

• —Challenge website: hilti-trimble-challenge.com/dataset-2026
• —Code, tools, calibration, and detailed documentation: GitHub repository
• —Paper: Hilti-Trimble-Oxford Dataset: 360 Visual-Inertial Benchmark with Floor Plan Priors for SLAM and Localization

Dataset overview

The dataset contains 30 recording sequences collected across multiple floors and dates at an active construction site.

It supports two principal tasks:

1. 1.SLAM: Estimate the camera trajectory in an arbitrary reference frame.
2. 2.Floorplan localization: Estimate the camera trajectory in the coordinate frame of the supplied building floor plan.

The recordings are intended to capture challenges encountered in practical construction-site robotics, including:

• —wide-field-of-view dual-fisheye imagery;
• —rolling-shutter cameras;
• —repetitive and texture-poor indoor environments;
• —changing lighting and scene conditions;
• —moving objects and ongoing construction work;
• —sequences with motion during initialization;
• —differences between planned floor plans and the as-built environment.

Sensor data

Each sequence was recorded using an Insta360 ONE RS 1-Inch 360 Edition camera containing two approximately 200° fisheye lenses and an integrated six-axis IMU.

The released ROS 2 bags contain:

The dataset provides the original dual-fisheye images rather than pre-stitched equirectangular panoramas. Because the two lenses have distinct optical centers, panorama generation introduces assumptions and may produce parallax-related geometric inconsistencies.

A reference stitching implementation and additional image-processing tools are available in the challenge repository.

Ground truth

Reference camera trajectories are provided for all sequences using the cam0 -> map convention: the pose expresses the position and orientation of cam0 in the map frame.

The trajectories were generated using a LiDAR-inertial mapping system rigidly attached to the camera rig, followed by LiDAR-to-camera transformation and IMU time alignment.

The raw LiDAR measurements used to create the reference trajectories are not included in the released dataset.

Because LiDAR recording begins slightly after camera recording, the reference trajectory does not cover approximately the first five seconds of each sequence.

For the localization task, an initial camera pose in the floor plan coordinate frame is also provided near timestamp 10005 s.

Dataset structure

Recordings are organized by floor, recording date, and run number:

data/
└── floor_X/
    └── YYYY-MM-DD/
        └── run_Z/
            └── rosbag/
                ├── rosbag.db3
                └── metadata.yaml

A sequence is identified using the following convention:

floor_X_YYYY-MM-DD_run_Z

For example:

floor_2_2025-10-28_run_1

The release includes the original calibration recordings and Kalibr configuration files for the two cameras and IMU (in camera_calibration folder). The kalibrimucamchain.yaml and kalibrimuchain.yaml files provide the calibration chains used with the dataset. The accompanying ROS 2 bags and AprilGrid configuration files are included to support calibration inspection and reproducibility.

Floor plan resources are distributed in three representations:

floorplans/
├── binary_masks/
│   ├── masks_no-window/
│   │   ├── BuchsIT_X_mask_nowindows.png
│   │   └── ...
│   └── masks_with-windows/
│       ├── BuchsIT_X_mask_windows.png
│       └── ...
├── dxf_format/
│   ├── floor_X.dxf
│   └── ...
└── png_format/
    ├── floor_X.png
    └── ...

The binary masks are available in variants with and without windows. DXF files provide vector floor plans, while the PNG files provide rasterized plans suitable for visualization and occupancy-map workflows. One sequence (floor_UG2_2025-12-02_run_1) does not have an associated floor plan and is therefore unsuitable for the floorplan-localization benchmark.

We additionally provide SEQUENCES.md file which specifies the difficulty of each sequence in the dataset.

Timestamp convention

A constant offset of 10000 s was applied to each recording to prevent negative IMU timestamps. Image timestamps therefore begin near 10000 s.

Users combining image, IMU, calibration, initial-pose, and reference-trajectory data should preserve this timestamp convention.

Challenge evaluation

The original challenge evaluated trajectory completeness and position accuracy for the SLAM and localization tasks. Predictions used the TUM trajectory format:

timestamp tx ty tz qx qy qz qw

The full evaluation protocol—including sequence exclusions, minimum trajectory coverage, coordinate conventions, scoring equations, file naming, and benchmark-specific requirements—is maintained in the challenge repository.

Citation

When using this dataset in academic work, please cite:

@misc{slamchallenge2026,
  title  = {{Hilti}-{Trimble}-{Oxford} Dataset: 360 Visual-Inertial Benchmark with Floor Plan Priors for SLAM and Localization},
  author = {Centanni, Samuele and Zhang, Yuhao and Tao, Yifu and Kindle, Julien and Neuhaus, Frank and Ko{\ss}, Tilman and Patel, Aryaman and Helmberger, Michael and Szyma{\'n}ska, Emilia and Gr{\"a}ber, Torben and Fallon, Maurice},
  year   = {2026},
  eprint = {2607.06464},
  url    = {https://arxiv.org/abs/2607.06464}
}

Maintainers and contact

For questions about the dataset, formats, calibration, or tools, open an issue in the GitHub repository.

For questions concerning the challenge website or evaluation infrastructure, contact:

challenge@hilti.com