Skip to content

Hilti x Trimble SLAM Challenge 2026

Related docs: SLAM Benchmarking Metrics and Datasets, OpenVINS, ORB-SLAM2 and ORB-SLAM3, SVO, Cartographer 3D, and AV / Indoor / Outdoor SLAM Decision Matrix.

Last updated: 2026-05-23

Executive Summary

The Hilti x Trimble SLAM Challenge 2026 is a construction-site visual-inertial benchmark from Hilti, Trimble, and the University of Oxford Dynamic Robot Systems Group. It differs from the older Hilti SLAM challenge line because it centers on a 360-degree camera with embedded IMU measurements and optional floor-plan priors, with separate tasks for free-frame SLAM and localization in the floor-plan reference frame.

This is a strong reference for industrial indoor localization because it tests active construction-site geometry: repetitive texture, changing lighting, partially built structures, floor-plan mismatch, rolling-shutter 360 imagery, dynamic initialization, and hidden ground-truth evaluation. It is not a direct AV benchmark, but it transfers well to warehouses, terminals, depots, hangars, construction sites, and other sites where a robot may have building plans but not a finished, static environment.

What It Contains

The official challenge page and repository describe:

  • 30 public recordings across multiple floors and underground levels.
  • Two challenge categories: SLAM in any reference frame and localization in the floor-plan map frame.
  • 360-degree visual-inertial data from an Insta360 One-RS 1-Inch Edition camera.
  • Embedded IMU measurements from the camera.
  • Floor-plan images, including variants with and without windows.
  • LiDAR-inertial ground truth generated by a rigidly attached Hesai XT32-based mapping device.
  • ROS 2 bags and helper tools for image decompression, image stitching, image inversion, floor-plan map serving, OpenVINS examples, and Stella-VSLAM examples.

The raw LiDAR data used for ground truth is not part of the released challenge input. That makes the benchmark specifically about visual-inertial SLAM/localization from the camera rig, with floor-plan context for the localization task.

Evaluation Model

SignalChallenge behaviorWhy it matters
Trajectory coverageRuns below 99 percent matched-pose coverage receive zero for that runPrevents methods from producing accurate fragments while failing most of the route.
Exponential position scoreErrors are converted to a bounded score per pose and summed by runRewards accurate localization while making large errors visibly costly.
Hidden evaluationSome run details are withheld from public feedbackReduces overfitting to the evaluation set.
Floor-plan localization taskCamera pose must be expressed in a map frame derived from floor-plan pixelsTests the practical problem of localizing against imperfect building priors.
No post-hoc scale adjustmentThe challenge notes that scale is not adjusted during scoringForces visual-inertial systems to maintain metric scale.

Domain Fit

DomainFitNote
Construction-site roboticsStrongThis is the benchmark's primary domain.
Warehouses and depotsStrong to conditionalFloor-plan priors, repetitive interiors, and changing layouts transfer well.
Airport terminals / hangarsConditionalUseful for indoor/hangar localization; not representative of open aprons.
Road AVWeakVehicle dynamics, weather, traffic, HD maps, and sensor suites differ.
Airside apronWeak to conditionalUseful only for indoor or terminal-edge visual-inertial failure analysis.

Failure Modes It Exposes

  • VIO initialization while the operator or platform is already moving.
  • Rolling-shutter and dual-fisheye geometry artifacts from 360-degree cameras.
  • Drift when floor-plan priors disagree with the as-built construction state.
  • False localization in repetitive corridors, columns, and unfinished rooms.
  • Scale inconsistency when the method cannot maintain metric visual-inertial pose.
  • Tooling assumptions that fail on ROS 2 bag formats, compressed images, or unusual camera models.

Implementation Notes

  • Keep SLAM and localization task results separate; one estimates trajectory in any frame, the other must align to the floor-plan map frame.
  • Report whether floor-plan priors are used, and how map mismatch is handled.
  • Treat the provided floor plans as imperfect priors, not surveyed truth.
  • Use the OpenVINS and Stella-VSLAM examples as reproducibility baselines, not as proof that a method is robust.
  • Note the CC BY-NC-SA 3.0 dataset/benchmark license before using data in commercial or redistribution workflows.

Limitations

  • The benchmark is construction-site focused, not road, airside-apron, port, mining, or agricultural autonomy.
  • The challenge deadline was 2026-05-15 for prize eligibility, though the evaluation system remains useful as a reference.
  • Most ground truth is hidden, and raw LiDAR ground-truth acquisition data is not released as normal benchmark input.
  • The camera is a consumer 360-degree device, so results may not transfer directly to calibrated industrial multi-camera rigs.
  • Dataset licensing is noncommercial share-alike.

Sources

Public research notes collected from public sources.