Hilti x Trimble SLAM Challenge 2026 in FiftyOne: One Anchor Gets You to the Door, Not the Route

The Hilti x Trimble SLAM Challenge 2026: a SLAM benchmark that ships its own floor plans

Key takeaways

• The Hilti x Trimble SLAM Challenge 2026 dataset in the FiftyOne Dataset Zoo covers 30 runs across 10 floors of one active construction site, recorded on eight dates between May and December 2025 with a single consumer 360 camera.
• Each Hilti x Trimble SLAM Challenge 2026 run ships one measured floor-plan pose, and that single pose is the only link between the run's SLAM trajectory and the building's floor plan.
• On floor_6 of the Hilti x Trimble SLAM Challenge 2026, anchoring each run to the floor plan with one rigid transform left same-day endpoints 1.05 meters apart and cross-date endpoints 5.9 meters apart on average.
• Whole-path chamfer distance between Hilti x Trimble SLAM Challenge 2026 floor_6 runs was about 4 meters for both same-day and cross-date pairs. Agreeing endpoints after single-anchor registration do not mean the walked routes agree.
• To evaluate single-anchor registration on any repeat-visit dataset, report an endpoint distance, a whole-path distance such as chamfer or Hausdorff, and the sample size.

Is one anchor point enough to register a SLAM trajectory to a floor plan?

Four visits to the same floor, six months apart

from mcap.reader import make_reader
from mcap_protobuf.decoder import DecoderFactory
from scipy.spatial.transform import Rotation

def read_poses(mcap_path, topic):
    with open(mcap_path, 'rb') as f:
        reader = make_reader(f, decoder_factories=[DecoderFactory()])
        for _, _, message, proto_msg in reader.iter_decoded_messages([topic]):
            p, q = proto_msg.pose.position, proto_msg.pose.orientation
            yield message.log_time * 1e-9, (p.x, p.y, p.z), (q.x, q.y, q.z, q.w)

# anchor: rotate/translate the run's own frame onto the floor-plan pose
R = (Rotation.from_quat(floorplan_quat) *
     Rotation.from_quat(first_gt_quat).inv()).as_matrix()
t = floorplan_pos - R @ first_gt_pos

What we measured: endpoint distance vs. whole-path chamfer distance

Building your own anchor-and-compare check

• Pull the first and last pose of each run in its own frame, plus whatever single shared-frame pose you do have (a GPS fix, an ArUco marker, a known doorway).
• Compute one rigid transform from the run's first pose to the shared anchor, and apply it to the whole run. Resist the urge to also correct the end point; that would hide exactly the error this check exists to find.
• Measure two different things, not one an endpoint-to-endpoint distance across runs, and a whole-path distance (chamfer, or Hausdorff for a stricter worst-case bound). A single anchor can pass the first test and fail the second, as it did here.
• Report both, and state sample size. An n=1 same-day pair is not proof of repeatability, it is one data point, exactly as honestly stated in measurements.json for this post.

Where a one-point anchor is the only option you have

• Facility and construction progress capture. Hilti's own SLAM challenges exist because large active sites cannot wire up real-time kinematic (RTK) GPS or ultra-wideband (UWB) anchors fast enough to keep up with changing floors; a walked survey anchored at one known doorway or marker per visit is the realistic workflow, not a limitation of the demo.
• Warehouse and retail robots between relayouts. A mobile robot re-localizing against a stored map after shelving moves typically gets one confident anchor (a fixed charging dock or fiducial) and has to trust its own odometry for everything else until the next anchor.
• Underground and tunnel inspection. GPS-denied by definition, so a surveyed shaft entrance or known junction is often the only absolute reference a crew gets per run, the same one-point-anchor situation this dataset reproduces on the surface.

Try the Hilti x Trimble SLAM Challenge 2026 dataset in FiftyOne