Robots Atlas>ROBOTS ATLAS
Robotics & Hardware

StrataSLAM Keeps Floors Apart. Without the Gate, 85.6% Collapse

Lady Robot22 November 2026 ยท 7 min read
StrataSLAM Keeps Floors Apart. Without the Gate, 85.6% Collapse

The Intelligent Robotics Lab at Universidad Rey Juan Carlos released StrataSLAM on November 9 โ€” an open-source multi-floor 3D SLAM for ROS 2. The system keeps one submap per floor and never accepts a loop closure between two storeys. In the repository's own tests, that single block cuts the floor-collapse rate from 85.6% to zero.

Key takeaways

  • Version 0.2.0, Apache-2.0 licence, ROS 2 Jazzy and Ubuntu 24.04, core in C++17
  • Scenario with the floor gate disabled: 85.6% collapsed storeys. With it on: 0
  • Sensors: RGB-D camera, 3D LiDAR or both. IMU, barometer and wheel odometry optional
  • KITTI drive of 692 m: ATE RMSE 0.4013 m, 52 ms per scan on one thread
  • Three identical floors without Structural anchors: Permanent building features visible from several floors โ€” a lift shaft, a stairwell, an atrium. They serve as the reference when the storeys are assembled.: 24.8 cm error instead of 6.9 cm

Picture the situation

A wheeled robot works a three-storey office building. It enters the elevator, the doors close, the cabin rises. For those few seconds the camera and LiDAR see only the inside of the cabin โ€” nothing that moves relative to them.

On stepping out, the robot stands in a corridor that looks identical to the one downstairs: the same door layout, the same carpet, the same signage.

But a problem appears: a conventional SLAM system reads that corridor as somewhere it has already been, closes the loop, and merges two floors into one.

One map remains โ€” flat and wrong. StrataSLAM targets exactly that moment.

85.6% โ†’ 0collapsed floors once the gate is on, across 20 random seedsStrataSLAM, scenarios S03 and S04

What exactly does StrataSLAM do?

The system does not hold a single map of the building. It holds one submap per floor and builds a global map from them.

Transitions between storeys are detected and classified โ€” stairs, ramp, elevator, flight between levels. An elevator ride is measured with the IMU even when the camera and LiDAR only see the cabin.

The floor gate matters most, though. No online code path accepts a loop closure between two floors. This is not a confidence threshold to be tuned, but a prohibition written into the architecture.

From the announcement: simulated buildings first, then real data. The run with the floor gate disabled is shown side by side with the same run with the gate on.

Transition
Robot enters the elevator
IMU measures the ride
Transition classified
Is the loop candidate on the same floor?
SAME FLOOR
Loop acceptedAllow
DIFFERENT FLOOR
Floor gate rejects itDeny
Result
One submap per floor + a global mapAllow

This is a simplified flow of one transition, not the system's full operating diagram.

The floors are assembled into a whole only later, offline. The strataslam fuse tool places them relative to one another in a gravity-aligned four-degree-of-freedom group written as SE(2)ร—โ„ โ€” even with no visual overlap between storeys. It draws on the transition constraints, the barometer, structural anchors and a weak floor-height prior.

How much does that gate actually buy?

The repository ships a negative control, so the figures compare directly.

Scenario S03 is two identical floors joined by stairs, camera only, no IMU. Across 20 Random seed: A number that initialises the random generator. The same seed reproduces a test run exactly, so results can be repeated and compared. the collapse rate was 0, the number of cross-floor loop candidates also 0, and the floor-height error 3 centimetres. Scenario S04 is the same run with the gate disabled โ€” and 85.6% collapsed storeys.

ScenarioConfigurationCollapsed floorsFusion error
S03two identical floors, stairs, gate on0%3 cm (floor height)
S04same run, gate off85.6%โ€”
S11three floors, stairs and elevator, with anchors0%6.9 cm / 0.075ยฐ
S11same inputs, without anchors0%24.8 cm / 1.43ยฐ

The second telling ablation concerns structural anchors. With three identical floors, stairs and an elevator, the fusion error came to 6.9 centimetres in translation and 0.075 degrees in rotation. Without anchors the same inputs gave 24.8 centimetres and 1.43 degrees โ€” more than three times worse.

On real data it looks like this: TUM RGB-D fr1/xyz (798 frames) gives an ATE RMSE of 1.43 centimetres, fr2/xyz (3,669 frames) 1.52 centimetres. A 692-metre KITTI drive produced 0.4013 metres across 506 keyframes, one loop closure, and 52 milliseconds per scan on a single thread.

DatasetPipelineFramesATE RMSEMax error
TUM RGB-D fr1/xyzvisual odometry7980.0143 m0.0551 m
TUM RGB-D fr2/xyzvisual odometry3,6690.0152 m0.0326 m
KITTI 2011_09_30_0027 (692 m)full LiDAR SLAM1,1060.4013 m0.7582 m

What do these names mean?

A few terms from the documentation need translating:

  • SLAM โ€” building a map and locating yourself within it at the same time. The robot draws the floor plan and marks itself on it, without having the plan beforehand.
  • Loop closure โ€” recognising that the robot has returned somewhere it already visited, and stitching the map at that point. The equivalent of matching the two ends of an unrolled tape.
  • Submap โ€” a separate, locally consistent fragment of the map. Here: exactly one per storey.
  • ATE RMSE โ€” the error between the computed and the true trajectory, taken as the root of the mean of the squares. One number saying how many metres the system was off by.
โ€ฆ
  • Four degrees of freedom โ€” translation on three axes plus one rotation about the vertical. That suffices because gravity fixes which way is up.

Together they describe one design choice: rather than fighting bad loops after the fact, do not permit them at all.

Why isn't a barometer enough?

It is the first thing that comes to mind: if height is the problem, measure it with pressure.

StrataSLAM does support a barometer, but strictly as an optional sensor and one input among several during offline fusion. On its own it does not solve the task, because the mistake is not only about height โ€” it is about the identity of a place. Two corridors on different floors can yield a near-identical image and a near-identical point cloud, while pressure in an air-conditioned building drifts across the day.

The second problem is practical. Even with height known perfectly, something still has to decide what to do with a loop candidate: accept it and risk merging floors, or reject it and risk the map drifting apart. The gate turns that into a hard decision rather than one more parameter to tune.

That is also why structural anchors and a weak floor-height prior are in the mix. The scenario with four identical floors shows where this still falls short: the floor-height error there averaged 45 centimetres.

Is it working yet?

Yes โ€” as open source at version 0.2.0, not as a finished product. The repository is public under Apache-2.0, needs ROS 2 Jazzy and Ubuntu 24.04, and the first build pulls a pinned commit of the small_gicp library.

The authors publish their failures too, which is not the rule in open SLAM projects. The elevator scenario without IMU or barometer passed 13 of 20 seeds โ€” seven times the translation error exceeded the 10-centimetre threshold. The gentle ramp (2.9 degrees, 30-metre flights) passed 19 of 20. In the performance figures the LiDAR frontend for 64 beams takes 243 milliseconds against its own budget of 50 milliseconds per scan โ€” an overrun the docs mark openly.

The announcement video shows simulated buildings first, then real data: a handheld LiDAR-plus-IMU rig through actual elevators, a Summit XL robot on ramps between levels, a TIAGo with LiDAR and an RGB-D camera, and the KITTI drive outdoors.

The system publishes map -> floor_<k> -> odom per REP-105, so Nav2 and RViz see the robot while it is still mapping. The strataslam doctor command checks topics, rates, TF and synchronisation before the first mapping run.

Bash
# ROS 2 Jazzy, Ubuntu 24.04 โ€” one YAML profile per robot
cp $(ros2 pkg prefix strataslam_ros)/share/strataslam_ros/profiles/ground_lidar.yaml ~/my_robot.yaml

# pre-flight check: topics, rates, TF, synchronisation
strataslam doctor ~/my_robot.yaml

# map, then fuse the floors into one building
ros2 launch strataslam_ros strataslam.launch.py profile:=$HOME/my_robot.yaml viewer:=true
strataslam fuse ~/strataslam_maps/<run>

The flow from the project documentation. Fusing the floors (strataslam fuse) is an offline step, run once mapping has finished.

Why does it matter?

Multi-storey buildings are the core environment for service and delivery robots, and simultaneously the most common reason a map falls apart. The value of this release lies not in record accuracy but in an explicit constraint: the ban on cross-floor loops sits in the architecture, not in a parameter. The negative control shipped alongside makes it possible to check what that ban is actually worth.

What's next?

  • The authors ask for issues and rosbags from users' own multi-floor buildings โ€” a stated direction for the project
  • CITATION.cff records a pending task to add a DOI once the tagged release is archived on Zenodo
  • The LiDAR frontend overruns its own timing budget nearly fivefold, making performance work an identified technical debt

Sources

Share this article