A quadruped robot system has been reported to explore a stairwell from the first floor to the fifth, climbing four consecutive corner staircases, in a real-world demonstration. The same system, called RAEM, was reported to achieve stable and exhaustive exploration across 20 independent simulation trials.
The paper is a preprint labeled arXiv:2608.25366v1 [cs.RO] 26 Aug 2026. It reports a framework for multi-floor exploration with a quadruped robot, combining Gazebo comparisons with selected real-world tests.
A layered map for a moving robot
RAEM combines local tomography, a way of representing nearby three-dimensional space from scans, and categorized 3D-grid maps with an elevation-aware global graph. It also includes staircase center alignment and dual path searching, making the architecture a hybrid local-global design.
The computational work is divided between the robot's graphics processor and central processor. GPU parallelization handles local tomography, local 3D-grid maps and slice-obstacle generation, while the CPU maintains the occupancy grid used for frontier identification.
Simulation tested multi-floor reach
In simulation, the evaluation used Gazebo on a laptop and a Unitree A1 quadruped equipped with a Mid-360 LiDAR. The comparison included TARE, FAEL, HPHS and RAEM, and each method was run 20 times in each environment from the same start position.
The baseline results were different from RAEM's reported multi-floor outcome. TARE and FAEL completed only the first floor in every scene and could not navigate higher. HPHS did not stably complete the first floor in scenes 2 and 3 and had no successful scene-4 attempt.
Across scenes 1 through 4, RAEM's reported mean exploration times were 174, 270, 670 and 477 seconds. Mean trajectory distances were 137, 195, 416 and 322 metres, while mean occupied volumes were 389, 393, 1,065 and 501 cubic metres, respectively.
These are descriptive results rather than a statistical estimate of performance more broadly. The analysis reports standard deviations and ranges for the simulation comparisons, but no confidence intervals or inferential tests, and the trial-level raw results are not provided.
The computer work was timed separately
The paper also measured local 3D-grid construction. Mean GPU time was lower than CPU time in every scene: 0.0027 versus 0.0268 seconds in scene 1, 0.0028 versus 0.0506 seconds in scene 2, 0.0027 versus 0.0500 seconds in scene 3, and 0.0027 versus 0.0504 seconds in scene 4.
Viewpoint generation showed a similar pattern in the reported means. Clustering took 0.0012 versus 0.0015 seconds in scene 1, 0.0024 versus 0.0029 in scene 2, 0.0026 versus 0.0053 in scene 3, and 0.0022 versus 0.0051 in scene 4, with clustering listed first each time. These were component timings, not whole exploration times.
Two tests focused on the stairs
A separate component test examined center alignment on selected simulation staircases. With alignment, the reported climbing trajectories were stable and nearly straight. Without it, one of 20 scene-2 trials ended in a catastrophic fall. The evaluation covered selected staircases, so it does not settle how the choice performs across other layouts.
A second comparison examined dual path searching in one corner-staircase scenario. In 20 independent scene-2 trials, dual path searching succeeded in all 20 attempts, while topology-only searching succeeded 12 times. The paper does not report an inferential test for that difference.
The robot was also run in buildings
For real-world testing, the system ran on a Unitree Go2 carrying a head-mounted Mid-360S LiDAR pitched 45 degrees downward. It used onboard NVIDIA Jetson Orin NX 16GB compute and Fast-LIO2 for real-time localization and mapping, with a reported effective LiDAR range of 4 metres.
The reported times and distances varied by setting. The stairwell run took 233.9 seconds over 104.9 metres at a maximum velocity of 1.0 metres per second. Building 1 took 290.0 seconds over 132.2 metres at 0.75 metres per second, building 2 took 375.0 seconds over 166.0 metres at 0.75 metres per second, and the lobby took 247.0 seconds over 99.9 metres at 1.0 metres per second.
The paper also reports complete exploration of both targeted floors in building 1 and building 2. In the lobby, movement continued without pauses while the system waited for planning results.
The evidence remains narrow
The real-world figures should be read as scenario demonstrations, not as a reliability rate. Repetition counts and variability for those runs were not reported. The broader test was also narrow: simulation used a Unitree A1 and the real-world work used a Unitree Go2, with demonstrations covering a stairwell, two buildings and a lobby. The findings describe performance in the tested configurations, but do not establish performance in untested ones.
Paper data and sources
Original title: RAEM: Robust Autonomous Exploration for Multi-Floor Environments with a Quadruped Robot
Authors: Zikang Yuan, Yuan Ren, Yian Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text