Organizations: AI Chip Center for Emerging Smart Systems, InnoHK Centers, Hong Kong Science Park, Hong Kong, China · HongKong University of Science and Technology, Hong Kong, China · Southeast University, Nanjing, China · City University of Hong Kong, Hong Kong, China · Huazhong University of Science and Technology, Wuhan, China · University of Hong Kong, Hong Kong, China · Wuhan University, Wuhan, China
In this paper, we propose RAEM, a robust autonomous exploration framework for quadruped robots operating in multi-floor environments. Most existing ground-robot exploration approaches rely on planar traversability representations, which cannot adequately represent the overlapping structures and cross-floor connectivity of multi-floor buildings. Although tomography-based representations provide effective traversability modeling for multi-floor navigation, maintaining a global tomography map incurs substantial computational overhead for online exploration with frequent replanning. Moreover, sparse and fragmented LiDAR observations in stairwells can degrade local traversability estimation, leading to irregular viewpoint placement and temporary topological disconnections. To address these challenges, RAEM adopts a hybrid local-global traversability representation, in which a local tomography map and an explicitly categorized local 3D grid map are used for online terrain analysis and connectivity evaluation, while an elevation-aware global topological graph is incrementally constructed from these local spatial representations for efficient cross-floor exploration planning. We further introduce a staircase center alignment strategy to reduce abrupt yaw variations during climbing and a dual path searching mechanism to recover guidance paths when the global topology is locally disconnected. Extensive simulation and real-world experiments demonstrate robust and computationally stable autonomous exploration across multi-floor structures, including continuous exploration of a five-floor stairwell.
Figures & tables
Fig. 1: Illustration of the proposed autonomous exploration framework for multi-floor environments, which consists of four main capabilities: (a) online elevation-aware topological graph construction in multi-floor structures; (b) centerline-constrained viewpoint generation on staircases to prevent the quadruped robot from falling due to yaw variations on stairs; (c) traversable grid-based path searching to compensate for topological failures in stairwells due to sparse and fragmented point clouds; and (d) automatic filtering out structurally inaccessible yet physically traversable regions.
Symbol
Description
MoccG
Global occupancy grid
MtomoL
Local tomography map
MgridL
Local 3D grid map
MterL
Inflated local 3D terrain map
S
Discrete tomogram slices
Tobs
Slice-wise obstacle trees
TABLE I: Nomenclature
Fig. 2: System overview of RAEM. The framework explicitly leverages GPU parallelization for the rapid generation of local tomography map, local 3D grid map, and the extraction of slice-based obstacles. Concurrently, the CPU manages a continuously updated occupancy grid for identifying frontiers. These detected frontiers, combined with the GPU-derived local spatial representations, directly drive the core modules, including elevation-aware topological graph construction, viewpoint generation, and dual path searching. To overcome the challenges posed by degraded tomography maps, RAEM uniquely incorporates a staircase centerline alignment method for safe viewpoint placement and a dual path searching approach to address topological fractures. The global planner subsequently determines the optimal visitation sequence of viewpoints, and the local planner generates trajectory points toward the next goal.
Fig. 3: Traversability analysis of local tomogram slices. Slices are built by projecting the local point cloud onto equidistant horizontal planes, capturing orange ceiling and blue ground layers. The squares represent the cell traversability (green: traversable, red: untraversable) considering the ground conditions and the ceiling height. The red dotted circles highlight cells conservatively marked as untraversable due to strict proximity to structural edges or point cloud map boundaries, despite being physically traversable. Beneath staircases, sparse point cloud observations combined with this conservative labeling generate massive falsely untraversable cells, severely compromising topological building, viewpoint generation, and path searching.
Fig. 4: Illustration of construction processes of the local 3D grid map MgridL . (a)-(c) illustrate the intermediate grid state assignments processed through the structural logic of individual slices. (d) and (e) demonstrate the structural logic across multiple slices, where previously unassigned spatial grids bounded between the ceiling of a lower slice and the ground of an adjacent upper slice are rigorously classified as occupied grids (red) representing solid architectural structures. (f) displays the final integrated map, highlighting a morphological horizontal dilation applied to the traversable grids (indicated by dark green arrows) to significantly reinforce the robustness of path searching for staircases.
Fig. 5: Overview of the slice-based obstacle extraction and distance-to-obstacle query. The multi channel tomogram slices explicitly encode traversable regions (green cubes) and untraversable obstacle boundaries (purple cubes). The untraversable cells within each slice Sk undergo a dimensional flattening operation, directly projecting the complex spatial structures into a dense 2D point cloud to construct a dedicated planar tree Tobsk . For any query point p , its corresponding slice index k is mathematically determined based solely on its vertical elevation component [p]z , subsequently enabling rapid and highly efficient distance-to-obstacle evaluations through Tobsk .
Fig. 6: Overview of the elevation-aware topological graph construction. (a) A spatial environment featuring overlapping elevation layers. (b) Uniformly sampling candidate locations across the 2D grid plane (i,j) . (c) Untraversable or unavailable cells on Plane 0 generate no vertices. (d) and (e) Structurally safe and traversable cells on Plane 1 and Plane 2 successfully instantiate topological vertices at their respective ground elevations. (f) The finalized topological graph interconnecting these generated vertices utilizing bidirectional collision-free spatial edges under maximum allowable edge degree limitations.
Fig. 7: Illustration of viewpoint generation. (a) Raw samples are evaluated where invalid points (green) are discarded, and valid points (red) are vertically projected to surface-level positions (orange). (b) Projected points are categorized according to the specific slice scopes they fall into (e.g., blue and brown subsets). After eliminating points located too close to obstacles (red crosses), the remaining safe points are clustered using collision-free circles to extract candidate viewpoints. (c) The optimal viewpoint p1 is selected for maximizing frontier visibility compared to p2 and p3 . (d) On staircases, an elevation-constrained region expansion from p1 extracts the tread centroid to serve as the new viewpoint position.
Fig. 8: Illustration of the dual path searching approach for staircases. (a) Graph-based searching fails due to topological discontinuity. (b) The fallback traversable-grid based searching successfully resolves a valid trajectory (yellow path) to the upper-floor viewpoint.
Size (m)
Floor Count
Inter-Floor Height (m)
scene_1
18.4 × 18.4 × 8.2
3
2.1
scene_2
13.0 × 22.4 × 8.2
3
2.5
scene_3
24.0 × 34.0 × 7.7
2
3.4
scene_4
17.6 × 16.8 × 6.2
4
1.5
TABLE II: Information Of Simulation Scenarios
scene_1
scene_2
scene_3
scene_4
Real-World
Rocc
4.0
4.0
Blocal
12 × 12 × 4
12 × 12 × 6
12 × 12 × 4
12 × 12 × 6
N
5
5
rtomo
0.2
0.2
dmin
0.5
0.5
θs
0.6
0.6
TABLE III: Parameter Configurations
Fig. 9: The variation of coverage rate with time duration in exploration progress for (a) scene_1 , (b) scene_2 , (c) scene_3 , and (d) scene_4 .
Fig. 10: The variation of coverage rate with traveling distance in exploration progress for (a) scene_1 , (b) scene_2 , (c) scene_3 , and (d) scene_4 .
Fig. 11: Visualization of exploration paths generated by RAEM across scene_1 , scene_2 , scene_3 , and scene_4 .
Exploration Time (s)
Trajectory Distance (m)
Occupied Volume (m 3 )
Success Number
Avg
Std
Max
Min
Avg
Std
Max
Min
Avg
Std
Max
Min
1st Floor
2nd Floor
3rd Floor
4th Floor
scene_1
TARE
63
2
67
59
62
1
65
60
102
3
107
95
20
0
0
FAEL
41
5
48
32
32
4
37
26
90
11
109
74
20
0
0
HPHS
41
2
47
37
39
2
46
35
89
1
90
87
20
0
0
Ours
174
3
184
166
137
2
141
134
389
4
397
379
20
20
20
scene_2
TARE
86
9
108
74
81
8
104
72
109
6
122
100
20
0
0
TABLE IV: Results of Simulations in four Environments
Fig. 12: Evolution of the computation time for a single planning iteration as time progresses in (a) scene_1 , (b) scene_2 , (c) scene_3 , and (d) scene_4 .
Fig. 13: Box plot with logarithmic scale of single computation time of each module for (a) scene_1 , (b) scene_2 , (c) scene_3 , and (d) scene_4 .
scene_1
scene_2
scene_3
scene_4
running on GPU
0.0027 s
0.0028 s
0.0027 s
0.0027 s
running on CPU
0.0268 s
0.0506 s
0.0500 s
0.0504 s
TABLE V: Average Computing Time for Local 3D Grid Map Construction Using CPU and GPU Architectures
Fig. 14: Box plots of the distribution of single execution times for local 3D grid map construction utilizing CPU and GPU across all four simulation scenarios.
scene_1
scene_2
scene_3
scene_4
clustering
0.0012 s
0.0024 s
0.0026 s
0.0022 s
sampling
0.0015 s
0.0029 s
0.0053 s
0.0051 s
TABLE VI: Average Computing Time of the Sampling and Clustering Viewpoint Generation Schemes
Fig. 15: Box plots of the distribution of single execution times for the sampling and clustering viewpoint generation schemes across all four simulation scenarios.
Fig. 16: Comparative visualization of the quadruped robot ascending staircases. (a) and (b) illustrate the selected staircase regions for evaluation. (c) and (d) demonstrate the erratic climbing trajectories characterized by yaw angle variations generated without the center alignment constraint. (e) and (f) display the nearly straight trajectories produced when the proposed strategy is applied.
Fig. 17: Sequential illustration of a lateral rollover failure caused by the absence of the center alignment constraint. (a) While the quadruped robot climbs facing left, a viewpoint distributed on the right is selected as the next navigation goal, compelling the robot to physically maneuver toward it. (b) and (c) To reach this new target, the robot is forced to substantially alter its yaw angle directly upon the staircase, a dangerous kinematic shift that destabilizes the platform and results in a severe fall.
Fig. 18: Sequential demonstration of the dual path searching mechanism solving a topological disconnection on a corner staircase. (a) A target viewpoint fails to link to the topological graph. (b) The traversable grid-based A-star search activates as a reliable fallback, successfully finding a supplementary path indicated by the solid white line. (c) The local planner generates a navigation trajectory based on this discovered route. (d) The topological graph is fully recovered and reconnected after the robot climbs and densely scans the staircase region in close proximity.
Fig. 19: Hardware configuration of the robotic platform and the trajectory visualization scheme for the real-world experiments. The system comprises a Unitree Go2 quadruped robot equipped with a Mid-360S LiDAR and an onboard NVIDIA Jetson Orin NX computing unit. The accompanying rainbow color gradient applied to the trajectory intuitively illustrates the continuous chronological progression of the exploration process.
Size (m)
Exp. Time (s)
Traj. Dist. (m)
Max Vel. (m/s)
stairwell
5.5 × 14 × 21.5
233.9
104.9
1.0
building_1
16 × 38 × 8
290.0
132.2
0.75
building_2
51 × 45 × 8
375.0
166.0
0.75
lobby
35 × 26 × 5
247.0
99.9
1.0
TABLE VII: Configuration and Performance Metrics of the Four Real-World Testing Scenarios
Fig. 20: Experimental results in the real-world stairwell scene. The proposed framework successfully enables the quadruped robot to autonomously explore from the first floor to the fifth floor.
Fig. 21: Experimental results in the real-world building_1 scene. Upon discovering a staircase within the designated exploration boundary, the quadruped robot autonomously climbs to the upper floor to seamlessly continue exploration.
Fig. 22: Experimental results in the real-world building_2 scene.
Fig. 23: Experimental results in the real-world lobby scene. Despite the sparsely occupied spaces generating numerous frontiers and candidate viewpoints that demand higher computational overhead, the quadruped robot maintains continuous movement, demonstrating the computational stability of the proposed framework.
Autonomous exploration of multi-floor buildings remains challenging for ground robots because conventional 2D and 2.5D maps cannot represent overlapping traversable surfaces such as stairs, ramps, and multiple reachable elevations. This letter presents a multi-floor exploration framework based on an incremental reachable graph. Built as a sparse graph over reachable support surfaces, the graph preserves potentially valid connectivity through tentative graph elements under sparse observations and enables stable, physically reachable frontier detection. To guide exploration beyond the currently mapped floor, we project task-zone priors from an explored floor to initialize a hypothetical graph on the target floor and reconcile it incrementally with incoming observations. A hierarchical planner then jointly reasons over confirmed and hypothetical structures for global guidance. In simulation, the proposed method demonstrates improved exploration efficiency and mapping completeness compared to evaluated baselines. Furthermore, onboard real-world experiments validate its practical feasibility and real-time performance.
Zero-shot Object Navigation (ZSON) has shown promise for open-vocabulary target search in unseen environments, yet most existing systems remain tied to planar representations and single-floor assumptions. These assumptions become inadequate in real buildings, where navigation involves floors, stairs, landings, and vertically overlapping spaces. This article presents TravExplorer, a cross-floor embodied exploration framework that couples zero-shot semantic guidance with traversability-aware 3-D planning. TravExplorer maintains a unified volumetric map that distinguishes occupied structures from robot-reachable support surfaces and extracts traversable frontiers from connected support surfaces, including floors, stairs, and landings. A FOV-aware active perception strategy further resolves incomplete observations during cross-floor traversal. To reduce semantic-reasoning latency, a lightweight guidance module aligns a probabilistic instance map from online open-vocabulary segmentation with a spatial value map from fast image-to-text matching. Based on these geometric and semantic memories, a hierarchical planner performs target-aware frontier touring over object hypotheses, traversable frontiers, and stair landmarks, and generates executable cross-floor motions through foothold-guided 3-D search and vertically constrained local trajectory optimization. Experiments over 4,195 simulated episodes on HM3D and MP3D demonstrate consistent advantages over representative ObjectNav baselines. Fifty real-world trials on a Unitree Go2 further validate open-vocabulary target search across single-floor and cross-floor indoor environments without prior maps or human intervention. The code will be released at https://github.com/wuyi2121/TravExplorer.
Autonomous planetary exploration requires robots to navigate unknown, uneven terrain while assessing risk, traversability, and energetic cost. Quadruped scouts are well suited for this task because they can traverse irregular surfaces and gather mobility-relevant information during locomotion. This paper presents a terrain-aware exploration framework that combines exteroceptive and proprioceptive mapping for a quadruped robot in lunar-like environments. An onboard RGB-D camera builds robot-centered elevation maps, estimates geometric traversability, and derives navigation costs for autonomous planning. In parallel, proprioceptive measurements provide interaction-aware terrain cues that complement geometry-based assessment. Local maps are incrementally registered into a global multi-layer representation, which is used by an exploration module to select targets in unexplored regions of interest. The targets are reached by an autonomous navigation system that guides collision-aware motion using the available map and cost layers. Simulation results on NVIDIA Isaac Sim show autonomous exploration, map expansion, and spatial association between terrain geometry and robot-terrain interaction. Subsequent navigation using this information exhibits lower average Cost of Transport (CoT) than initial exploration.
Alberto Sanchez-Delgado, João Carlos Virgolino Soares, Victor Barasuol +1
Dynamic Legged Systems lab (DLS), Istituto Italiano di Tecnologia (IIT), Genoa, Italy