Efficient Autonomous Navigation of a Quadruped Robot in Underground Mines on Edge Hardware

Yixiang Gao; Kwame Awuah-Offei

cs.ROMar 4, 2026

Efficient Autonomous Navigation of a Quadruped Robot in Underground Mines on Edge Hardware

Authors: Yixiang Gao, Kwame Awuah-Offei

Organizations: Department of Mining and Explosive Engineering, Missouri University of Science and Technology, Rolla, USA

Abstract

Embodied navigation in underground mines faces significant challenges, including narrow passages, uneven terrain, near-total darkness, GPS-denied conditions, and limited communication infrastructure. While recent learning-based approaches rely on GPU-accelerated inference and extensive training data, we present a fully autonomous navigation stack for a Boston Dynamics Spot quadruped robot that runs entirely on a low-power Intel NUC edge computer with no GPU and no network connectivity requirements. The system integrates LiDAR-inertial odometry, scan-matching localization against a prior map, terrain segmentation, and visibility-graph global planning with a velocity-regulated local path follower, achieving real-time perception-to-action at consistent control rates. After a single mapping pass of the environment, the system handles arbitrary goal locations within the known map without any environment-specific training or learned components. We validate the system through repeated field trials using four target locations of varying traversal difficulty in an experimental underground mine, accumulating over 700 m of fully autonomous traverse with a 100% success rate across all 20 trials (5 repetitions x 4 targets) and an overall Success weighted by Path Length (SPL) of 0.73 \pm 0.09.

Figures & tables

Panel of Figure 1 — Fig. 1 : (a) A third-person view of the environment and the robot during an autonomous navigation mission in an underground mine with no GPS, WiFi, or cellular signal. (b) The prior map point cloud overlaid with a real-time LiDAR scan and terrain classification (white = traversable, red = obstacle); insets show the fisheye camera view (near-dark, true lighting) and thermal image. (c) Localized map with traversed path (dashed) and robot current position; photo insets show the robot at the entrance (right), mid-journey (center), and approaching the goal (left).

Panel of Figure 2 — Fig. 2 : System architecture of the autonomous navigation stack. Dashed lines indicate offline data loaded at startup. FAST-LIO2 fuses LiDAR and IMU inputs to produce odometry and motion-undistorted scans. Two parallel branches—NDT localization (drift correction against a prior map) and terrain analysis (ground/obstacle segmentation)—feed the FAR Planner, which computes a global path tracked by a regulated pure pursuit controller.

Panel of Figure 3 — Fig. 3 : Boston Dynamics Spot with the custom navigation payload. The sensor suite includes a Velodyne VLP-16 LiDAR, a Yahboom IMU 3 3 3 We used an external IMU because accessing the robot’s internal high-frequency IMU requires an extra license. , and a TOPDON TC001 thermal camera. All processing runs onboard on an Intel NUC 13 mini PC.

Fig. 4 : Goal and start locations on the mine elevation map (20 trials total). (a) Goals 1–3 share a common start near the entrance. (b) Goal 4 (Entrance): all 5 start locations marked, each from a distinct position in the mine. Dashed lines show the geodesic shortest paths used to compute the reference distance $\ell$ for SPL evaluation.

Panel of Figure 5 — Fig. 5 : Qualitative timeline and trajectory overview. Left (a–c): Goal 3 (Entrance $\rightarrow$ Deep)—the robot navigates through narrow, sloped, pitch-black passages. Right (d–f): Goal 4 (Deep $\rightarrow$ Entrance)—starting from a dark location, the robot returns to the entrance.

Fig. 6 : Executed paths (solid red) vs. geodesic shortest paths (dashed blue) on the elevation map for all 20 trials across the four target locations. $p_{i}$ is path length and $\ell_{i}$ is geodesic distance.

Goal	Succ.	Path (m)	$p/\ell$	SPL	Time (s)
G1 (Easy)	5/5	$13.1$ 0.2	$1.16$ 0.02	$0.87$ 0.01	$27$ 1
G2 (Inter.)	5/5	$25.5$ 1.7	$1.41$ 0.09	$0.71$ 0.05	$60$ 6
G3 (Deep)	5/5	$51.3$ 1.0	$1.47$ 0.03	$0.68$ 0.01	$127$ 7
G4 (Entr.)	5/5	$53.5$ 8.8	$1.53$ 0.09	$0.66$ 0.04	$131$ 20
All	20/20	35.9	1.39	$0.73$ 0.09	86

TABLE I : Navigation performance (mean

\pm

std over 5 repetitions per goal; 20 trials total). SPL per Anderson et al. [ 21 ] . “All” reports the mean of per-goal means.

Quadruped Robot Navigation

See all 21 papers

MineXplore: An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment
Jun 3, 2026Abhishek S, Badrikanath Praharaj, Sreeram MVRL Benchmarks Robot Navigation
BuildMachineLabs
TRAIN: Terrain-Aware Interactive Navigation for Quadruped Robots via Reinforcement Learning
Feb 13, 2026Wei Zhu, Irfan Tito Kurniawan, Ye Zhao +1Quadruped Robot Navigation Terrain-Aware Robot Locomotion
Department of Robotics, Graduate School of Engineering, Tohoku University, 980-8579, Sendai, Japan. · Laboratory for Intelligent Decision and Autonomous Robots, Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30313, USA.
MulDP: Multimodal Diffusion Policy for Autonomous Quadruped Parkour Navigation across Complex Terrains
Sep 3, 2026Kangmai Hu, Yueqi Zhang, Peng Zhai +5Robot Policy Learning Robot Navigation
College of Intelligent Robotics and Advanced Manufacturing, Fudan University, Shanghai 200433, China

Jun 3, 2026J/K move · Enter open · S save

MineXplore: An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment

Abhishek S, Badrikanath Praharaj, Sreeram MV

BuildMachineLabs

RL Benchmarks Robot Navigation Robotics RL Exploration Robot Navigation Benchmarks

Open paper AskPDF arXiv

Underground mines present extreme conditions for autonomous robot navigation: GPS is denied, lighting is degraded, and tunnel topology is loop-rich and non-convex. Simulation benchmarks grounded in real production-mine geometry and compatible with GPU-accelerated learning pipelines do not yet exist in the open-source ecosystem. We present MineXplore, an open-source MuJoCo-based navigation benchmark derived from the Leung et al. 2017 Chilean underground copper mine dataset. The environment reconstructs a 104,423 sq.m tunnel network through an six-stage contour-to-MJCF pipeline incorporating octagonal wall cross-sections, LiDAR-sourced jagged wall geometry, three terrain friction zones, a global 5 degree incline, and periodic spot lighting. Geometric fidelity is validated at an Intersection over Union (IoU) of 0.9538 against the source survey map, and surface texture similarity scores 79.4% across six structural dimensions. A single-agent PPO baseline trained via RLlib across five independent random seeds achieves a best rolling coverage of 88.89% (3 of 5 seeds reaching the 90% coverage target), confirming that MineXplore supports stable and reproducible policy learning under realistic underground sensing and topology.

Abstract

Quadruped Robot Navigation

Explore similar work

MineXplore: An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment

TRAIN: Terrain-Aware Interactive Navigation for Quadruped Robots via Reinforcement Learning

MulDP: Multimodal Diffusion Policy for Autonomous Quadruped Parkour Navigation across Complex Terrains