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

Yixiang Gao  ·  Kwame Awuah-Offei

Department of Mining and Explosive Engineering
Missouri University of Science and Technology

Missouri S&T

Pre-print, under review

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 × 4 targets) and an overall Success weighted by Path Length (SPL) of 0.73 ± 0.09.

Key Results

100% Success Rate (20/20)
0.73 SPL (Anderson et al.)
717 m Autonomous Traversal
40 W Edge Compute (No GPU)
Swipe sideways to view full table.
Goal Trials Success Path (m) SPL Time (s)
G1 (Easy) 5/5 100% 13.1 ± 0.2 0.87 ± 0.01 27 ± 1
G2 (Intersection) 5/5 100% 25.5 ± 1.7 0.71 ± 0.05 60 ± 6
G3 (Deep) 5/5 100% 51.3 ± 1.0 0.68 ± 0.01 127 ± 7
G4 (Entrance) 5/5 100% 53.5 ± 8.8 0.66 ± 0.04 131 ± 20
All 20/20 100% 35.9 0.73 ± 0.09 86

All Trial Paths

Executed paths (red) vs. geodesic shortest paths (blue dashed) for all 20 trials across 4 goal locations, overlaid on the mine elevation map.

Tap or click image to enlarge.

System Architecture

Tap or click image to enlarge.

Hardware

Boston Dynamics Spot with custom navigation payload
Swipe sideways to view full table.
ComponentModelNotes
RobotBoston Dynamics Spot
ComputeIntel NUC13 (i7-1360P)12C/16T, 32 GB RAM, no discrete GPU, 40W sustained
LiDARVelodyne VLP-1610 Hz, 360° FoV, 0.75–30 m range
IMUYahboom100 Hz, 6-axis
ThermalTOPDON TC00130 Hz (recorded only, not used for navigation)

Navigation Demos

Each demo shows one representative trial at 5× speed. The robot navigates autonomously from start to goal using only LiDAR and IMU — no cameras, no GPS, no network.

Mission 1 demo
Goal 1 (Easy) — 11 m, straight traverse
Mission 2 demo
Goal 2 (Intersection) — 18 m, through intersection
Mission 3 demo
Goal 3 (Deep) — 35 m, narrow & sloped passages
Mission 4 demo
Goal 4 (Entrance) — 35 m avg, arbitrary start to entrance

Field Footage

Third-person phone footage from the underground mine. The full mission video shows a complete Goal 4 trial at 5× speed; the arrival clips below show the robot reaching each goal location.

Goal 1 arrival (5×)
Goal 2 arrival (5×)
Goal 3 arrival (5×)
Goal 4 arrival (5×)
Phone footage (5×)
3D LiDAR view (5×)

Citation

@misc{gao2026efficientautonomousnavigationquadruped, title={Efficient Autonomous Navigation of a Quadruped Robot in Underground Mines on Edge Hardware}, author={Yixiang Gao and Kwame Awuah-Offei}, year={2026}, eprint={2603.04470}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2603.04470}, }

Acknowledgment

This research was funded by a grant from the Centers for Disease Control's National Occupational Safety and Health (grant no. U60OH012350-01-00).