Efficient Autonomous Navigation of a Quadruped Robot in Underground Mines on Edge Hardware
A quadruped robot achieved a 100% success rate in dark, GPS-denied mines using only a low-power Intel NUC computer.
Researchers Yixiang Gao and Kwame Awuah-Offei have published a paper detailing a breakthrough in robotic autonomy for extreme environments. They developed a complete navigation system enabling a Boston Dynamics Spot quadruped robot to operate fully autonomously in the challenging conditions of underground mines, which feature narrow passages, uneven terrain, total darkness, and no GPS signal. Crucially, their system runs entirely on a low-power Intel NUC edge computer, eliminating the need for bulky GPU-accelerated hardware or constant network connectivity, which are often unavailable in such remote locations. This represents a significant shift from recent learning-based approaches that require extensive data and powerful compute.
The technical stack integrates LiDAR-inertial odometry for positioning, scan-matching for localization against a prior map, terrain segmentation, and a visibility-graph global planner with a local path follower. After a single mapping pass, the robot can navigate to any goal within the known map without any environment-specific AI training. In rigorous field trials at an experimental mine, the system was tested with four target locations of varying difficulty. It successfully completed all 20 trials (5 repetitions per target), accumulating over 700 meters of fully autonomous travel and achieving a 100% success rate with an SPL (Success weighted by Path Length) score of 0.73. This proves robust, real-time perception-to-action control is possible on constrained hardware, paving the way for practical deployment of inspection and logistics robots in hazardous industrial settings.
- System runs fully on a low-power Intel NUC edge computer with no GPU or network requirements, enabling operation in infrastructure-poor areas.
- Achieved a 100% success rate across 20 trials, autonomously traversing over 700 meters in a dark, GPS-denied underground mine environment.
- Uses a classic robotics stack (LiDAR odometry, prior maps) instead of learned AI models, allowing deployment after a single mapping pass with no extra training.
Why It Matters
Enables practical, cost-effective robotic inspection and logistics in hazardous industries like mining, construction, and disaster response without relying on cloud compute.