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

This paper presents a fully autonomous navigation stack for a quadruped robot running on low-power edge hardware without GPUs or network connectivity, which successfully navigated complex underground mine environments with a 100% success rate across 20 trials using a combination of LiDAR-inertial odometry, map-based localization, and classical planning algorithms.

The Gist

Imagine you are trying to guide a four-legged robot dog through a pitch-black, narrow, and bumpy underground mine. There is no GPS (like Google Maps), no Wi-Fi, no internet, and it's so dark you can't see your hand in front of your face. Most robots would get lost, confused, or require a supercomputer the size of a refrigerator to figure out where to go.

This paper describes a clever solution that lets a Boston Dynamics "Spot" robot do this job all by itself, using a tiny, low-power computer (like a high-end laptop) strapped to its back.

Here is the story of how they did it, explained with some everyday analogies.

1. The Problem: The "Dark Maze" Challenge

Underground mines are the ultimate test for robots. They are like a maze made of rock, with:

• No Light: Cameras (which most robots use) are useless here. It's like trying to drive a car with your eyes closed.
• No GPS: You can't ask for directions because the satellite signal doesn't reach underground.
• No Internet: You can't send the data to a cloud server to solve the problem; the robot has to think for itself.
• Rough Terrain: The ground is uneven, with rocks, holes, and low ceilings.

Most modern robots try to solve this with "AI" (Artificial Intelligence) that needs massive, expensive graphics cards (GPUs) and huge amounts of training data. But the researchers wanted something simpler, cheaper, and more reliable that could run on a small computer.

2. The Solution: The "Old-School" Smart Navigator

Instead of using a "black box" AI that learns by trial and error, the team built a navigation system using classic, logical rules. Think of it like giving the robot a very strict, step-by-step instruction manual rather than teaching it to "guess" the way.

The system works in four main steps, like a relay race:

Step A: Feeling the Way (LiDAR + IMU)

Since the robot can't see, it uses a LiDAR sensor. Imagine a bat using echolocation, but instead of sound, it shoots out invisible laser beams 10 times a second to map the walls and floor in 3D.

• The Analogy: It's like walking through a dark room with a cane, tapping the walls to know where you are.
• The Boost: They also added an IMU (an accelerometer, like the one in your phone that knows which way is up). This helps the robot know exactly how it's tilting or turning, even if the laser beams get a little jumpy.

Step B: The "Check-In" System (Localization)

Even with a cane, you can get lost if you walk too far (this is called "drift"). To fix this, the robot carries a digital map of the mine in its memory.

• The Analogy: Imagine you are walking in a dark forest, but you have a mental picture of the trees. Every few steps, you stop, look at the trees around you, and say, "Ah, that tree looks like the one on my map. I am here."
• How it works: The robot constantly compares its laser scan of the current tunnel against the pre-loaded map. This corrects any mistakes and keeps it on the right track without needing a GPS.

Step C: Reading the Floor (Terrain Segmentation)

The robot needs to know what is safe to walk on. Is that pile of rocks a path, or is it a wall?

• The Analogy: Imagine the robot is wearing special glasses that turn the floor green (safe to walk) and the walls/rocks red (do not walk).
• How it works: A mathematical filter looks at the laser data and instantly decides: "This is flat ground," or "This is a wall." It ignores the ceiling (which is low in mines) so the robot doesn't get confused by overhead pipes.

Step D: The Route Planner (Visibility Graph)

Once the robot knows where it is and what is safe, it needs a route to the goal.

• The Analogy: Instead of drawing a straight line (which might go through a wall), the robot draws a "string" connecting all the open corners of the tunnel. It finds the shortest path along these strings.
• The Trick: They didn't make the robot calculate this from scratch every time. Before the mission, a human drove the robot through the mine once to create a "roadmap" (a visibility graph). The robot just loads this roadmap and follows it, only adjusting if it sees a new obstacle (like a fallen rock).

3. The Results: The Perfect Score

The team tested this system 20 times in a real experimental mine in Missouri. They sent the robot to four different locations, ranging from an easy straight hallway to a deep, dark, winding tunnel.

• Success Rate: 100%. The robot reached every single goal without human help.
• Distance: It traveled over 700 meters (about 7 football fields) completely autonomously.
• Hardware: It did all this on a tiny computer (Intel NUC) that uses very little power, with no GPU and no internet.

Why This Matters

This paper proves that you don't need a supercomputer or a "learning" AI to navigate dangerous places. By using smart, logical rules and good sensors, you can build a robot that is:

1. Reliable: It doesn't get confused by darkness.
2. Efficient: It runs on a small battery-friendly computer.
3. Ready for Industry: It's simple enough to be used in real mines to check for safety hazards without risking human lives.

In short: They taught a robot dog to navigate a pitch-black cave by giving it a laser "cane," a mental map, and a strict set of rules, proving that sometimes the simplest, most logical approach is the most powerful.

Technical Summary

Here is a detailed technical summary of the paper "Efficient Autonomous Navigation of a Quadruped Robot in Underground Mines on Edge Hardware" by Yixiang Gao and Kwame Awuah-Offei.

1. Problem Statement

Autonomous navigation in underground mines presents a unique set of challenges that render many state-of-the-art robotic solutions ineffective:

• Environmental Constraints: Near-total darkness (no ambient light), GPS-denied conditions, narrow passages, and uneven terrain with elevation changes.
• Infrastructure Limitations: Lack of wireless communication (WiFi/cellular) prevents offloading computation to the cloud or remote servers.
• Hardware Limitations: Industrial deployment requires low-power, robust hardware. Recent learning-based approaches (e.g., Vision-Language-Action models) rely on high-end GPUs and massive datasets, which are incompatible with the power budgets and latency requirements of edge devices in mines.
• Operational Need: Unlike exploration tasks, mine operations often require reliable, repeated point-to-point navigation (e.g., inspection, reconnaissance) rather than mapping unknown spaces.

The core problem addressed is achieving fully autonomous, real-time navigation for a quadruped robot in these harsh conditions using only a low-power CPU (no GPU, no network, no learned components).

2. Methodology

The authors propose a classical, interpretable navigation stack running on ROS 2, integrated into a Boston Dynamics Spot quadruped. The system operates entirely on an Intel NUC 13 edge computer (13th Gen i7, 40W power limit) without discrete GPUs.

A. Hardware Platform

• Robot: Boston Dynamics Spot.
• Sensors:
  • Velodyne VLP-16 LiDAR: 16-channel, 360° FOV, 10 Hz. Used for odometry and localization.
  • Yahboom IMU: 6-axis, 100 Hz, fused with LiDAR for state estimation.
  • Thermal Camera: Used for situational awareness but not for control (due to lack of training data integration).
• Compute: Intel NUC 13 (CPU-only processing).

B. Software Pipeline

The system consists of four tightly integrated subsystems:

1. State Estimation & Localization (Two-Stage):

   • Local Odometry: Uses FAST-LIO2 to fuse LiDAR and IMU data. It performs motion undistortion and point-to-plane scan matching against an incremental map, outputting 6-DoF odometry at 10 Hz.
   • Global Drift Correction: Uses Normal Distributions Transform (NDT) to align real-time scans against a prior point cloud map (captured during a single teleoperated pass). This corrects accumulated drift without maintaining a full pose graph, ensuring global consistency.

2. Terrain Perception:

   • Ceiling Filtering: Removes points above 1.5m to ignore roof structures common in mines.
   • Segmentation: Applies an Approximate Progressive Morphological Filter (PMF) to classify points as "traversable ground" or "obstacle" (walls, rubble). This is crucial for distinguishing flat floors from vertical structures in 2D planning.

3. Global Planning:

   • FAR Planner: A visibility-graph-based planner. A pre-computed visibility graph (846 nodes) is loaded at startup, allowing immediate path planning to any goal without an exploration phase.
   • Dynamic Updates: During traversal, live LiDAR scans update the graph to route around new obstacles (e.g., fallen rock) not present in the prior map.

4. Local Path Following:

   • Regulated Pure Pursuit: Tracks the global path with a lookahead distance of 0.5m.
   • Velocity Regulation: Adjusts speed based on path curvature and proximity to the goal to ensure stability on uneven terrain and prevent stalling.

3. Key Contributions

• CPU-Only Edge Navigation: Demonstrated a fully autonomous navigation stack running in real-time on a low-power Intel NUC without GPU acceleration or network connectivity.
• Zero-Learning Approach: The system relies entirely on classical, interpretable algorithms (LiDAR-inertial odometry, NDT, PMF, Visibility Graphs), requiring no environment-specific training or domain adaptation. It generalizes to arbitrary goals within a known map after a single mapping pass.
• Robustness in Darkness: Validated that LiDAR-only perception is sufficient for navigation in pitch-black underground environments where camera-based methods fail.
• Open Source: The authors released the full navigation stack and field data.

4. Experimental Results

The system was validated in the Missouri S&T Experimental Mine over 20 trials (5 repetitions × 4 target locations) covering 717 meters of autonomous traversal.

• Success Rate: 100% (20/20 trials). The robot reached within 1.0m of the target in every trial without human intervention.
• Success Weighted by Path Length (SPL): 0.73 ± 0.09.
  • Goal 1 (Easy): 0.87 SPL (near-optimal path).
  • Goal 4 (Arbitrary Start): 0.66 SPL (demonstrating generalization to varying start points).
• Path Efficiency: Path ratios ($p/\ell$) ranged from 1.16 to 1.53. The deviation from the geodesic shortest path is attributed to the visibility graph structure and conservative curvature-limited tracking.
• Latency: End-to-end perception-to-action latency had a median of 176 ms (mean 195 ms), well within the 400 ms planning cycle. NDT localization runs asynchronously, ensuring it does not block control responsiveness.
• Drift Correction: NDT provided a median XY correction of 1.1 cm per step, bounding accumulated drift to 31–103 cm over long paths. Without NDT, drift would have exceeded the success threshold on longer missions.

5. Significance and Future Work

Significance:
This work proves that high-reliability autonomous navigation in hazardous, infrastructure-denied environments does not require high-power GPUs or data-hungry learning models. By leveraging classical geometric methods on low-power edge hardware, the system offers a practical, deployable solution for industrial mine inspection and reconnaissance. It addresses the "prohibitive" constraints of underground mines (no light, no network, limited power) effectively.

Limitations & Future Work:

• Prior Map Dependency: The system currently requires a pre-existing map; it cannot operate in completely unknown environments.
• Dynamic Obstacles: The system does not yet handle dynamic obstacles (e.g., moving personnel) in real-time avoidance, though it can route around static new obstacles.
• Parameter Tuning: Terrain segmentation parameters are tuned for specific mine geometries and may require adjustment for different environments.
• Future Directions: The authors plan to integrate exploration for unknown environments, add dynamic obstacle avoidance, and utilize the thermal camera for miner tracking.