Resilient Terrain Navigation with a 5 DOF Metal Detector Drone

Patrick Pfreundschuh, Rik Bähnemann, Tim Kazik, Thomas Mantel, Roland Siegwart, Olov Andersson

This paper, recently published by researchers at ETH Zurich, tackles one of the most dangerous and technically demanding tasks in robotics: autonomous landmine detection. While drones have long been proposed as a solution for humanitarian demining, the physics of metal detection make it a "Goldilocks" problem—the sensor must be kept very close to the ground and perfectly parallel to the surface, even as the terrain rolls and dips.

TL;DR

The researchers developed a 5-DOF (Degrees of Freedom) tricopter capable of independent position and tilt control, allowing it to "hug" uneven terrain with a metal detector. By fusing LiDAR, GNSS, and IMU data through a resilient factor graph and using a smart receding-horizon planner, the drone can autonomously map buried metallic objects even under tree canopies or in featureless fields where GPS and LiDAR often fail.

What Problem Does This Paper Address?

Traditional drones are "underactuated." To move forward, a standard quadrotor must tilt its entire body. If you strap a metal detector to it, that sensor tilts too, moving away from the ground and losing its signal. Furthermore, landmines are often found in "denied" environments—under trees (no GPS) or in wide-open, featureless fields (where LiDAR-based SLAM struggles to find "anchors").

To be useful for demining, a drone needs:

  1. Mechanical Dexterity: The ability to tilt the sensor to match the ground slope without necessarily moving in that direction.
  2. Resilient Localization: A way to stay on track when sensors like GPS or LiDAR become unreliable.
  3. Smart Planning: A way to navigate complex slopes smoothly rather than jerkily reacting to every bump.

Methods

The team's approach is built on three pillars:

  • The Hardware (5-DOF MAV): They used the Voliro-T, a tricopter with tilting rotors. Unlike a standard drone, it can control its pitch and heading independently of its translational movement. This allows the rigidly attached metal detector to stay parallel to the ground while the drone moves at a constant height.
  • Resilient State Estimation: They implemented a factor graph-based fusion system. Instead of just switching between GPS and LiDAR, the system performs online extrinsic calibration. It estimates the transformation between the local odometry and the global GNSS frame in real-time. If the LiDAR "slips" (common in tall grass) or GNSS drops out (under trees), the factor graph maintains a consistent trajectory.
  • Receding-Horizon Planner: Rather than a "greedy" planner that tries to perfectly align with the ground at every single millisecond (which leads to erratic, battery-draining motion), their planner looks ahead several meters. It optimizes for a "smooth" yaw and pitch trajectory that satisfies ground-alignment constraints while minimizing unnecessary turning.

Datasets

The authors evaluated their system using:

  • Simulation: A scaled-down "crater landscape" in Gazebo to test the planner against baseline approaches.
  • Real-World Experiments: Conducted in outdoor environments at ETH Zurich, including a dense tree canopy (GNSS-denied) and a featureless field (LIO-challenging).
  • Target Mapping: They buried 9 metallic gas cartridges in an undulated grassy field to validate the autonomous mapping capability.

Key Insights and Ideas

  • Smoothness vs. Perfection: One of the most interesting findings is that "perfect" alignment is the enemy of efficiency. By allowing for a tiny, controlled alignment error (staying within the sensor's 7.5° tolerance), the drone can fly much smoother and faster than if it tried to track every micro-slope of the terrain.
  • Handling Sensor Degeneracy: The paper highlights that LiDAR and GNSS are complementary. LiDAR fails in "geometric voids" (flat fields), while GNSS fails in "signal voids" (forests). Their factor graph approach treats the relationship between these sensors as a dynamic variable, allowing the drone to navigate through "blind spots" of either sensor.
  • 5-DOF is the Sweet Spot: While 6-DOF drones exist, the authors show that 5-DOF (omitting roll) is sufficient for surface tracking, reducing mechanical complexity while providing the necessary decoupling of orientation and translation.

Key Results

  • Efficiency: In simulations, the proposed planner was 10.6% faster than standard look-ahead baselines and significantly smoother, reducing yaw changes by roughly 24° per step.
  • Localization Robustness: In real-world tests involving a total loss of GNSS and significant LiDAR drift, the system maintained a landing accuracy of 0.15m, whereas standard LiDAR-Inertial Odometry (LIO) deviated by over 1.5m.
  • Successful Detection: The system successfully identified and mapped 9 out of 9 buried targets in a complex, undulated environment. In a second test with a physical obstacle (a pole), it successfully avoided the obstacle and mapped 8 out of 9 targets (missing only the one directly next to the obstacle for safety).

Limitations and Future Work

The authors noted that the metal detector used was a "black box" that performed internal calibrations they couldn't control, leading to inconsistent sensitivity. They suggest using sensors that provide raw signal data in the future.

Additionally, while the obstacle avoidance works, it is currently a reactive layer. Integrating obstacle avoidance directly into the "coverage" part of the planner (re-planning the path on the fly rather than just nudging the drone sideways) would make the system even more robust for highly cluttered environments like forests or ruins.

Summary updated

Abstract

Micro aerial vehicles (MAVs) hold the potential for performing autonomous and contactless land surveys for the detection of landmines and explosive remnants of war (ERW). Metal detectors are the standard detection tool but must be operated close to and parallel to the terrain. A successful combination of MAVs with metal detectors has not been presented yet, as it requires advanced flight capabilities. To this end, we present an autonomous system to survey challenging undulated terrain using a metal detector mounted on a 5 degrees of freedom (DOF) MAV. Based on an online estimate of the terrain, our receding-horizon planner efficiently covers the area, aligning the detector to the surface while considering the kinematic and visibility constraints of the platform. As the survey requires resilient and accurate localization in diverse terrain, we also propose a factor graph-based online fusion of GNSS, IMU, and LiDAR measurements. We validate the robustness of the solution to individual sensor degeneracy by flying under the canopy of trees and over featureless fields. A simulated ablation study shows that the proposed planner reduces coverage duration and improves trajectory smoothness. Real-world flight experiments showcase autonomous mapping of buried metallic objects in undulated and obstructed terrain.
BibTeX
@inproceedings{pfreundschuh2023resilient,
  author = {Patrick Pfreundschuh and Rik Bähnemann and Tim Kazik and Thomas Mantel and Roland Siegwart and Olov Andersson},
  title = {{Resilient Terrain Navigation with a 5 DOF Metal Detector Drone}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  doi = {10.1109/icra48891.2023.10161253},
}

Loading related content...

Figure