Omnidirectional Solid-State mmWave Radar Perception for UAV Power Line Collision Avoidance

AI summary

Overview

Research area: Robotics — UAV perception, mmWave radar sensing, and detect-and-avoid for aerial vehicles operating near power infrastructure.

Technical level: Advanced (assumes familiarity with radar field-of-view concepts, autopilot state estimation, and vector-based avoidance control).

Scope: This paper builds and field-tests a spherical, radar-only perception layer made of six solid-state mmWave modules on a small quadcopter, characterizing radar behavior around power lines and deriving a lightweight avoidance algorithm from it.

What This Paper Is About

Power lines are thin, feature-poor, and hard to judge in distance, both for human UAV pilots watching a compressed video feed and for autonomous systems relying on cameras. That makes collisions likelier, and because multirotors can move in any direction, a sensor that only looks forward leaves dangerous blind spots. This paper surrounds a small UAV with six mmWave radar devices to get sensing coverage in all directions, studies how those radars actually behave when pointed at power lines, and uses that behavior to build an avoidance algorithm that is tested on real decommissioned power lines.

Key Contributions

1. A lightweight, solid-state multi-radar perception architecture that synthesizes omnidirectional (spherical) coverage around a small UAV frame, using one long-range sensor forward and five shorter-range antenna-on-package sensors covering the remaining directions.
2. A characterization of how mmWave radar behaves with power line targets in this omnidirectional arrangement, including measured versus expected fields of view and distance-measurement error statistics for each of the six sensors.
3. A robust power line avoidance algorithm that exploits radar's specific interaction with power lines — chiefly the finding that a detection approximates the closest point on the wire, which reduces the avoidance problem to a two-dimensional one.
4. Extensive real-world validation on an approximately 35 meter stretch of decommissioned 3-phase transmission line, covering high-speed approaches, oblique approaches, descents between conductors, and a 1.2 mm diameter wire.

Main Findings

• Detection range: With the long-range front sensor, power lines are typically detected from beyond 10 meters away. The shorter-range sensors detect them from approximately 7 meters.

• Avoidance speeds: Avoidance maneuvers are performed at velocities upwards of 10 m/s when approaching with the front sensor, and at up to 5 m/s in the directions covered by the shorter-range sensors.

• Thin-wire capability: The system detected and avoided a steel wire just 1.2 mm in diameter — the smallest gauge available during testing.

• Distance accuracy: Across the turntable validation tests, the overall mean distance error was 0.0607 m and the overall RMSE was 0.0663 m, with the overall minimum at −0.0004 m and maximum at 0.1136 m. The authors describe an average error around 6 cm as acceptable for obstacle avoidance, and note further testing is required to determine whether this is a constant offset, a percentage offset, or a differently scaling error.

• Field-of-view discrepancy: All five IWR6843AOPEVM sensors measured much narrower azimuth FoV than the expected 120° — 76° for top and bottom, 75° for left, right, and rear. Elevation measured 108° (top), 106° (bottom), 104° (left and right), and 135° (rear) against an expected 120°. The front IWR6843ISK sensor measured slightly wider than expected: 41° elevation (expected 30°) and 122° azimuth (expected 120°).

• Blind spots that did not appear outdoors: The narrower-than-expected azimuth FoV created narrow blind spots in the XY plane during laboratory testing — only four sensors detect the target during a full revolution in any given plane. However, real-world tests approaching the power line from an oblique angle succeeded without issue, which the authors interpret as the blind spots possibly being an artifact of the corner-reflector laboratory setup.

• The "closest point" phenomenon: While a radar boresight is within roughly ±30° of perpendicular to the power line, the detected point is approximately equal to the closest point on the wire. The paper calls this P_a and uses it to treat detections as closest points directly, without extra processing.

• Trajectory safety margin: In a 16 second avoidance test plotted in the plane perpendicular to the power lines, detections from the various sensors were consistent throughout, and the UAV never came closer than 1 meter to a detected power line.

• Independence from RTK: Ground truth positioning came from an RTK-capable GNSS receiver providing ±2 cm accuracy, but the paper states the system has also been successfully tested without RTK.

Methodology in Plain English

The team took a Holybro QAV250 quadcopter with a 250 mm wheelbase and mounted six Texas Instruments mmWave radar devices around the frame, pointing along the ±X, ±Y and ±Z directions. The front device is a long-range unit (IWR6843ISK, 60–64 GHz, 120° azimuth and 30° elevation FoV), chosen because forward flight is where the highest speeds occur; it is rotated so its wide azimuth coverage sweeps up and down, keeping forward targets in view even when the UAV pitches steeply at speed. The other five are shorter-range antenna-on-package units (IWR6843AOPEVM, 60–64 GHz, 120° azimuth and 120° elevation FoV), reduced to just the main board to save weight. The left and right units are angled 15° toward the front — 75° and −75° relative to the front sensor — to compensate for the front sensor's narrower horizontal coverage. All radars run stock firmware in 3D mode (3RX, 4TX) at 10 Hz, interfaced through a custom ROS2 node. Data and power run through a USB hub and a Raspberry Pi 4 running Ubuntu 20.04 and ROS2 Foxy; a Pixhawk 4 Mini autopilot runs PX4 v1.13.0 with RTPS. Two DC-DC converters regulate the 4S LiPo battery to 5 V, one for the sensors and one for the Pi.

To check whether coverage really was spherical, they mounted the UAV on a turntable three times — once per plane — with a metallic corner reflector on a thin wooden stick placed exactly 1 meter from the rotation axis, then slowly rotated the UAV and mapped all six sensors' measurements into the UAV frame.

To learn how radar behaves near power lines, they hovered the UAV in front of a wire and commanded a slow yaw, recording detections alongside the UAV's relative rotation. This let them compare the angle between the detected point and the closest point against the angle between the radar boresight and the closest point, which produced the P_a finding.

The avoidance algorithm splits into three scenarios within the plane K_p perpendicular to the wires. At medium speeds, the system computes tangents to each detection inside an avoidance sphere and steers along them, scaling each tangent by how directly the UAV is heading at the wire; corrections are computed twice, once for the UAV's current velocity and once for the user-commanded velocity, then summed and clamped so the output never exceeds the user's speed. At high speeds, an e-brake horizon cone expands along the velocity direction with length ||v_d||² / (2 a_max) + s_margin and a 2:1 length-to-width ratio, giving a half-angle of atan(1/4) = 0.245 radians; detections inside it trigger immediate maximum braking until speed drops. At very close range, detections inside a smaller safety sphere trigger a rejection velocity that grows stronger the closer the detection is to the sphere's center. The final output velocity is sent to the autopilot.

Why This Matters

This is a radar-only system — no camera, no image processing — which matters because power line detection has to work in all weather and at long range, and because the paper shows direction information about the wires can be extracted directly from radar data rather than requiring a camera pipeline as in prior work. It is also the omnidirectional piece that earlier single-radar power line systems lacked, and the paper argues that no prior UAV radar obstacle-avoidance work has both spherical coverage and real-world power line validation.

Real-world applications:

• Safety layer for manually piloted UAVs, warning or overriding a pilot whose distance judgment of wires is poor from the ground or through a compressed video downlink.
• Autonomous inspection flights along transmission corridors, where the same radar stack can act as a protection layer on top of an existing autonomy stack.
• Emergency landing and autonomous takeoff/landing near power lines, since the system has upward- and downward-facing sensors and can navigate between conductors with only a downward velocity command.
• Thinner-wire environments such as telephone wires, since the system avoided a 1.2 mm steel wire, which was the thinnest gauge available during testing.

Industry relevance: the work targets the intersection of drone operations and power transmission infrastructure, where collision consequences range from harmless to UAV crashes, infrastructure damage, and disruption to power distribution. The authors note the platform was also tested successfully without RTK, which matters for cost and for operations where precise GNSS correction is unavailable.

Future Directions

• Characterizing the distance error more precisely: the paper explicitly states that further testing is required to determine whether the roughly 6 cm average error is a constant offset, a percentage offset, or a differently scaling error.
• Resolving the field-of-view discrepancy: whether the narrower measured azimuth FoV of the antenna-on-package sensors is a genuine hardware property or an artifact of the corner-reflector target. The paper notes real power line testing seems to greatly diminish the discrepancy, and suggests substituting another IWR6843AOPEVM for the front IWR6843ISK to gain FoV at the cost of range.
• Testing even thinner wires: 1.2 mm was the smallest gauge available, so the true lower limit of detectable wire diameter is not established.
• Broader environmental validation: the paper's field results come from a single approximately 35 meter decommissioned stretch of 3-phase transmission line; behavior across other geometries, live-line conditions, and weather is not reported.

Target Audience

Robotics and UAV perception researchers working on collision avoidance and radar sensing; engineers integrating detect-and-avoid layers into commercial drone autopilots, particularly in the power utility and infrastructure inspection sector; and readers interested in how physical sensor behavior — rather than heavy processing — can be exploited to make an avoidance algorithm simple and lightweight.

Authors’ abstract

Detecting and estimating distances to power lines is a challenge for both human UAV pilots and autonomous systems, which increases the risk of unintended collisions. We present a mmWave radar-based perception system that provides spherical sensing coverage around a small UAV for robust power line detection and avoidance. The system integrates multiple compact solid-state mmWave radar modules to synthesize an omnidirectional field of view while remaining lightweight. We characterize the sensing behavior of this omnidirectional radar arrangement in power line environments and develop a robust detection-and-avoidance algorithm tailored to that behavior. Field experiments on real power lines demonstrate reliable detection at ranges up to 10 m, successful avoidance maneuvers at flight speeds upwards of 10 m/s, and detection of wires as thin as 1.2 mm in diameter. These results indicate the approach's suitability as an additional safety layer for both autonomous and manual UAV flight.