PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Quadruped-Provided Localization and Active Tracking for Micro-UAVs

Alejandro Lorite Mora, Andrés Faíña

arXiv:2610.06215v1cs.RO

TL;DR

Micro-UAVs can access confined spaces, but their payload limits constrain autonomous sensing and computation. This paper moves localization to a quadruped with an actively positioned arm camera, which supplies the drone’s map-frame pose. Lab flights achieved 12–16 mm external pose accuracy and reduced trajectory error with active tracking.

  • Problem

    Micro-UAVs can reach confined spaces, but their payload limits sensing and computation, and lightweight onboard navigation methods degrade in weak visual texture.

    Source: §1 · p. 2

  • Method

    A quadruped-mounted arm camera observes drone AprilTags and combines that pose with the quadruped’s localization to provide a map-frame pose while the arm tracks the moving drone.

    Source: §1 · p. 3, §5 · p. 11

  • Results

    12–16 mm external pose accuracy after calibration-offset removal; active tracking reduced trajectory error from 11.0 cm standing still to 6.9 cm with arm and body following.

    Source: §4 · p. 10, §5 · p. 11

  • Takeaways & Limitations

    The quadruped’s motion can improve the pose it supplies by keeping the drone in the close range where AprilTags are most accurate.

    Source: §4 · p. 10, §5 · p. 11

    Evaluation used lab motion-capture ground truth and presurveyed tags, and did not test planning or flight through a real confined opening.

    Source: §5 · p. 11

Abstract

from arXiv · show

Micro unmanned aerial vehicles (micro-UAVs) are small enough to reach confined spaces that larger robots cannot access, but too small to carry the sensing and computing power required for autonomous flight. We move the localization stack entirely off the aerial platform onto a quadruped robot with a 7-degree-of-freedom (DOF) arm, which supplies the micro-UAV (27 g bare, 42 g with fiducial markers) its full 6-DOF pose. A camera at the arm's end-effector detects AprilTag fiducial markers on the drone and composes that observation with the quadruped's own self-localization to place the drone in a shared map frame, so the ground robot localizes its partner, rather than only tracking it relative to the camera. The arm acts as an actively-controlled observer, repositioning to keep the drone in view as both robots move; the drone carries only an inertial measurement unit and fuses the external pose to fly commanded setpoints. In lab flights the external pose is accurate to 12-16 mm, enough to fly the drone autonomously within 2-5 cm of motion-capture-fed control. Having the quadruped actively follow the drone reduces the tracking error from 11.0 cm to 6.9 cm by holding the camera in the close range, where the markers are most accurate.

1 INTRODUCTION

Micro-UAVs can reach confined spaces, but their limited payload constrains onboard sensing and computation; existing external localization also leaves a gap in supporting a sensing-constrained drone while both robots move. This work uses a quadruped-mounted, actively positioned camera to provide the drone’s map-frame pose, with preliminary lab results showing accurate autonomous flight and improved tracking when the quadruped follows.

Motivation

  • Micro-UAVs can enter narrow, cluttered spaces, but limited payload capacity restricts sensing and computation needed for autonomous navigation.Larger UAVs can carry advanced sensors and processors but cannot access some confined or hazardous spaces.Source: §1 · pp. 1–2

Background

  • On sub-30 g platforms, stereo or LiDAR with VIO/SLAM exceeds available payload, computation, and power, while optical-flow and learned-navigation methods degrade in weak visual texture.The passage identifies poorly lit or featureless environments as settings where conventional VIO and LiDAR SLAM degrade.Source: §1 · p. 2

Related Work

  • Prior cooperative ground–air systems retain UAV localization sensors or impose update constraints, while moving the observer to sustain external localization of a separate sensing-constrained robot remains insufficiently studied.The proposed system instead gives the micro-UAV only an IMU and supplies its absolute 6-DOF pose from the quadruped while both robots move.Source: §1 · pp. 2–3

Contribution

  • After removing a constant calibration bias, external pose estimates localized the micro-UAV to 12–16 mm, enabling autonomous commanded trajectories within a few centimeters of motion-capture-fed control.Active arm tracking improved rather than degraded pose estimates by keeping the drone close, where fiducial accuracy is higher.Source: §1 · p. 3

2 METHOD

The method places the perception and control stack on a quadruped carrying a camera-equipped arm, while the micro-UAV serves as a setpoint-following actuator. The system composes AprilTag observations with the quadruped’s localization to produce a map-frame drone pose and actively repositions the observer to keep the drone visible.

2.1 System Overview

  • 2.1 System Overview: A Spot with a 7-DOF arm and end-effector camera observes a four-tag Crazyflie, while the quadruped hosts sensing, estimation, and control and the drone executes setpoints.The Crazyflie is 27 g bare and 42 g with its tag chassis; the integrated hardware also includes a Jetson AGX Orin.Source: §2.1 · pp. 3–4, Figure 1 · p. 4

2.2 Drone Localization in the Map Frame

  • 2.2 Drone Localization in the Map Frame: The pipeline composes the quadruped’s map localization with the arm camera’s AprilTag observation to estimate the drone’s full 6-DOF map-frame pose.The pose is emitted per detection without extrapolation; the onboard EKF coasts on the IMU through short gaps, and a watchdog commands landing if updates stop.Source: §2.2 · pp. 4–5, §2.1 · p. 4, Figure 2 · p. 5

2.3 Active Tracking

  • 2.3 Active Tracking: The arm predicts the drone’s next position, points the camera toward it, and uses inverse kinematics to re-center the view, optionally moving the quadruped as well.The controller targets an approximately 0.9 m stand-off as a preferred operating point, not a hard range limit; the algorithm also skips unsafe-close updates and clamps targets to arm reach.Source: §2.3 · pp. 5–6

2.4 Experimental Setup

  • 2.4 Experimental Setup: Indoor experiments used an OptiTrack system covering the workspace to provide 6-DOF ground truth at 120 Hz and, in selected runs, a comparison pose source.The lab working volume was 3 × 3 × 2 m, within the larger 4 × 4 × 2.5 m tracked volume.Source: §2.4 · p. 6

3 EXPERIMENTS AND RESULTS

The experiments measure AprilTag pose accuracy, test whether that pose supports autonomous flight, and assess whether quadruped tracking changes flight error. Active tracking lowers median trajectory error, while resolution presents an accuracy–throughput trade-off.

3.1 Experiment 1: AprilTag Pose Estimation Accuracy

  • In Experiment 1, 29 flights (134.8k samples) compared map-frame AprilTag poses against OptiTrack at 640×480 and 1280×720.Both raw delivered error and error after removing the constant setup offset were reported.Source: §3.1 · p. 7
  • 12 mm aligned median position error at 1280×720 versus 16 mm at 640×480, while orientation error was 3.4° versus 8.8°, respectively.The raw position error was approximately 50 mm at both resolutions because it was dominated by an approximately 46 mm fixed offset.Source: §3.1 · p. 7, Figure 3 · p. 8
  • 640×480 yielded approximately 40 Hz AprilTag poses versus 17 Hz at 1280×720, with gaps over 100 ms in 1.1% versus 15.7% of intervals.The throughput cost outweighed the millimeter-level accuracy gain at measured distances, so closed-loop experiments used 640×480.Source: §3.1 · pp. 7–8

3.2 Experiment 2: Tracking Error – MOCAP vs. AprilTag Pose Source

  • Experiment 2 compared AprilTag-fed and motion-capture-fed control over 18 flights, scoring trajectory error against commanded waypoints using median and IQR.The evaluation used raw closed-loop error without extrinsic alignment because it had no sensor-calibration offset to remove.Source: §3.2 · p. 8
  • AprilTag trajectory error rose from a median 7.8 cm at 1.5–2.0 m to 11.0 cm at 2.5–3.0 m, versus approximately 5.5 cm under motion capture.The resulting 2–5 cm penalty widened with distance, while AprilTag pose remained sufficient for autonomous flight.Source: §3.2 · p. 8

3.3 Experiment 3: Active Tracking

  • Experiment 3 compared a still quadruped, arm tracking, and arm-and-body tracking across eight flights at 640×480.The tracking conditions intentionally used different ranges: approximately 0.9 m for tracking and 2.7 m when still.Source: §3.3 · p. 9
  • 6.9 cm with arm-and-body tracking and 7.1 cm with arm tracking, versus 11.0 cm with the quadruped still, for median trajectory error.Tracking held the camera near 0.9 m, where AprilTag localization is more accurate; per-flight error remained bounded.Source: §3.3 · p. 9

4 DISCUSSION

Active tracking keeps AprilTags in their more accurate close range, reducing trajectory error; the discussion also identifies calibration, latency, and visibility limits. The system supplies a planner-ready map-frame pose, but its evaluation remains limited to lab localization and tracking.

4 DISCUSSION

  • The raw AprilTag–MOCAP gap is 46 mm, mostly fixed calibration error; after alignment, intrinsic accuracy is 12–16 mm.Flying on AprilTag pose adds a median 3–5 cm of tracking error over MOCAP, yet completes the trajectory autonomously.Source: §4 · pp. 9–10
  • 11.0 cm with the quadruped still versus 6.9 cm with arm and body tracking; Fig. 5 presents tracking error by motion condition and over time.The per-flight error stays bounded throughout each run.Source: Figure 5 · p. 10, §4 · p. 10
  • AprilTag pose reaches the controller a median 72 ms old (p95 ≈105 ms), with image delivery and rectification dominating latency.If pose age exceeds the 500 ms AprilTag watchdog timeout, the drone stops trusting it and lands.Source: §4 · p. 10
  • Occlusion at narrow openings, marker detection gaps, and accuracy loss at distance or oblique angles constrain operation; tests have not run outside the lab.The drone’s ≈7 min battery also caps mission length.Source: §4 · pp. 10–11
  • The system provides a planner-ready metric pose without motion capture or fixed infrastructure in GNSS-denied, confined environments.The evaluation covers localization and tracking rather than planning.Source: §4 · pp. 10–11

5 CONCLUSION

A quadruped with an actively positioned arm provides the full map-frame pose of a micro-UAV that has no localization sensing of its own. Lab experiments report 12–16 mm external-pose accuracy after calibration offset removal and lower trajectory error with active tracking than with the quadruped still.

5 CONCLUSION

  • A quadruped’s 7-DOF arm and camera estimate the micro-UAV’s full 6-DOF map-frame pose while both robots move, and the drone’s EKF combines it with IMU data.The 42 g drone has no exteroceptive sensors and flies under autonomous control.Source: §5 · p. 11
  • 11.0 cm with the quadruped still versus 6.9 cm with arm and body tracking; external-pose accuracy is 12–16 mm after removing a constant calibration offset.Flying on AprilTag pose yields 8–11 cm median path error, about 2–5 cm worse than motion-capture pose.Source: §5 · p. 11
  • Evaluation used lab motion-capture ground truth and presurveyed tags, and did not test outdoor operation, real confined openings, or planning.Marker detection can drop under fast yaw or long and oblique views; occlusion also occurs at narrow gaps.Source: §5 · p. 11
Loading 2610.06215v1…