PaperlayerView exact revision on arXiv

Source-linked AI summary

HARP: Handoff-Aware Recovery Planning for Active Target Acquisition in Multi-UAV Handoff

Heegyeong Kim, Alice James, Avishkar Seth, Endrowednes Kuantama, Richard Han

arXiv:2610.33640v1cs.RO

TL;DR

During multi-UAV handoff, a receiver with no prior target observations must acquire its own usable view while the sender continues tracking. HARP uses sender information to plan and check recovery motions, including admissible prefixes when full trajectories cannot be admitted. Simulations and replays demonstrated sender-guided first acquisition, and field evaluation validated recovery geometry under FoV limitation and occlusion.

  • Problem

    During handoff, a receiver without target-observation history must obtain a usable view before transfer can complete, even while the sender continues tracking.

    Source: §I · p. 1

  • Method

    HARP registers bounded sender predictions in the receiver’s map, distinguishes FoV failure from occlusion, and checks failure-specific recovery motions for admissibility.

    Source: §I · pp. 1–2

  • Results

    Closed-loop simulations and controlled replays demonstrated sender-guided first acquisition, while field evaluation validated FoV and occlusion recovery geometry in a physical two-UAV setting.

    Source: §VI · p. 8

  • Takeaways & Limitations

    When a complete trajectory cannot be admitted, HARP can execute a sampled admissible nonzero prefix and replan using updated observations.

    Source: §III · p. 2, §VI · p. 8

    The finite search provides no global completeness or minimum-time optimality guarantee.

    Source: §III · p. 2

Abstract

from arXiv · show

Continuous aerial tracking often requires multiple drones to maintain target coverage over long periods. During handoff, the receiving UAV may lack prior target observations and must actively acquire a usable view without disrupting the sender's ongoing coverage. This work introduces HARP, a handoff-aware recovery planner that registers bounded sender predictions in the receiver's local map, distinguishes field-of-view failure from physical occlusion, and generates failure-specific yaw, translation, and parallax viewpoints. Candidate motions are checked for receiver clearance, sender-fiducial observability, protected sender--target visibility, temporal validity, and deadline compliance. When a complete trajectory cannot be admitted, HARP may execute a sampled admissible nonzero prefix and replan from updated evidence. In simulation, HARP acquired the target in all 49 standard trials and in 15 of 16 stress trials. Field trials validated HARP's recovery geometry under field-of-view limitation and environmental occlusions.

I. INTRODUCTION

During handoff, the receiver may have no usable target view even as the sender continues tracking, so acquisition requires the receiver to recover its own view without disrupting sender coverage. HARP frames this as sender-guided first acquisition using bounded sender information and receiver-side evidence.

I. INTRODUCTION

  • HARP addresses first acquisition by a receiver with no target history, using the sender’s continuing track to recover a usable view without interrupting sender coverage.The receiver’s view may fail because the target is outside its FoV or physically occluded.Source: §I · p. 1
  • HARP registers bounded sender target predictions in the receiver’s map, distinguishes FoV failure from physical occlusion, and proposes failure-specific recovery viewpoints.It searches yaw first for FoV recovery, considers translation if needed, and uses target–blocker parallax for occlusion recovery.Source: §I · p. 1

II. PROBLEM FORMULATION

  • The task is to acquire the target before an external deadline while assuming continued sender observation and protecting the sender’s view during receiver motion.Acquisition requires a fresh, held sequence of actual receiver observations; path identity verification and responsibility transfer are outside the evaluated task.Source: §II · p. 2
  • Sender messages provide target position and velocity predictions with uncertainty bounds and a finite lifetime, while sender motion is predicted with fixed orientation during its validity interval.The receiver registers the sender frame using navigation and fiducial PnP; planning uses registered estimates, fiducial geometry, and local depth evidence, not target truth or complete scene geometry.Source: §II · p. 2
  • Candidate motions are checked for receiver clearance, sender-fiducial observability, protected sender–target visibility, motion limits, model validity, and deadline compliance.FoV recovery requires endpoint image readiness; occlusion recovery additionally requires complete target-support visibility through known-free viewing segments.Source: §I · p. 1, §II · p. 2
  • Admissibility checks are conditional on calibrated geometry, consistent registration and navigation, target containment within declared support, and the sender-motion model; they do not guarantee continuous-time safety under arbitrary errors or dynamic obstacles.Unknown map space cannot certify free clearance or map-supported visibility.Source: §II · p. 2

III. HARP METHOD

HARP selects a failure-specific recovery proposal, checks it for sampled admissibility, and admits motion for execution. When a full trajectory is not admitted, it may execute a nonzero prefix and replan with updated evidence.

III. HARP METHOD

  • When full-trajectory admission fails, HARP may retain a useful nonzero prefix for execution, then use updated evidence for subsequent attempts.The prefix endpoint need not satisfy the recovery condition, and the finite search provides no global completeness or minimum-time optimality guarantee.Source: §III · p. 2
  • HARP registers sender support, diagnoses viewing conditions, and selects one failure-specific proposal for sampled admissibility checking.The method’s geometric interpretation includes parallax recovery, prefix admission, and execution logic.Source: §III · p. 2
  • HARP uses sampled admissibility checking and execution admission rather than claiming complete search or minimum-time optimality.Updated evidence supports subsequent attempts within a finite search.Source: §III · p. 2

A. Registered Sender Information and Local Mapping

HARP registers sender-provided target motion in the receiver’s local frame and builds a local map from depth observations, maintaining validity and timing constraints on those inputs.

A. Registered Sender Information and Local Mapping

  • Sender motion is registered from navigation and matched fiducial pose samples, with camera-to-body extrinsics completing the transform composition.Inputs must be valid, unexpired, correctly framed, ordered, and temporally consistent.Source: A. Registered Sender Information and… · pp. 2–3
  • Component-wise sender uncertainty defines a transformed target support whose eight vertices preserve the registered support’s orientation.New valid messages replace the active model; expired models are not extrapolated.Source: A. Registered Sender Information and… · p. 3
  • Valid depth hits mark free rays and occupied endpoints, while clear-range observations add only free evidence; invalid observations add none.Map shifts retain overlapping evidence and initialize new coverage as unknown; snapshots preserve voxel arrays and timestamps.Source: A. Registered Sender Information and… · p. 3

B. Perception-First Diagnosis

HARP diagnoses target-view conditions using synchronized camera and map evidence, distinguishes field-of-view failure from occupied or unknown obstruction, and evaluates eligible future horizons independently.

B. Perception-First Diagnosis

  • Fresh held acquisition takes priority; otherwise, failed complete-support image readiness yields FFOV, occupied intersections yield FOCC, and unknown-only views yield FUNK.A known-free view without detection remains waiting for acquisition.Source: B. Perception-First Diagnosis · p. 3
  • The map-intersection test covers the complete target-view volume with conservative boundary handling, excluding target-endpoint-only contacts.The complete viewing volume is defined for the target-camera center.Source: B. Perception-First Diagnosis · p. 3
  • Future proposals classify each eligible source-relative horizon independently, so current diagnosis does not gate every future branch.Unknown-only horizons generate no active exploration proposals, though another future horizon may support FoV or occlusion recovery.Source: B. Perception-First Diagnosis · p. 3

C. Failure-Specific Viewpoint Proposals

HARP selects recovery viewpoints according to whether diagnosis identifies field-of-view failure or physical occlusion, then checks endpoint geometry and ranks candidates for admission.

C. Failure-Specific Viewpoint Proposals

  • For FoV recovery, HARP searches yaw at the current position first, then a finite translation lattice if no yaw endpoint meets readiness and marker-observability requirements.Translation endpoints must have known-free clearance; yaw-only body clearance is checked during trajectory admission.Source: C. Failure-Specific Viewpoint Proposals · p. 3
  • For FOCC, HARP uses the nearest intersecting occupied component’s projected silhouette and selected boundary edges to propose parallax recovery viewpoints.Positive-depth blocker anchors and normals toward the free image side define the recovery geometry.Source: C. Failure-Specific Viewpoint Proposals · p. 3
  • The parallax proposal enforces pixel-margin separation for all eight support vertices, rejects excessive displacement or zero-displacement solutions, and verifies exact separation after reprojection.No displacement floor or certified linearization-error bound is imposed.Source: C. Failure-Specific Viewpoint Proposals · pp. 3–4
  • A full OCC endpoint requires complete known-free target visibility, while unknown visibility may allow only a prefix-capable goal; OCC candidates precede FoV candidates.Only the selected candidate is checked for trajectory admissibility, with no same-snapshot fallback after rejection.Source: C. Failure-Specific Viewpoint Proposals · p. 4

D. Trajectory Construction and Sampled Admissibility

HARP constructs a smooth trajectory to a selected recovery goal and samples it against time-varying target and sender geometry. If full admission fails, a nonzero prefix may be retained, but finite sampling does not guarantee continuous-time safety.

D. Trajectory Construction and Sampled Admissibility

  • A cubic smoothstep trajectory connects the current receiver state to the selected recovery goal, with duration also used for proposal reachability.Samples follow configured time, distance, and yaw increments and include both endpoints.Source: D. Trajectory Construction and Sampled… · p. 4
  • Every sampled receiver pose must retain known-free clearance, preserve the protected sender–target viewing volume, and keep the sender marker observable.Target and sender geometry are queried at the planning time plus each sample time.Source: D. Trajectory Construction and Sampled… · p. 4
  • If the first failing sample occurs after the initial sample, HARP may retain the nonzero prefix ending immediately before it, even though that endpoint may not satisfy recovery conditions.The prefix is admissible relative to the sampled nominal reference; finite-sample checks do not guarantee the executed closed-loop trajectory continuously.Source: D. Trajectory Construction and Sampled… · p. 4

E. Receding Execution and Acquisition

HARP revalidates and guards recovery motions during execution, then returns to sensing and replanning when acquisition or continuation fails. Its simulation evaluation uses held receiver detections and a 120-second episode timeout.

E. Receding Execution and Acquisition

  • At authorization, retained motion duration must fit the applicable target and sender validity intervals and acquisition deadline.Before execution, HARP also checks artifact freshness, identity, start-state consistency, and temporal validity.Source: E. Receding Execution and Acquisition · p. 4
  • Before issuing motion commands, HARP checks receiver clearance, marker observability, and sender–target sightline protection; new sender information or map updates trigger revalidation or replanning.FoV motions and parallax prefixes may reuse stored samples, while continuing a full parallax motion requires regenerating blocker-dependent geometry.Source: E. Receding Execution and Acquisition · p. 4
  • Planning-time admissibility and execution-time guarding are separate, so runtime checks do not guarantee every executed intermediate pose satisfies all planning constraints.Authorization may generate intermediate poses and time-indexed predictions that differ from the finite nominal samples checked during planning.Source: E. Receding Execution and Acquisition · p. 4
  • Fresh held acquisition ends recovery, while failed continuation, stale evidence, prefix completion, or arrival without acquisition returns the receiver to hold and replanning.If no admissible motion is available, HARP withholds further recovery motion and continues sensing.Source: E. Receding Execution and Acquisition · p. 4

A. Platform and Evaluation Criteria

  • The ROS-based simulator uses a simplified UAV model, static voxel geometry, synthetic depth, and fiducial corners for PnP registration.Target truth and complete scene geometry are used only to generate simulated observations and for offline evaluation.Source: A. Platform and Evaluation Criteria · p. 4
  • In simulation, held acquisition requires three consecutive valid receiver detections, each with at least 20 visible pixels and visible fraction ≥0.60.Detections must have increasing timestamps and age ≤0.5 s; the episode timeout is 120 s from the first accepted registered model.Source: A. Platform and Evaluation Criteria · p. 5

B. Scenarios and Comparison Design

The evaluation defines FoV, occlusion, and cluttered-scene scenarios, then tests configuration sensitivity and targeted internal comparisons. Trial populations and comparison procedures are reported separately.

B. Scenarios and Comparison Design

  • F1–O2 comprise FoV recovery and local-blocker scenarios, while C3D tests a moving target among cluttered three-dimensional structures.F1 tests attitude change; F2 requires translation; O1 uses a stationary target with a local blocker, and O2 combines occlusion with target motion.Source: B. Scenarios and Comparison Design · p. 5
  • R1 increases target-acceleration support, R2 coarsens the OCC yaw grid, and R3 applies both changes relative to primary configuration R0.These arms probe sensitivity to faster support growth and coarser occlusion search.Source: B. Scenarios and Comparison Design · p. 5
  • The evaluation contains 160 trials across R0–R3, with canonical F1–O2 results reported separately from the C3D stress population.R0 has one F1 regression run and 16 trials each for F2, O1, O2, and C3D.Source: B. Scenarios and Comparison Design · p. 5
  • The eight-run internal ablation pilot changes one HARP component at a time and is intended to expose mechanism-level differences, not provide a statistical comparison.The listed variants disable active recovery, remove translation, admit only complete artifacts, or use a point target representation.Source: B. Scenarios and Comparison Design · p. 5
  • The comparison design also includes four-attempt O1 Full/No-prefix repetition, a prospective O1 geometry matrix, and separate captured-input replay for identical-input decisions.Reported latency ranges are F2 0.7, O1 2.8–3.4, O2 16.4–62.8, and C3D 6.5–10.2 s; path entries are medians.Source: B. Scenarios and Comparison Design · p. 5

V. RESULTS

HARP achieved held acquisition in the canonical trials and most C3D stress trials, while controlled comparisons examined viewpoint changes, support-dependent admission, and prefix-enabled progress. Results also identify limits: prefix admission cannot help when proposal generation yields no candidate.

A. First Acquisition and Receiver Viewpoint

  • 49/49 canonical R0 trials and 15/16 C3D stress trials achieved held acquisition; latency and path statistics cover acquired runs only.F1 acquired in 1.3 s with zero translation and −2.517° yaw; F2 acquired in 0.7 s, and O1 median latency and path length were 3.1 s and 0.2225 m.Source: A. First Acquisition and Receiver… · p. 5
  • O2 acquired in every R0 trial; its 62.8 s latency tail involved only 0.0962 m of receiver travel, with delay concentrated before the first recorded EXECUTING state.The recorded sequence does not isolate a single source of delay.Source: A. First Acquisition and Receiver… · p. 5
  • In a replay of O2 run 0008, moving-camera samples became acquisition-eligible, while the fixed diagnosis pose remained below the 0.60 visible-fraction threshold throughout.With target, sender, scene, and timestamps unchanged, the intervention isolates receiver viewpoint change as necessary for acquisition along this episode.Source: A. First Acquisition and Receiver… · p. 5

B. Receiver Motion and Support-Dependent Admission

  • For the same F2 candidate, complete-support checking retained a 1.420527 s prefix after finding arrival-FoV failure, rather than admitting the 1.515229 s full artifact.The comparison demonstrates a change in admitted action extent, not a general acquisition advantage over centroid checking.Source: B. Receiver Motion and… · p. 6, Figure 4 · p. 6

C. Prefix-Enabled Progress and Geometry Sensitivity

  • In O1 replay, prefix permission admitted 0.852091 s of a trajectory whose full 1.609506 s artifact failed the selected-horizon cap; without permission, no motion was admitted.Live repetitions and the geometry study also showed progress with prefix permission, including acquisition in three of four eligible geometry conditions versus none for No-prefix.Source: C. Prefix-Enabled Progress and… · p. 6, Table III · p. 6, Table IV · p. 7
  • In G04, both modes failed before trajectory checking because finite proposal generation produced no admissible candidate, so prefix permission could not provide progress.Two geometry-study conditions were excluded because D2 had already observed the target before recovery.Source: C. Prefix-Enabled Progress and… · p. 6

D. Configuration Sensitivity

Configuration sensitivity varied by scenario, and the examined checks showed no recorded violations while remaining limited to sampled evidence. Computation time varied with scene geometry, supporting event-driven rather than fixed-rate planning.

D. Configuration Sensitivity

  • R2 retained acquisition throughout F2 and O1 but had occasional O2 and C3D timeouts, while R1 and R3 failed throughout F2 but mostly acquired in O2 and C3D.Unequal screening and confirmation allocations, and unisolated rejection causes, limit these findings to sensitivity within the tested configurations.Source: D. Configuration Sensitivity · p. 7

E. Constraint Evidence and Computational Cost

  • No violations appeared in the recorded checks for receiver clearance, sender-fiducial observability, protected sightlines, and model validity, but sampled checks do not establish continuous-time safety.In a representative C3D run, target-centroid containment held for 5/8 examined selected-arrival queries, with a maximum miss of 0.156769 m; executionwide containment could not be reconstructed.Source: E. Constraint Evidence and… · p. 7
  • OCC generation medians ranged from about 143 ms in two Full O1 repetitions to 456–467 ms in G04, while one frozen O2 record reached 4.11 s.The authors characterize HARP as event-driven rather than claiming a fixed-rate 10 Hz planning loop.Source: E. Constraint Evidence and… · p. 7

F. Flight Validation

Field validation tested HARP’s recovery geometry under field-of-view loss and occlusion, while the conclusion identifies broader planning and evaluation limits for future work.

F. Flight Validation

  • Five of ten flights proposed rotation-only FoV recovery, with yaw magnitudes of approximately 22–36°; other cases already placed target support inside the receiver FoV.The inside-FoV cases could still require further visibility assessment because of occlusion.Source: F. Flight Validation · p. 7
  • The three cases proposed an 18° yaw rotation, a 0.675 m lateral displacement, and a 0.450 m lateral displacement with a 15° yaw adjustment.Full target support first met image-margin and occlusion-clearance conditions at approximately 2.25 s, 2.80 s, and 1.95 s, respectively; proposed arrivals were 2.36 s, 2.89 s, and 1.96 s.Source: F. Flight Validation · pp. 7–8, Figure 7 · p. 7
  • 2.25 s was the median time to satisfy geometric visibility conditions, and 0.450 m was the median proposed translation distance across three model-based cases.All three cases satisfied the conditions in approximately 2.80 s.Source: F. Flight Validation · p. 8

VI. CONCLUSION

  • HARP’s field evaluation validated FoV and occlusion recovery geometry in a physical two-UAV setting, while future work targets planning efficiency, continuous-time safety guarantees, and broader deployment conditions.The identified conditions include larger, more dynamic environments and uncertainty in communication, sensing, localization, and target motion.Source: §VI · p. 8
Loading 2610.33640v1…