DARP-3D: UAV Coverage & 3D Scheduling

Updated 4 July 2026

The paper presents a terrain-aware extension for multi-UAV coverage that adjusts waypoint altitude and camera angles to improve 3D reconstruction metrics.
DARP-3D is also applied as a specialized online Dial-a-Ride in 3D space, offering deterministic and randomized competitive guarantees.
The method retains the underlying 2D partitioning while integrating lightweight 3D refinements, demonstrating robust performance in both simulated and real-world environments.

DARP-3D is a context-dependent research term rather than a single universally standardized method. In its most explicit contemporary use, it denotes a terrain-aware extension of the Divide Areas for Optimal Multi-Robot Coverage Path Planning algorithm for multi-UAV coverage path planning, where 2D region partitioning is preserved but waypoint altitude and camera orientation are adapted from a pre-scanned 3D model of the region of interest (Karakontis et al., 23 Jul 2025). In a separate online scheduling literature, the same label is used as a specialization of online Dial-a-Ride to 3D Euclidean space, analyzed through the Mimic framework with deterministic and randomized competitive guarantees (Bienkowski et al., 2021). The acronym DARP also appears in several unrelated fields, so the meaning of DARP-3D is determined by domain context rather than by a single canonical definition.

1. Nomenclature and scope

The term has at least two technically substantive uses in the provided literature. One belongs to multi-UAV coverage path planning and photogrammetric reconstruction; the other belongs to online scheduling for pickup-and-delivery systems. Several additional papers use the acronym DARP for unrelated mechanisms and explicitly either do not define a 3D-specific variant or discuss only a possible extension.

Usage	Domain	Status in source
DARP-3D	Terrain-aware multi-UAV coverage path planning	Explicitly named algorithm
DARP-3D	Online Dial-a-Ride in 3D Euclidean space	Explicit specialization
DARP	Imitation learning, DRAM refresh, SSL, 3D detection	Different acronym expansion

This ambiguity matters because the UAV interpretation is an algorithmic extension of a coverage planner, whereas the scheduling interpretation is a metric specialization of an online competitive-analysis framework. Conflating them obscures both the underlying optimization problem and the claims attached to the name.

2. Terrain-aware DARP-3D for multi-UAV coverage

In the UAV planning literature, DARP-3D extends a 2D multi-UAV coverage planner to account for 3D structure in the region of interest. The motivating problem is that many commercial multi-UAV coverage path planning systems treat the region of interest as a flat 2D plane and therefore generate fixed-altitude, nadir-view missions. The reported consequences are occlusions, poor capture of vertical surfaces, and incomplete 3D reconstructions. DARP-3D addresses this by retaining the baseline DARP partitioning and modifying waypoint altitude together with per-waypoint camera yaw and pitch, using a pre-scanned point cloud of the scene (Karakontis et al., 23 Jul 2025).

The method is explicitly offline. It assumes a pre-scanned 3D model, obtained in the paper through a sparse initial mission followed by reconstruction with OpenDroneMap. Inputs include the baseline 2D paths $P=\{P_1,\dots,P_n\}$ , one per UAV; a point cloud $M$ ; and user parameters such as $z_{\text{offset}}$ , $\Delta z$ , step, $r_0$ , and $\Delta r$ . The operational variables are the terrain-adjusted altitude $H(p)$ for each waypoint and the camera yaw $\psi(p)$ and pitch $\theta(p)$ . The formulation is heuristic and two-stage rather than a single global optimization.

Several constraints are central. Terrain standoff is enforced through altitude adjustment above the estimated ground. The camera is mounted below the airframe, so upward pointing is disallowed. Inter-UAV deconfliction is inherited from DARP’s area partitioning, and DARP-3D does not alter partition topology. The paper frames the practical objective as improved 3D reconstruction quality—measured through precision, recall, and $F_1$ against ground truth—while preserving feasible multi-UAV mission execution (Karakontis et al., 23 Jul 2025).

A common misconception is that DARP-3D replaces the DARP partitioner. It does not. The partitioning and workload distribution remain those of the underlying 2D planner; the 3D contribution is terrain-following altitude scheduling and viewpoint steering.

3. Algorithmic pipeline, camera steering, and empirical behavior

The UAV version follows a two-stage pipeline. Stage 1 runs a baseline 2D planner, demonstrated with DARP, executes an initial mission, and reconstructs a coarse 3D model. Stage 2 applies terrain-aware refinement to the baseline waypoints. The altitude-adjustment module uses KD-tree queries over the point cloud to estimate terrain elevation under each waypoint, sets the waypoint altitude with a vertical offset, and inserts intermediate waypoints when the height difference from the previous waypoint exceeds $M$ 0. The camera-angle module performs a sphere search around each waypoint, keeps only points below the UAV, expands the search radius until such points are found, computes an average altitude over the retained points, selects the point whose $M$ 1-coordinate is closest to that average, and then computes yaw $M$ 2 and pitch $M$ 3 to aim at that target point. The source describes the angle computation procedurally rather than by explicit closed-form equations (Karakontis et al., 23 Jul 2025).

The reported computational complexity of the refinement step is $M$ 4 in the number of waypoints in the longest path. On dense paths over a $M$ 5 region of interest, full computation finished within 2 minutes. This suggests that the added 3D reasoning is lightweight relative to photogrammetric reconstruction and mission execution, although the method remains dependent on the availability of a pre-scan.

The evaluation spans four Unreal Engine and AirSim environments—Dubai Rock, Pallet Pile, Village, and Cathedral—and a real-world flight test using DJI hardware. For fair comparison, the experiments keep the total image budget equal between baseline DARP and DARP-3D, with DARP-3D splitting the budget between a sparse first pass and a terrain-aware second pass. The most pronounced gains occur in scenes with strong vertical structure.

Environment	DARP $M$ 6 at 5 cm	DARP-3D $M$ 7 at 5 cm
Rock	68.57%	82.31%
Cathedral	21.07%	65.33%
Village	7.27%	22.35%

The same study reports that, at 5 cm, Rock improves from 69.63% to 82.65% in precision and from 67.55% to 81.97% in recall; Cathedral improves from 30.58% to 71.07% in precision and from 16.07% to 60.44% in recall; Village improves from 7.60% to 23.69% in precision and from 6.96% to 21.15% in recall. Pallet Pile shows comparable performance across methods, with only limited benefit, which the authors attribute to low vertical variation and high sensitivity in small piles. A fixed 60° pitch baseline improves over nadir imaging but remains well below DARP-3D on terrains with significant vertical structure (Karakontis et al., 23 Jul 2025).

The real-world demonstration used three DJI Phantom 4 Pro UAVs at the NTUA School of Mining & Metallurgical Engineering. The baseline day used 80 m altitude, 3 m/s speed, 80% sidelap, 80% frontlap, $M$ 8 pitch, equal area distribution, and 132 images. The DARP-3D campaign used a sparse first pass with 50% frontlap and 31 images, then a terrain-aware second pass with 80% frontlap and 146 images. The resulting model is reported to fill many gaps that were empty in the baseline reconstruction.

4. DARP-3D as online Dial-a-Ride in $M$ 9

In online scheduling, DARP-3D denotes online Dial-a-Ride specialized to 3D Euclidean space. The metric space is $z_{\text{offset}}$ 0 with Euclidean distance, travel time equals distance under unit speed, each request $z_{\text{offset}}$ 1 has release time $z_{\text{offset}}$ 2, pick-up point $z_{\text{offset}}$ 3, drop-off point $z_{\text{offset}}$ 4, and weight $z_{\text{offset}}$ 5, and there may be one or multiple servers with capacity $z_{\text{offset}}$ 6. The objective is weighted total completion time, $z_{\text{offset}}$ 7, where $z_{\text{offset}}$ 8 is the time at which request $z_{\text{offset}}$ 9 is fully completed (Bienkowski et al., 2021).

The analysis places DARP variants in the class of $\Delta z$ 0-resettable scheduling problems with $\Delta z$ 1. The key auxiliary object is a $\Delta z$ 2-schedule $\Delta z$ 3 on the released prefix $\Delta z$ 4, evaluated through the prize-collecting surrogate

$\Delta z$ 5

Here, completed requests contribute their weighted completion times inside the auxiliary schedule, while uncompleted released requests are charged as if completed at time $\Delta z$ 6. The Mimic algorithm repeatedly computes an auxiliary schedule $\Delta z$ 7 minimizing $\Delta z$ 8, executes it during a phase, and then resets the server to the origin before the next phase.

For DARP in this framework, $\Delta z$ 9, and the deterministic schedule boundaries are $r_0$ 0. The specialized deterministic routine waits until $r_0$ 1, then for each phase computes $r_0$ 2, executes it on $r_0$ 3, and resets on $r_0$ 4. The paper proves that this yields a 4-competitive deterministic algorithm. With a randomized offset $r_0$ 5, the competitive ratio becomes

$r_0$ 6

again for $r_0$ 7 (Bienkowski et al., 2021).

The guarantees are dimension-free in the sense that they depend only on metric properties and the reset bound. As a result, Euclidean $r_0$ 8 is a direct instantiation rather than a distinct proof technique. The framework covers any number of vehicles, any capacity, and both preemptive and non-preemptive DARP, but it does not include time windows in its guarantees. A central practical limitation is that computing the auxiliary schedule $r_0$ 9 is NP-hard, so the result is information-theoretic rather than a polynomial-time deployment guarantee.

Outside those two uses, DARP is an overloaded acronym. In 3D object detection, Pyramid R-CNN introduces Density-Aware Radius Prediction as a module inside the pyramid RoI head; it adapts the neighborhood radius for RoI feature aggregation but is not called DARP-3D (Mao et al., 2021). In imitation learning, Difference-Aware Retrieval Policies are introduced as a retrieval-based alternative to behavior cloning, and the paper explicitly states that it does not define a 3D-specific variant or “DARP-3D,” although it discusses principled 3D adaptations for $\Delta r$ 0 poses, point clouds, and voxel or RGB-D inputs (Pfeifer et al., 8 Jun 2026).

The same acronym appears in memory systems as Dynamic Access Refresh Parallelization, combined with SARP for DRAM refresh-access overlap; those papers explicitly state that they do not define DARP-3D, though they discuss how the mechanism could conceptually extend to 3D-stacked DRAM such as HBM (Chang et al., 2017, Kim et al., 2020). In semi-supervised learning, DARP denotes Distribution Aligning Refinery of Pseudo-label, and the source states that no 3D-specific variant is presented even though the method is modality-agnostic and could be applied to 3D classification or segmentation (Kim et al., 2020).

This terminological overload explains why “DARP-3D” can refer either to a named UAV planning algorithm or to a 3D metric specialization in online Dial-a-Ride, while closely related 3D papers using DARP do not adopt the same label.

6. Limitations, assumptions, and enduring significance

The UAV planner’s limitations are explicit. It requires a pre-scan and is therefore not an online adaptive planner. It does not perform explicit occlusion reasoning or ray-casting; the downward-hemisphere search is a pragmatic proxy for visible structure. Upward camera pointing is disallowed because of the camera mount. Gains diminish on nearly flat scenes, and abrupt yaw or pitch changes between waypoints can reduce overlap and leave gaps. The paper identifies smoother camera transitions and better overlap consistency as natural future directions. Its practical significance lies in showing that substantial 3D reconstruction gains can be obtained without redesigning the core DARP partitioner, and in doing so through an implementation that integrates with DJI mission files, AirSim, OpenDroneMap, CloudCompare, and an open-source release in TerraPlan (Karakontis et al., 23 Jul 2025).

The scheduling use has a different limitation profile. Its results are strong in competitive-analysis terms—4-competitive deterministically and $\Delta r$ 1-competitive with randomization—but the auxiliary optimization is NP-hard, and the guarantees exclude time windows. Its significance is therefore theoretical and structural: it shows that online Dial-a-Ride in 3D Euclidean space fits a unified resettable scheduling framework and inherits improved bounds previously unavailable for online TRP and DARP (Bienkowski et al., 2021).

Taken together, DARP-3D is best understood not as a single method but as an overloaded label with two substantive meanings in current research. In UAV coverage planning it denotes a concrete terrain-aware algorithm layered on top of 2D DARP; in online scheduling it denotes a Euclidean 3D specialization of Dial-a-Ride analyzed by Mimic. The term’s technical content is therefore inseparable from the literature in which it appears.